LATEST
AI Tools

Cerebras Tungsten: High-Level Code for the 900,000-Core Wafer

Updated November 17, 2026
13 min read
Editor's Note: This article was human-reviewed for accuracy
Close-up of a computer server processor and RAM assembly

Quick Answer & Key Takeaways

  • • Cerebras released Tungsten on November 16, 2026, a language and compiler that lets scientists program its 900,000-core Wafer-Scale Engine in high-level code instead of hand-tuned intrinsics.
  • • It arrives with Aviary, a portfolio of scientific applications and tutorials, and a research-preview SDK that runs in a simulator with no Cerebras hardware needed.
  • • The wafer systems themselves carry no published price, and every headline speed figure in the launch comes from Cerebras.

Programming the largest chip in the world used to mean writing code that sat close to assembly. Every non-scalar program had to be wired to the hardware by hand, and only specialists could do it. On November 16, 2026, Cerebras released Tungsten, a language and compiler that aims to change that. It lets scientists program the company's 900,000-core Wafer-Scale Engine in code that reads close to the textbook form of the algorithm, and it arrives with Aviary, a portfolio of scientific applications and tutorials written in the new language. The release is aimed squarely at researchers, not at general developers, and the hardware it targets is still enterprise-only. What follows is what Tungsten does, what it costs, and where the launch claims need a second look.

What Tungsten actually is

Tungsten is a programming language and compiler for Cerebras's Wafer-Scale Engine, the wafer-sized chip that packs 900,000 cores onto a single piece of silicon. The engine is a dataflow machine: compute sits on top of local memory, the cores are joined by an on-wafer fabric, and data arriving at a core triggers the work that consumes it. Cerebras describes the cores as working like a continuum rather than as separate processors, which is what lets a single wafer hold a whole simulation instead of splitting it across thousands of separate chips.

The release has three parts. Tungsten itself defines what each processing tile computes and the data it sends or receives. Paint, a companion language, describes where those programs run and how the tiles connect across the wafer. Aviary is the application portfolio: stencil codes, particle transport, sparse solvers, fast Fourier transforms, molecular dynamics, plasma simulation and high-order hydrodynamics, plus tutorials that teach the language step by step. Everything is distributed through Cerebras Labs as the Labs SDK, which is labeled a research preview.

That framing matters. Tungsten is not a general-purpose coding tool and it is not a chatbot wrapper. It is a systems programming release for scientific computing, the kind of software that national laboratories run on machines most readers will never touch. Its audience is researchers who simulate plasmas, weather, materials and nuclear processes, and the people building AI models that use those simulations as training data.

The problem Tungsten solves

Before Tungsten, programming the Wafer-Scale Engine meant working through the machine's low-level features directly. Earlier compilers exposed the vector, stream and event hardware as intrinsics, so scalar code was fine but anything more complex had to call into the hardware explicitly. Cerebras's own account of that era is blunt: developers configured descriptors by hand, wired tasks to events by hand, and routed the fabric by hand. The company says it wrote a lot of code that way itself, and that the code was fast but took deep expertise and was hard to read, change or reuse.

The practical effect was a bottleneck. The hardware could run simulations hundreds of times faster than a conventional supercomputer on some problems, but only a small group of specialists could write the programs. A physicist with a new simulation idea first had to become a wafer programmer, or find one. Tungsten is Cerebras's attempt to remove that step, so the person who understands the science can write the code without first learning the machine's internals.

This is a familiar pattern in computing. Compilers stopped asking programmers to allocate CPU registers decades ago, and Tungsten applies the same idea to the wafer's more complex resources. The compiler carries a full model of the architecture and decides how access patterns, loop definitions, fabric sockets and microthreads are assigned across the whole program, swapping them in and out as the program runs. When a program needs restructuring later, the change is an edit to high-level code rather than a rewrite of hand-wired hardware calls.

How the compiler works

Tungsten's notation is built to read like mathematics. In the Jacobi stencil example from the tutorial, each line of code is a loop over a whole block of data, and each line compiles to a single vector instruction. The operator <+- accumulates into a destination, working like += in a conventional language. There is no descriptor setup and no task wiring in the source, only the stencil itself. The Gray-Scott reaction-diffusion tutorial is about 50 lines of Tungsten, which gives new users a small entry point before they tackle the full scientific codes.

Under the surface, the compiler maps the code onto access-pattern registers that walk multidimensional arrays and fabric streams, so address arithmetic costs no instructions. A single instruction can take a whole loop as its operand. Cores call functions on each other by sending messages across the fabric, and event-based control flow replaces polling and branching. Because the hardware carries the loop and control overhead, a core runs near peak even on short inner loops, and spreading a problem thin across the wafer keeps getting faster past the point where other machines stall.

Paint handles the other half of the job. It lets developers compose layouts from individual tiles and rectangular regions, bind program sockets to communication channels, and specify routes through the on-wafer fabric. Separating the computation from its layout means a program can be resized or rearranged without rewriting the numerical kernel. The two languages compile into a machine image rather than generating Cerebras Software Language source, so the output is a runnable wafer program, not another layer of code to maintain.

For comparison, the open inference world moved in a similar direction this year: our SGLang Consortium coverage describes how NVIDIA, AMD, Intel and Google united behind a shared inference runtime, and DeepSeek's DeepInfer opened up its production inference engine. Tungsten is the scientific-computing version of the same idea: move the hard machine knowledge into the compiler so more people can use the hardware.

Aviary: the applications and the numbers

Aviary is where Cerebras makes its performance case, and every number in this section comes from the company. There is no independent benchmark of Tungsten yet, and the figures below should be read as Cerebras's own claims about its own hardware.

ApplicationWhat Cerebras reportsContext
SNAP (nuclear radiation transport proxy)1.6 PFLOP/s on a single CS-4Cerebras says this is ahead of 10,800 GPUs on the El Capitan supercomputer
HPCG (sparse-workload supercomputing benchmark)1.04 PFLOP/s on a single systemHPCG is the benchmark used to rank supercomputers on real-world sparse workloads
Wafer-scale molecular dynamics748 times faster than FrontierFrontier is one of the world's largest supercomputers; the comparison is from a 2025 Journal of Chemical Physics paper
Particle-in-cell plasma simulation60 million particles at 7,200 time steps per second430 billion particle updates per second on one system, per Cerebras
Global tsunami simulationRan on 64 wafersCerebras says such deployments scale to 2,048 systems

The applications grew out of Cerebras's work with Sandia, Lawrence Livermore and Los Alamos national laboratories under the NNSA's Advanced Memory Technology program. The molecular dynamics codes build on the collaboration's 2024 Gordon Bell Prize finalist work, though Cerebras notes the Aviary codes are not the same code as that result, and that the earlier result already used an early version of Tungsten.

Two caveats belong next to the table. First, the figures compare one wafer-scale system against whole supercomputers, which is the comparison Cerebras wants you to see, but it is a vendor comparison with no third-party verification published so far. Second, the numbers describe runs on physical wafer hardware, while the Labs SDK most readers can access only simulates the hardware. Simulator runs check that a program behaves correctly; they do not establish measured throughput on a physical wafer.

The agent-written plasma simulation

The launch's most talked-about detail is not a benchmark but a workflow. A plasma physics simulation, an electron beam entering an argon plasma, ran across 98 percent of a Cerebras wafer, and a coding agent wrote every line of it. One researcher and one coding agent built the particle-in-cell code in about two weeks through a chat window. The agent worked from published algorithms, validated its reference implementation against the open-source XOOPIC code, modeled the problem on the EMPIRE plasma code, and produced the tests, figures and documentation while finding and fixing its own bugs.

Cerebras says the full run advanced 60 million particles at 7,200 time steps per second, or 430 billion particle updates per second on a single system. The company presents this as evidence that the Tungsten reference is complete enough for agents to work with, so a researcher can try the language with almost no startup cost. Michael James, Cerebras's co-founder and chief architect of advanced technologies, still tells users to learn the language themselves. His line in the announcement is worth quoting directly: "Someone has to keep the agents on track."

The point is easy to overstate. An agent wrote a simulation for hardware that only a handful of institutions own, and the result was validated against established codes, which is exactly the kind of checking that makes the story credible rather than a demo. But it is one application, written for one machine, and it does not show that agents can write arbitrary scientific codes on Tungsten. The honest reading is narrower: for a well-specified physics problem with published algorithms and a reference implementation to check against, an agent plus a researcher can now produce wafer code in weeks instead of months.

What it costs

There is no published price for Tungsten, for the Aviary portfolio, or for the Labs SDK. Cerebras distributes the SDK as a research preview and researchers request access through Cerebras Labs, rather than downloading it freely. Cerebras has not announced an open-source licence for Tungsten, so do not treat it as open-source software. The wafer hardware itself, the CS-3 and CS-4 systems, is sold under enterprise contracts with no public price, which is standard for machines of this size.

That leaves the question most readers actually have: what does it cost to touch Cerebras at all? The company's separate inference product, Cerebras Inference, does have a public rate card, and it gives a sense of the company's pricing posture. The official pricing page lists gpt-oss-120b at $0.35 per million input tokens and $0.75 per million output tokens, and Qwen 3.8 27B at $0.99 per million input tokens and $1.49 per million output tokens. At November 17, 2026 market rates (1 USD = 0.758 GBP, 1.43 CAD), that is about £0.27 and C$0.50 per million input tokens for gpt-oss-120b, rising to about £1.13 and C$2.13 per million output tokens for Qwen 3.8 27B. Those rates are approximate and cover the inference API only. They are not Tungsten prices, and they buy nothing on a wafer.

For hardware context, NVIDIA's DGX Spark, a desktop-scale AI system, went on sale at $4,999 for the 64GB version, as covered in our DGX Spark pricing breakdown. A wafer-scale system sits in a different category entirely: it is a data-center installation sold to national laboratories and large AI teams, not a desktop purchase, and its price is negotiated rather than listed.

How to try it

The realistic route for almost everyone is the simulator. Researchers can request the Labs SDK from Cerebras Labs, then build and simulate the examples without access to a physical Wafer-Scale Engine. The SDK includes the compiler, the simulator, the tutorials and the full application portfolio. Aviary's application pages include source files, explanations of the algorithms and data exchange, and tests that build and simulate the programs and compare their results with NumPy references.

The development tools include Vim and Visual Studio Code integrations for Tungsten and Paint, with syntax highlighting and Unicode-symbol completion in VS Code, plus build tools, pipeline traces, memory inspection and per-tile debugging. That tooling is aimed at the same two audiences: scientists iterating on a simulation locally, and coding agents iterating before a workload moves onto Cerebras hardware.

Start with the Gray-Scott tutorial if you are new to the language. At about 50 lines it is the smallest complete example, and it produces the reaction-diffusion patterns Cerebras shows in its launch material. Then read one full application page end to end before writing your own code, because the pages explain how data moves between tiles, which is the part of wafer programming that has no analogue in ordinary code.

Five things to know before you build on it

First, the Labs SDK is a research preview, not a finished product. Access is by request, the label signals that interfaces and behavior can change, and there is no published timeline for a general release. Treat anything you build on it as exploratory until Cerebras says otherwise.

Second, simulator results are not hardware results. A program that runs correctly in the simulator has passed a behavior check, not a performance test. Do not quote simulator runs as speed figures, and do not compare them with the PFLOP/s numbers in the launch material.

Third, every speed figure in this article comes from Cerebras. SNAP at 1.6 PFLOP/s, HPCG at 1.04 PFLOP/s and molecular dynamics at 748 times Frontier are the company's own measurements on its own hardware, with no independent benchmark published yet. They are a reason to pay attention, not a verdict.

Fourth, the design optimizer is not part of the release. Cerebras describes it as coming soon to Aviary: an AI model coupled to a physics simulation acting as its world model, with both running on the wafer, so the model can test designs against real physics. The company is showing it at its SC26 booth. Until it ships, the agentic-design story is a preview, not a product.

Fifth, hardware access is the real gate. Tungsten lowers the programming barrier, but the machine it programs is still a wafer-scale system sold to laboratories and large teams. For most readers the simulator is the whole experience, and that is fine: the language, the tutorials and the application source are the parts of this release that are actually reachable today.

Is Cerebras Tungsten free?

There is no published price, and access is by request through Cerebras Labs as a research preview. Cerebras has not announced an open-source licence for Tungsten. The wafer hardware it targets is sold under enterprise contracts with no public price.

Is Tungsten open source?

Not as announced. Cerebras describes the release as a Labs SDK in research preview, available by request. No open-source licence has been announced for Tungsten or Paint, so treat the code as proprietary research software until the company says otherwise.

Can I run Tungsten without a Cerebras wafer?

Yes. The Labs SDK includes a simulator, so researchers can build and run the tutorials and applications without physical hardware. Simulator runs validate behavior, not speed.

How fast is Tungsten really?

That is not yet independently known. Cerebras reports SNAP at 1.6 PFLOP/s on one CS-4, HPCG at 1.04 PFLOP/s, and molecular dynamics 748 times faster than Frontier, but all three figures are company-reported and no third-party benchmark has been published.

What is Paint?

Paint is Tungsten's companion language. Tungsten defines what each tile computes; Paint describes where those programs run and how the tiles connect across the wafer. Together they compile into a machine image for the Wafer-Scale Engine.

Cerebras Tungsten is a serious release for a narrow audience. It removes the hand-wired programming that kept wafer-scale computing in the hands of specialists, and it pairs that with a portfolio of real scientific codes and a simulator anyone in research can request. The speed claims are all Cerebras's own, the design optimizer is still to come, and the hardware remains out of reach for most. But the direction is clear: the company wants scientists and their coding agents writing wafer code, and Tungsten is the first version of that bet. Request the SDK from Cerebras Labs and start with the Gray-Scott tutorial if you want to see what high-level wafer code looks like.

Sources

13 min read
109 views
edit_note Written by

Ratul Ahamed

Founder & Editor-in-Chief

Ratul Ahamed is the founder and editor of AI Core Updates, an independent site covering AI tools, models and pricing. He is a web developer based in Bangladesh and also runs All In One Tools, a free collection of browser-based utilities. Contact: [email protected]

Read all stories by Ratul Ahamed arrow_forward
verified_user Human-Reviewed Editorial

forum Comments (0)

Leave a Reply

chat_bubble_outline

No comments yet. Be the first to join the conversation!

Weekly AI Digest

Get our weekly AI tools digest

One email a week analyzing new AI tools, model updates, and practical workflows. Tested and summarized with zero fluff. Unsubscribe any time.

bolt
Hands-on AI Tool Reviews
Pricing math, limits, and alternatives.
verified_user
Strict Anti-Spam Policy
1-click unsubscribe at any time.
mark_email_unread

Subscribe to Digest

Sent weekly • No spam

person_outline
mail_outline

No spam ever. You can unsubscribe with a single click at any time.

cookie

Cookie Settings

We use cookies to enhance your browsing experience, serve personalized content, and analyze our traffic. By clicking "Accept All", you consent to our use of cookies. Learn more