Generating Beautiful UIs May 08, 2026

Introducing Multi-LoRA on Cerebras Inference May 06, 2026

Today, we are launching Multi-LoRA—multi-adapter support for Low-Rank Adaptation—on Cerebras Inference in private preview. Multi-LoRA lets teams use many LoRA adapters with a single shared base model, so they can specialize model behavior for different domains, tasks, customers, and workflows. It advances our mission of making Cerebras Inference the fastest and simplest way to run specialized AI applications. LoRAs are lightweight adapters trained to specialize a base model. Instead of fine-tuning all of the base model’s parameters, teams train a much smaller set of adapter weights that can be applied at inference time. This makes specialization practical and cost efficient without requiring a separate full model for each variant. How Multi-LoRA works on Cerebras Inference Cerebras Inference handles the serving infrastructure behind the endpoint. We manage the base model and adapter serving path, so teams can focus on building the…

MoE at Scale: Making Sparse Models Fast on Real Hardware September 03, 2025

In this video we discuss scaling MoE models on modern hardware and address key optimization challenges. If you can’t open the video displayed above, please use this link to open it on YouTube: https://youtu.be/MXo9LEYzwkg Mixture-of-Experts (MoE) models allow you to increase total parameter count without proportional increase in compute, letting you train bigger and better models efficiently (Soboleva, 2025a). You might wonder if extracting theoretical benefits from MoE models requires significant engineering work. After all, your part 3 implementation (Soboleva and Tiwari, 2025) trained perfectly fine on a small acceleration node (and even your laptop). An important point here is that you used only 4 experts and 124M backbone parameters, but production systems like DeepSeek-V3, Qwen3, etc., use hundreds of experts and huge backbones. Try scaling to their sizes with our previous implementation on the GPU, and you will quickly…

MoE Math Demystified: What Does 8x7B Actually Mean? October 14, 2025

This video breaks down MoE inference arithmetic and deployment bottlenecks across different hardware setups. If you can’t open the video displayed above, please use this link to open it on YouTube: https://youtu.be/gHpDBoyCOrE What does 8x7B actually mean? You probably thought it meant 8 experts with 7B active parameters per token. We did too. Turns out it is actually 13B active parameters. But wait, where does 13B come from? This is exactly the kind of confusion this post clears up (skip to the answer). We'll explain what those numbers actually mean for inference by answering how much memory you need, how many GPUs, and what the commonly hit bottlenecks are in production deployment. We'll show that single-GPU deployment is memory-bound, multi-GPU setups are communication-bound, and specialized hardware like Cerebras WSE is compute-bound. Originally, we set out to write a simple post…

Case Study - Cognition x Cerebras December 10, 2025

Dec 10 2025 Case Study - Cognition x Cerebras The Dawn of Real-Time Coding Agents TL;DR Powered by Cerebras Inference, Cognition's SWE-1.6 and the SWE-grep family deliver frontier-level coding performance up to ~5x faster than on GPU, with a smoother agent experience that keeps developers in flow while they explore codebases, ship features, and debug complex systems. The Challenge AI is redefining software development, turning natural language prompts into working code. But for an AI coding assistant to be useful, it must feel instantaneous and handle large, complex projects seamlessly. Until now, AI coding on GPU meant frustrating delays - 20 to 30 second generation times that broke a developer's concentration. Even slight lags forced context-switching. Developers were stuck choosing between smaller, faster models that lacked skill and larger models that were too slow. The industry needed a solution that…

Figma - MultiAgents April 16, 2026

Everything is easier now. I have been toying around with agent orchestration for a while now. I’m currently running 10-20 agents around the clock.AI agents are now capable of bringing my ideas to life. Like many developers, I’ve been feeling the token anxiety. I can do much more now than ever before, and every time I have a spare minute I want to kick off another agent session. - I see a cool product I don’t want to pay for? Codex will build it for me. - I have a silly idea I want to see come to life? Codex will build it for me. - I get mildly annoyed doing the same thing over and over? Codex pls. If you have an army of infinitely patient, intelligent, and helpful agents waiting for your next command, why shouldn’t we take…

Lessons learned from building multi-agent workflows April 16, 2026

I pay my upfront subscription ($200/month), write what I hope is the right prompt (prompt AND context engineer), and wait. 35 minutes later, it’s still 'synthesizing', 'perusing', 'effecting', and 'germinating' (who came up with these). By the end, I have files of bad code, a bloated context window, and I’m counting the remaining tokens on my left hand. Okay, I grab an apple, compact, type some heavy handed verbal abuse, re-explain everything from scratch, and pray the next attempt gets further than the last one…. only to be disappointed by the same result. By now, the spark and joys of AI coding are long dead. Stop being a one-shot Sloperator This is the single-agent ceiling. Every developer building with AI agents hits it the moment their project graduates from a 3D HTML snake game to anything more practical. This happens…

The Debate of MCP vs. CLI Centers on Speed April 06, 2026

MCP had a formative year. Then it had a turbulent week. Perplexity CTO Denis Yarats walked on stage at Ask 2026 and announced that Perplexity was moving away from MCPs… and back to APIs and CLIs. Immediately, Twitter split into two camps. Not surprising, given MCP grew from an Anthropic open standard in November 2024 to industry-wide adoptions with over 97 million monthly downloads in just thirteen months(1) across a range of companies and platforms. Yet Perplexity, a prominent AI company, chose to walk away from it. MCP's overhead isn't arbitrary. The protocol works by(2) guiding model interactions down specific, auditable paths: every tool call carries its full schema definition, every auth handshake runs end to end, and every step waits for the previous one to complete before the next begins. That predictability is exactly what enterprise deployments need. But…

Why speed wins: faster inference is about more than just quicker answers–it’s the new path to accuracy February 19, 2026

Feb 19 2026 Why speed wins: faster inference is about more than just quicker answers–it’s the new path to accuracy Watching extraordinary athletes compete at the Winter Olympic games in Milano-Cortina these last two weeks, is a reminder that world-class performance demands excellence across many fronts—and is hard to sustain indefinitely. Biathlon, which originated in the 1700s as a race-and-shoot event between ski patrol units at the Sweden-Norway border, offers a particularly good example. Athletes cross-country ski at near-maximum effort and then immediately transition into target shooting. The sport doesn’t reward athletes who are “fast” or “accurate” in isolation—it crowns the best combination of skiing speed and marksmanship under fatigue, weather, and pressure. Raw speed is not only necessary to stay ahead of competitors, but also to provide enough margin to shoot clean and avoid costly time penalties. The sport…

Partner Spotlight: Armis + Cerebras Enable Teams Build and Secure Software Faster March 27, 2026

Mar 27 2026 Partner Spotlight: Armis + Cerebras Enable Teams Build and Secure Software Faster At Cerebras, we’ve always believed that speed changes what’s possible. In software development, that means more than faster generation or faster inference. It means faster iteration, faster validation, and faster action. That’s why we’re excited to spotlight Armis, whose Armis Centrix™ for Application Security unifies application security across the software lifecycle. With Armis and Cerebras, teams can identify and remediate vulnerabilities faster while reducing noise and focusing on the risks that matter most. The timing matters. Armis launched Armis Centrix™ for Application Security on February 10, 2026, positioning it as an AI-powered platform for detection, contextualization, and remediation across the software development lifecycle. In its launch materials, Armis argued that AI-assisted coding and continuous development pipelines are exposing the limits of fragmented AppSec point tools:…

Cerebras is coming to AWS March 13, 2026

The world’s fastest inference is coming to the world’s leading cloud. Today we're announcing that Amazon Web Services is deploying Cerebras CS-3 systems in AWS data centers. Available via AWS Bedrock, the new service will offer leading open-source LLMs and Amazon’s Nova models running at the industry’s highest inference speed. In addition, AWS and Cerebras are collaborating on a new disaggregated architecture that pairs AWS Trainium with Cerebras WSE to deliver 5x more high-speed token capacity in the same hardware footprint. The Need for Fast Inference AI is reshaping software development. Code is increasingly written by AI agents rather than by human developers. Unlike conversational chat, agentic coding generates approximately 15x more tokens per query and demands high-speed token output to keep developers productive. The result is an urgent and growing need for more fast inference across the industry. Cerebras…

Jais 2: A Blueprint for Sovereign AI December 09, 2025

Arabic is spoken by more than 400 million people, yet Arabic-centric Large Language Models (LLMs)still lag behind English-optimized frontier models. Building on the experience gained with the original Jais models, G42’s Inception, the Institute of Foundation Models at MBZUAI and Cerebras Systems introduce Jais-2—a new family of Arabic-centric LLMs, that represent the most capable and culturally aligned Arabic LLMs to date. Jais 2 models were trained end-to-end and deployed for production grade inference on Cerebras wafer-scale clusters, bringing frontier-level capability to models purpose-built for Arabic speaking nations. The Jais 2 chat application runs at 2,000 tokens per second, making it one of the fastest LLMs in the world. Jais 2 serves as a blueprint for sovereign AI, showing how nations can develop highly capable, culturally aligned models at lower cost, higher speed, and without the complexity of large GPU clusters.…

March 20, 2026 Why the AI Race Shifted to Speed Read blog post

For most of 2025, the AI race was about model intelligence. In the past three months, the race has shifted. Model intelligence is still critical, but across every major frontier lab, inference speed has become a new and urgent focus: - Google unveiled Gemini 3 Flash. Built for agentic coding, it runs 3x faster than Gemini 3 Pro. - Anthropic released a 2.5x-faster edition of Claude Opus 4.6 for speed-critical coding use cases. - OpenAI announced a partnership with Cerebras to release GPT-5.3-Codex-Spark, running at over 1,200 tokens/s, making it the fastest OpenAI coding model to date. Why has inference speed suddenly become so important? Because the rate at which a model generates tokens now directly affects the rate of model iteration for the major labs and the rate of building software for the broader economy. In February, both OpenAI…

The world’s fastest GLM-4.6 – now available on Cerebras November 18, 2025

Today, Cerebras is releasing GLM-4.6 — our most capable model yet on the Cerebras Inference API. GLM-4.6 brings major upgrades across reasoning, tool use, and coding, combining exceptional intelligence with an unmatched speed of 1,000 tokens per second on Cerebras. For many tasks, GLM-4.6 is comparable to Sonnet 4.5 in output while running 17x faster and 25% cheaper on Cerebras. GLM-4.6 is available today with our pay-as-you-go developer tier starting at $10 or our Cerebras Code plan starting at $50/month. GLM-4.6 GLM-4.6 is widely regarded as one of the world’s top open coding models. GLM-4.5 is ranked #1 model for tool calling on the Berkeley Function Calling Leaderboard (BFCL), ahead of Opus 4.1. GLM-4.6 performs on par with Sonnet 4.5 on LM Arena’s web-development leaderboard, based on thousands of user votes. Across real-world usage, developers highlight four defining strengths: -…

The GPU Is Being Split in Half March 26, 2026

The entire way we run AI inference is being rearchitected right now. AWS and Cerebras just announced a partnership around it. NVIDIA spent $20 billion acquiring Groq to catch up. Jensen Huang stood on stage at GTC 2026 and effectively validated what companies like Cerebras have been saying for years: general-purpose GPUs aren't enough for inference at scale. The thing they're all converging on is called disaggregated inference. And if you're a developer building anything on top of LLMs, this is going to change how fast your products feel, how much they cost to run, and what's even possible to build. Your GPU Is Doing Two Very Different Jobs When you send a prompt to an LLM, the model doesn't just "think" and return text. It runs two completely separate operations, back to back, on the same hardware. Phase 1:…

Introducing OpenAI GPT-5.3-Codex-Spark Powered by Cerebras February 12, 2026

Today, we’re announcing that OpenAI’s new GPT-5.3-Codex-Spark model, powered by Cerebras, is available in research preview. This marks the first release in our collaboration between Cerebras and OpenAI. Codex-Spark is designed for real-time software development where responsiveness matters as much as intelligence. Powered by the Cerebras Wafer-Scale Engine, it runs at over 1,000 tokens/s, enabling near-instant feedback in live coding environments. Agentic coding has fundamentally changed software development. For the first time, machines can autonomously work for hours or days without human supervision. But this mode of interaction can also leave developers feeling out of the loop with long wait times and less opportunity to direct the work. As software development is iterative, developers need to inject taste, direction, and sensibility along the way. Codex-Spark is designed for this kind of real-time, iterative work. It is fast, responsive, and steerable,…

How to stop your autoresearch loop from cheating March 19, 2026

TLDR: We let an AI agent run overnight. By morning, it had abandoned our experiment and started its own. Across 71 experiments on two very different problems: training optimization and model compression, we learned that autoresearch can reliably surface real findings when the loop is tightly scoped. Loosen the guardrails, and the agent drifts within hours. The bottleneck isn't intelligence. It's everything around it. Everything we built/ran is open-source: - codex-autoresearch-harness, Bash wrapper that forces Codex into a research loop with built-in A/B testing (Experiment 1) - reap-expert-swap, Expert pruning + dynamic swapping to fit Kimi-k2.5 in BF16 (2.5 TB) onto 8× RTX 3090s (Experiment 2) We left an AI agent running overnight on two research experiments. When we checked in the next morning, it had stopped doing what we asked. Instead of optimizing memory usage, it had gone off…

Stop Shipping AI Slop: How Codex Spark Changes The Way You Code March 04, 2026

In the past few years, we've developed series of interesting workflows. Think Ralph loops and multi-agent orchestration systems. The idea is writing very descriptive prompts and running 8-hour sessions, or having 10 instances running on your machine at all times. Most of this complexity spawned from one issue: LLMs are slow. If you prompt and wait, you'll get less done than if you prompt and move on to the next task. Spark is fast. Codex Spark changes how developers work with AI. A coding model generating 1,200+ tokens/second makes real-time collaboration possible, but it also requires a different approach. At this speed, sloppy interactions have consequences, and working with LLMs needs to be much more deliberate. This guide is a practical playbook for how we've been using GPT-5.3-Codex-Spark. Know when to use Codex vs Spark Codex now spans two complementary…

GLM-4.7: Frontier intelligence at record speed — now available on Cerebras January 08, 2026

Today, we’re announcing GLM-4.7, the latest GLM family model released from Z.ai, now available on Cerebras Inference Cloud. This model combines speed with frontier intelligence, for coding, tool-driven agents, multi-turn reasoning, and more. Frontier Intelligence GLM-4.7 is a clear step up from GLM-4.6. Against leading closed models, GLM-4.7 demonstrates comparable high-quality code generation and editing, reliable tool use, and consistent multi-turn reasoning. All at up to an order of magnitude higher speed and price-performance! On benchmarks that reflect real developer workloads, GLM-4.7 now ranks as the top open-weight model, leading DeepSeek-V3.2 across a broad set of advanced developer benchmarks, including SWEbench, τ²bench, and LiveCodeBench. Coding improvements in day-to-day development work are the most immediately visible advance from GLM-4.6 to 4.7. With more accurate solutions, cleaner structure, and stronger multilingual output, GLM-4.7 is noticeably more intelligent while stable over long, iterative…

2026: Fast Inference Finds its Groove January 06, 2026

I met my wife learning to dance Argentine tango. In tango you cannot fake your way through the steps. You have to feel the rhythm, listen to the moment, and respond in real time. Push too hard and the whole thing breaks. Find the groove and everything opens up. In 2025, AI had its own tango moment. For most of the last decade, people measured AI progress by the size of the model they could train. Bigger clusters, bigger budgets, bigger number. The ground shifted. The industry began to understand, not in theory but in practice, that inference speed is not a bragging point. It is the real constraint that determines what AI systems can do in the world. At Cerebras, we have believed this for years. We built the largest chip ever made because it was the only way…

Cerebras October 2025 Highlights November 03, 2025

October was a month of momentum for Cerebras. With new launches, global events, and groundbreaking collaborations, we continued to expand access to wafer-scale AI around the world. Try OpenAI gpt-oss-safeguard-120b at Cerebras speed Cerebras is the fastest inference provider for OpenAI's newest model. Enabling real-time reasoning about AI safety policies with full configurability and zero black-box limits. 🦺 Policy-based & transparent: Bring your policy, get explainable classifications 📖 Open-weight & configurable: Apache 2.0 license; weights freely available 🚀 Run it in real time: Moderation, document triage, and agent guardrails at wafer-scale speed Join the private preview on Cerebras Inference Cloud to experience open-weight safety AI at wafer-scale speed. The Fastest AI Inference, just $10 away With Cerebras Inference: Pay Per Token, you can start building on wafer-scale compute for as little as $10 — no contracts, no friction, no GPU…

Thinking Inside the Box: The Implicit Chain Transformer for Efficient State Tracking December 12, 2025

Dec 12 2025 Thinking Inside the Box: The Implicit Chain Transformer for Efficient State Tracking Motivation Large Language Model (LLM) decoders have demonstrated remarkable capabilities in open-ended generation, reasoning, and human-computer interaction. However, the standard autoregressive formulation suffers from a representational bottleneck: to generate the next token, the model must implicitly re-derive the underlying semantic context by attending to the entire history. This statelessness renders standard Transformers surprisingly brittle on tasks necessitating the maintenance of a running state—such as calculating the sum of a list of numbers modulo X or performing graph traversal. In this work, we introduce the Implicit Chain Transformer (ICT), a novel architecture designed to bridge this gap. By propagating a learnable "intent" latent vector forward across time steps, our method enables the model to explicitly update and contextualize a running state, rather than solely relying on…

Cerebras February 2026 Highlights November 03, 2025

- OpenAI Codex-Spark launches, powered by Cerebras - UAE and India Advance Sovereign AI Infra with Cerebras - ExomeBench: a new benchmark for clinical genomics - Café Compute takes over Boston, New York, and Seattle OpenAI Codex-Spark launches, powered by Cerebras This marks the first release in our fast-inference collaboration with OpenAI, coming just one month after announcing the partnership. Codex-Spark is built for real-time software development, where responsiveness matters as much as intelligence. It’s exceptionally fast at targeted edits, logic revisions, and frontend iteration. Powered by the Cerebras Wafer-Scale Engine, it runs at over 1,000 tokens per second, providing developers with rapid feedback and higher productivity. Rolling out now to ChatGPT Pro users. We welcome your feedback and look forward to shipping even more capable offerings this year. UAE and India Advance Sovereign AI Infrastructure At the AI Impact…

ExomeBench: A Benchmark for Clinical Variant Interpretation in Exome Regions February 23, 2026

Feb 23 2026 ExomeBench: A Benchmark for Clinical Variant Interpretation in Exome Regions 1. What is ExomeBench? We are excited to announce the public release of ExomeBench, a reproducible benchmark for clinically relevant variant interpretation in exome regions. This benchmark is designed to help researchers evaluate and improve models for health-relevant predictions, complementing existing tools and datasets in genomics. This post summarizes the benchmark tasks, baseline results, and how to get started. There has been tremendous progress in DNA and genomics modelling with transformer-based models, such as Nucleotide Transformer[1] and Evo[2,3]. These models are typically evaluated on structural and functional genomics tasks, such as predicting regulatory elements, chromatin accessibility, or other sequence-level properties, and they achieve impressive performance on these benchmarks. However, as most existing benchmarks focus on tasks related to general sequence modeling, it is unclear how well these…

Cerebras CS-3 vs. Groq LPU September 19, 2025

TL;DR The Cerebras CS-3 outperforms Groq’s LPU-based solution across almost all key metrics, delivering ~6x higher inference speeds on frontier LLMs, enabling more generation in the same amount of time, with higher accuracy and lower power consumption – at similar cost. With Cerebras, developers can build the fastest and most intelligent conversational AI, real-time code generation, instant reasoning, and agentic applications. Performance: Advantage Cerebras Today’s large language models are bottlenecked by slow GPU inference. With complex reasoning and agentic models, for example, it can take 20–30 minutes to generate an answer. The simple reason: low effective memory bandwidth. LLM generation is limited by how fast you can move model weights from memory to compute for each token. Cerebras and Groq both achieve faster LLM inference than Nvidia GPUs by addressing this memory bandwidth bottleneck. Cerebras’ wafer-scale engine stores the entire…

Cerebras CS-3 vs. Nvidia DGX B200 Blackwell September 19, 2025

Cerebras delivers the world’s fastest AI infrastructure TL;DR The Cerebras CS-3 system is 21x faster, 1/3 lower cost, and 1/3 lower power than Nvidia’s flagship DGX B200 Blackwell GPU—making previously impractical use cases a reality, including conversational AI, real-time code generation, instant reasoning, and agentic applications. Performance: Advantage Cerebras Today’s large language models are bottlenecked by slow GPU inference. With complex reasoning and agentic models, for example, it can take 20–30 minutes to generate an answer. The simple reason: low effective memory bandwidth. LLM generation is limited by how fast you can move model weights from memory to compute for each token. Cerebras keeps that traffic on-chip, with dramatically higher memory bandwidth than a GPU’s “high-bandwidth” memory (HBM) and GPU interconnect, so both output speed and end-to-end latency are significantly faster. In a third-party report by SemiAnalysis (Source), a competitive…

StackAI × Cerebras: enabling the fastest inference for enterprise AI agents January 28, 2026

Jan 28 2026 StackAI × Cerebras: enabling the fastest inference for enterprise AI agents StackAI is a low-code enterprise platform for building and deploying AI agents in regulated industries, powering workflows like compliance reviews, underwriting, and claims automation. As customers moved from simple copilots to complex, multi-step agentic workflows, StackAI needed an inference layer that could deliver sub-second latency across diverse model sizes and use cases. By integrating Cerebras, StackAI gives enterprises fast, flexible, production-grade inference—so high-stakes workflows like claims triage, compliance checks, and credit decisioning feel instantaneous. Together, StackAI and Cerebras enable real-time, scalable agentic automation across finance, healthcare, and the public sector. The Challenge StackAI supports hundreds of use cases, from document-heavy processes to real-time operational decision-making, all on one secure platform, and each is built on structured retrieval, multi-step reasoning, and integrations across dozens of enterprise systems.…

The Year of Latency Debt (And How Big Tech Is Paying It Down) January 28, 2026

I typed a single sentence into one of the world's most advanced language models: "Write a function to parse JSON out of markdown code blocks" Then I waited. The cursor blinked. I shifted in my chair. "Thinking..." I checked Instagram stories. By the time the model was done, I’d already gotten pulled into a meeting. The response was beautiful. The experience was far from ideal. And if you've been building with frontier AI models, you've probably felt this too. This is the best technology humans have ever built, and using it often feels like watching paint dry. What is ‘Latency Debt’? In software engineering, "technical debt" refers to the accumulated cost of shortcuts and slop code that works today but creates problems tomorrow. Engineers move fast, auto-accept AI suggestions, and defer the cleanup. Latency debt works the same way. Over…

Fast inference is going mainstream — the Cerebras ecosystem is scaling access January 28, 2026

Jan 28 2026 Fast inference is going mainstream — the Cerebras ecosystem is scaling access The broadband moment for AI inference Ultra‑low‑latency inference is shifting from a differentiator to a key requirement for AI-powered applications. At the same time, access through the Cerebras ecosystem is expanding across models, clouds, and developer tooling. Fast inference is no longer a niche advantage; it is becoming foundational infrastructure. As low‑latency AI experiences move from demos into daily workflows, the industry is entering a new phase where latency directly determines which applications are viable. Recent announcements across the AI ecosystem make this shift unmistakable. Ultra‑low‑latency inference is now a platform priority, not a marginal optimization. When models respond instantly, users stay engaged longer, agents can reason in tighter loops, and entirely new classes of applications become possible. Cerebras has focused on low‑latency inference well…

This new model is smarter than Sonnet 4.5…and 20X faster? January 08, 2026

So, you need speed, intelligence, and great economics… introducing GLM 4.7, the first open model that delivers all three. Why developers are switching At Cerebras, we’ve seen overwhelming demand from developers for GLM 4.7. The migration to GLM 4.7 is driven by three key factors: cost, speed, and intelligence. - Cost: GLM 4.7 is more affordable than models like Claude Sonnet 4.5, achieving high-proficiency intelligence at a fraction of the cost. - Speed: On Cerebras, GLM 4.7 achieves output speeds of up to1500+ tokens per second, making it 20x faster than closed-source competitors like Sonnet 4.5. This significantly reduces latency in agentic workflows, allowing for rapid iteration and execution in development environments. - Intelligence: GLM 4.7 is the strongest open-source coding models available today. It’s remarkably skilled at tool use, achieving 96% on 𝜏²-Bench Telecom, which makes it suitable for…

OpenAI Partners with Cerebras to Bring High-Speed Inference to the Mainstream January 14, 2026

Jan 14 2026 OpenAI Partners with Cerebras to Bring High-Speed Inference to the Mainstream OpenAI and Cerebras have signed a multi-year agreement to deploy 750 megawatts of Cerebras wafer-scale systems to serve OpenAI customers. This deployment will roll out in multiple stages beginning in 2026, making it the largest high-speed AI inference deployment in the world. This partnership was a decade in the making. OpenAI and Cerebras were both founded around the same time with radically ambitious visions for the future of AI: OpenAI set out to create the software that powers AGI while Cerebras upended conventional wisdom about chip making to build a wafer scale AI processor that defied Moore’s Law. Our teams have met frequently since 2017, sharing research, early work, and a common belief that there would come a moment when model scale and hardware architecture would…

Scaling SWE Agent Data Collection with Dockerized Environments for Execution November 24, 2025

Nov 24 2025 Scaling SWE Agent Data Collection with Dockerized Environments for Execution By :Gune S,Sahil Lathiya,APOORV PANDEY,Mritunjai Chandra,Vijay Srinivas,Ganesh Venkatesh Introduction We are focused on building a high-quality platform for agentic flow training, including support for Reinforcement Learning (RL), datasets, and ML recipes. We announced our RL platform a few weeks ago, and this project represents our dataset infrastructure initiative. Why This Matters Challenge: Training effective AI agents for software engineering requires: - High Quality Datasets: Diverse Corpus of high quality data which consist of PR title/description, issue title/description, base commit, patches, unit test files and other metadata provided by github - Reproducible Environments: Consistent execution environments across different repositories - High-Quality Signals: Clear pass/fail signals for learning (FAIL_TO_PASS tests) - Scale: Thousands of diverse, real-world software engineering tasks - Validation: Verified test outcomes with proper testing Our Solution:…

Cerebras at NeurIPS 2025: Nine Papers From Pretraining to Inference December 04, 2025

Dec 04 2025 Cerebras at NeurIPS 2025: Nine Papers From Pretraining to Inference Cerebras is excited to be at NeurIPS 2025—and what a year it's been. We launched our inference API last August, opened a new data center in Oklahoma City, and watched demand for Cerebras Inference explode with the latest state-of-the-art open weight models. Our research team has been hard at work too, and this year they're presenting nine papers probing the foundational questions of modern AI practice: where does compute get wasted during training? How should reasoning models allocate tokens at inference? When do smaller models beat bigger ones? The work spans pretraining to inference—new findings on scaling laws, training efficiency, and smarter orchestration of test-time compute. Below is an overview of each paper, what we found, and who should care. Links to the full arXiv papers are…

Router Wars: Which MoE Routing Strategy Actually Works August 04, 2025

MoE Fundamentals | Router Wars | Debugging Dead MoE Models | MoE at Scale | MoE Math Demystified Here’s what nobody tells you about Mixture-of-Experts (MoE): the router can single-handedly destroy your model. You can have perfect expert network architecture, tuned hyperparameters, and unlimited compute, but if your router collapses, you’re back to dense model performance regardless of number of experts you choose. The router’s job sounds simple – it needs to decide which expert handles each token. In practice, it’s where most MoE implementations go wrong. With wrong strategy you can spend weeks debugging and be completely lost. So which routing strategy should you use and what to expect from it? Let’s examine the most common approaches, their real-world tradeoffs, and what works in practice. The Routing Landscape: Oh So Many Flavors… Table 1: MoE routing reality. Behind the…

Debugging Dead MoE Models: A Step-by-Step Guide August 19, 2025

This video shows a complete step-by-step walkthrough of training a small MoE model and debugging router issues. If you can’t open the video displayed above, please use this link to open it on YouTube: https://youtu.be/phXUzFt7hrs?si=6FAHkA00Tvpjz4LR What We Expect to See at the End So I bet when you hear Mixture-of-Experts (MoE), you immediately think “another thing that only Google can afford to train”, right? That’s exactly the myth we want to bust today. Yes, the famous MoE models are huge - we’re talking trillion parameter scale (Kimi Team, 2025). But this is like avoiding neural networks because GPT-4 exists. You can create a perceptron network from scratch in less than 20 lines of code. Unfortunately, it is a common myth that training MoE models is not really accessible to majority of people. In fact as we were working on this…

Scaling Code-Repair Agents with Reinforcement Learning: Extending OpenHands for Real-World Repositories November 24, 2025

Nov 24 2025 Scaling Code-Repair Agents with Reinforcement Learning: Extending OpenHands for Real-World Repositories By: Harsh Gupta,Nishit Neema,Shivashish Naithani,Srinjoy Mukherjee,Sapan Shah,David Bick,Gokul Ramakrishnan,Ganesh Venkatesh Introduction Training effective code-repair agents requires more than just large language models—it demands a robust infrastructure that can efficiently interact with real-world codebases at scale. After building a comprehensive dataset of real-world issue-resolution instances packaged as Docker images, we faced a critical challenge: existing agentic frameworks like OpenHands were tightly coupled to benchmark-specific assumptions that prevented their use in large-scale reinforcement learning (RL) training loops. This post details our technical journey in transforming OpenHands from a SWE-Bench evaluation tool into a general-purpose RL training platform capable of handling thousands of diverse Python repositories. The Challenge: Breaking Free from Benchmark Constraints OpenHands provides a sophisticated agentic scaffold through its CodeActAgent, but its evaluation pipeline was architected specifically…

Rox × Cerebras: Real-time speed for agentic sales workflows November 25, 2025

Rox provides enterprise revenue agents on top of your data warehouse. The platform uses a secure knowledge graph to combine internal and external sources of data, including all the raw data from customer interactions, your CRM, product usage, and more, plus public sources of data across the web. This data is then used by AI agents to automate go-to-market workflows. These workflows include everything a seller does in their day-to-day, from account research, autofilling RFPs, meeting prep, and outreach to monitoring deal risks and moving deals through the pipeline. Users interact with Rox via web, Slack, macOS, iOS, and a conversational interface called Command. The challenge Rox orchestrates multiple AI agents specialized in high-accuracy, multi-step research. That rigor is valuable, but it can feel slow when users are on a sales call, reviewing a deal in Slack, or prepping on…

OpenAI GPT-OSS 120B Benchmarked – NVIDIA Blackwell vs. Cerebras November 06, 2025

A year ago, Cerebras launched its inference API—setting a new benchmark for AI performance. While GPU-based providers were generating 50 to 100 tokens per second, Cerebras delivered 1,000 to 3,000 tokens per second across a range of open-weight models such as Llama, Qwen, and GPT-OSS.At the time, some skeptics argued that beating NVIDIA’s Hopper-generation GPUs was one thing, but the real test would come with its next generation Blackwell GPU. Now in late 2025, cloud providers are finally rollingout GB200 Blackwell systems,it’s time to revisit the question: who’s faster in AI inference—NVIDIA or Cerebras? The Open-Weight Showdown: GPT-OSS 120B OpenAI’s GPT-OSS-120B is today’s leading open-weight model developed by a U.S. company, widely used for its strong reasoning and coding capabilities. Based on benchmarks by Artificial Analysis, most vendors today run GPT-OSS-120B in the 100 to 300 tokens per second range,…

Building Instant RL Loops with Meta Llama Tools and Cerebras October 27, 2025

Oct 27 2025 Building Instant RL Loops with Meta Llama Tools and Cerebras In this post, we’ll show how to use two open-source tools from Meta’s Llama ecosystem, Prompt-Ops and Synthetic-Data-Kit, with Cerebras Inference to build fast, RL-style workflows that optimize prompts and distill reasoning datasets in real time. Reinforcement learning (RL) is built around one powerful concept: feedback loops. An agent interacts with an environment, takes actions, receives rewards, and updates its behavior to improve over time. This idea of experiment → measure → improve isn’t limited to training new models. You can apply the same reinforcement principles at the inference layer: optimizing prompts, generating synthetic data, and iterating rapidly with measurable feedback. The faster you can complete each iteration loop of experimentation, the faster your system improves. Cerebras stands apart by building our own hardware purpose-built for serving…

The Fastest AI Datacenters will run on Cerebras: Meet OKC September 22, 2025

In the heart of Oklahoma, where determination and ingenuity have shaped communities for generations, a new chapter of innovation is unfolding. Today in Oklahoma City, I stood with our team and cut the ribbon on Cerebras’ newest AI datacenter—a facility built not just to power artificial intelligence, but to shape its future. Growing up, I often thought about my father, who was raised on a remote mining site in Australia. Life on a mine was hard—the work grueling, the distances vast—but it taught resilience, teamwork, and big dreams. His stories convinced me that transformative work doesn’t only happen in big cities. It happens wherever people are willing to roll up their sleeves and build something extraordinary together. Standing here in Oklahoma City, I feel that same spirit alive in this community. Built for Breakthroughs In 2023 we built our first…

REAP: One-Shot Pruning for Trillion-Parameter Mixture-of-Experts Models October 16, 2025

Oct 16 2025 REAP: One-Shot Pruning for Trillion-Parameter Mixture-of-Experts Models TL;DR: We introduce REAP (Router-weighted Expert Activation Pruning), a new one-shot method for compressing Mixture-of-Experts (MoE) language models. Our key finding is that for generative tasks like code generation pruning low-impact experts is fundamentally better than merging them. REAP removes up to 50% of experts from models as large as 1 trillion parameters while largely maintaining baseline model quality. For instance, with the Qwen3-480B-Coder-FP8 model, REAP at 50% pruning retains 97.6% of its baseline non-agentic coding ability and 96.7% on the agentic SWE-Bench benchmark. We are open-sourcing the complete codebase and pruned model checkpoints on HuggingFace to encourage further research. Leveraging Expert Redundancy for MoE Compression Sparsely-activated Mixture-of-Experts (SMoE) models achieve their high quality by decoupling their total parameter count from their computational cost [1]. This allows them to leverage…

Cerebras Inference: Now Available via Pay Per Token October 13, 2025

The fastest AI inference in the world is now just $10 away. Today, we’re making Cerebras Inference available to everyone through pay-per-token pricing. Start building on the world’s fastest AI infrastructure for as little as $10 — no contracts, no friction, just add your credit card and go. We believe our developer tier delivers the most compelling inference API in the industry. Run the world’s leading open weight models from Qwen3 235B Instruct and Thinking, GPT OSS 120B, and Qwen3 Coder 480B — all at 20x the speed of closed source model providers running on GPUs. Moreover, we’ve heard your calls for higher rate limits – our developer tier has over 10x higher limits than our free tier, so you can build, iterate, and scale without friction. Cerebras Code Revamped with Higher Rate Limits Our self-serve pay-per-token tier is the…

Cerebras API Certification Partner Program for LLM API Providers September 22, 2025

Introduction As the demand for scalable, low-latency large language model (LLM) inference surges, API platform providers face the challenge of delivering breakthrough performance while maintaining enterprise-grade security, governance, and ease of integration. The Cerebras API Certification Partner Program addresses these needs by enabling LLM API partners to seamlessly integrate Cerebras’s wafer-scale inference capabilities, validate their implementations against rigorous standards, and jointly bring next-generation AI services to market. Program Vision The program’s core objective is to democratize ultra-fast AI inference across the enterprise stack. By certifying API providers who meet strict performance, security, and operational criteria, Cerebras ensures that end users can rely on sub-50 ms inference at scale. Certified partners benefit from technical enablement, co-marketing, and a clear path to evolving from basic integration to full-stack, production-ready solutions. Program Overview The Cerebras API Certification Partner Program empowers LLM API platforms…

Cerebras and Docker Compose: Building Isolated AI Code Environments September 17, 2025

Developers can now run Cerebras inference inside Docker containers, deployed with Docker Compose, to create safe and reproducible environments for AI-generated code. By combining Docker’s containerization with Cerebras’ inference speed, teams can build and evaluate new code quickly while keeping experiments isolated and repeatable. Cerebras and Docker: Speed Meets Safety Cerebras runs the fastest AI inference in the world. The tool can generate code at more than 2,500 tokens per second. An easy and practical way to run Cerebras’ code is with Docker Compose. Compose simplifies running complex, multi-container applications, like agentic applications combining the agent loop, tools, and other supporting services. Run one command, and, with a single configuration file, start every service in your product without depending on specific agentic framework details. So, Compose is well equipped to allow users to deploy Cerebras agents alongside local models. Developers…

OpenAI GPT OSS 120B Runs Fastest on Cerebras August 06, 2025

OpenAI’sGPT OSS 120B model is now available on Cerebras. The first open weight reasoning model by OpenAI, OSS 120B delivers model accuracy that rivals o4-mini while running at up to 3,000 tokens per second on the Cerebras Inference Cloud. Reasoning tasks that take up to a minute to complete on GPUs finish in just one second on Cerebras. OSS 120B is available today with 131K context at $0.25 per M input tokens and $0.69 per M output tokens. GPTOSS120B is a 120 billion parameter mixture-of-expert model that delivers near parity performance with OpenAI’s popular o4mini on core reasoning benchmarks. It excels at chain of thought tasks, tackling coding, mathematical reasoning, and health related queries with class leading accuracy and efficiency. With its public weights release under Apache 2.0, it offers transparency, finetuning flexibility, and the ability to run on the…

Cerebras Launches OpenAI’s gpt-oss-120B at a Blistering 3,000 tokens/sec August 05, 2025

Cerebras is a day one launch partner for OpenAI’s new open-weight model, gpt-oss-120B, now available on Cerebras Cloud. Developers can run the model at 3000 tokens per second at full 128k context with streaming, high-throughput inference that scales from prototype to production. Cerebras makes it possible to integrate gpt-oss-120B into demanding workloads—including agentic reasoning, knowledge retrieval, and long-context generation—with ease and speed. Performance and Pricing - Throughput: 3000 tokens per second - Input: $0.25 per million tokens - Output: $0.69 per million tokens About the Model: OpenAI’s gpt-oss-120B gpt-oss-120B is OpenAI’s most capable open-weight model, released under the Apache 2.0 license. It uses a Mixture-of-Experts architecture with 117 billion total parameters, 5.1 billion active parameters per token, and a 128-expert configuration across 36 layers. The model supports a 128k context window, enabling complex multi-turn reasoning and long-form memory. The model…

Qwen3 Coder 480B is Live on Cerebras August 01, 2025

Alibaba's Qwen3 Coder 480B Instruct model is now available on Cerebras. Qwen3 Coder is one of the top coding models in the world with coding ability that rivals Claude 4 Sonnet and Gemini 2.5. Running on the Cerebras Wafer Scale Engine, Qwen3 Coder reaches an unprecedented 2,000 tokens per second. Coding problems that take 20 seconds on Sonnet 4 finish in just one second on Cerebras. To make Qwen3 Coder widely accessible, we are also launching Cerebras Code – two monthly subscription plans with generous rate limits at $50 and $200 per month. Just two weeks after launch, Alibaba’s Qwen3 Coder 480B has soared in adoption, reaching #2 in OpenRouter’s coding model leaderboard, overtaking Gemini 2.5, DeepSeek V3, Kimi K2, and Claude 4 Opus. It’s widely praised as the first model that matches Claude 4 Sonnet – the industry’s leading…

Introducing Cerebras Code August 01, 2025

We are launching two new plans designed to make AI coding faster and more accessible: Cerebras Code Pro ($50/month) and Code Max ($200/month). Both plans give you access to Qwen3-Coder, the world’s leading open-weight coding model—running at speeds of up to 2,000 tokens per second, with a 131k-token context window, no proprietary IDE lock-in, and no weekly limits! Cerebras Makes Code Generation Instant Even with the best frontier models, you still end up waiting around for completions. And as coding workflows get more agentic, the latency adds up fast. You’re not just waiting once. You have to wait on every LLM call across multi-step edits, tool use, retries, and planning. At 2,000 tokens per second, code generation becomes instant. And starting at $50/month, anyone can use Cerebras Code and enjoy fast code generation that keeps you in flow. Powered by…

From Zero to Sudoku Hero: An RL Adventure August 01, 2025

Abstract To tackle complex, real-world problems, Large Language Models (LLMs) need to learn how to reason, plan, and adapt. Our recent work on test-time scaling, CePO, demonstrated that even medium-sized models (<= 32B parameters) can outperform much larger frontier models by using adaptive planning, tool use, and self-correction [1]. We believe we can push these capabilities even further by teaching LLMs to break down challenging tasks into smaller steps, advancing when successful and backtracking when they hit a wall. This post presents our work on teaching these skills using online Reinforcement Learning (RL). Our journey begins with an ideal proxy for this kind of challenge: Sudoku. While its rules are simple, solving difficult puzzles requires significant planning and the ability to backtrack from incorrect assumptions, making it a perfect testbed for teaching an LLM the foundational skills of long-horizon reasoning.…

Qwen3 235B 2507 Instruct Now Available on Cerebras July 29, 2025

Alibaba's Qwen3 235B 2507 Instruct model is now available on Cerebras. The world’s leading non-reasoning model – Qwen3 235B Instruct runs at over 1,400 tokens per second – 11x faster than the leading GPU cloud. We serve the model with 131K context and FP8 weights from our US based data centers. Priced at $0.60 per million input tokens and $1.20 per million output tokens, Qwen3 235B 2507 on Cerebras delivers best-in-class intelligence, speed, and price-performance. Qwen3 235B2507 Instruct Following developer feedback, the Qwen team developed two separate models based on Qwen3 235B – a thinking and non-thinking version. Qwen3-235B-A22B-Instruct-2507 is the non-thinking model, achieving state-of-the-art results among non-reasoning models. It outperforms GPT-4.1, Claude Opus 4, DeepSeek V3, and Kimi K2 in the Artificial Analysis Intelligence Index – a blended score across seven benchmarks representing general knowledge, reasoning, coding, and STEM.…