AI Models & Platforms
Cerebras Releases Tungsten for High-Level Wafer-Scale Programming

Key PointsTungsten shifts wafer resource allocation from manual intrinsics to compiler-managed…
View more
- Tungsten shifts wafer resource allocation from manual intrinsics to compiler-managed cost models and optimization passes.
- Separating computation in Tungsten from wafer layout in Paint lets developers resize deployments without rewriting numerical kernels.
- A coding agent wrote a plasma simulation that ran across 98% of a Cerebras wafer in about two weeks.
Cerebras Systems released Tungsten, a high-level language and compiler for its Wafer-Scale Engine, on November 16, 2026, alongside Aviary, a portfolio of scientific applications and tutorials. The announcement says a plasma simulation written entirely by a coding agent has already run across 98% of a Cerebras wafer.
The release targets the programming burden rather than a change to the processor itself. Earlier compilers exposed non-scalar vector, stream and event capabilities through intrinsics, requiring developers to configure descriptors, connect tasks to events and route the fabric manually. Cerebras’s objective is to let scientists program the wafer at a high level without becoming hardware specialists or giving up performance.
Compiler Manages Wafer-Level Resources
Cerebras describes the Wafer-Scale Engine as a dataflow machine with 900,000 cores, local memory and an on-wafer communications fabric; arriving data triggers the work that consumes it. Access-pattern registers walk multidimensional arrays and fabric streams, so address arithmetic consumes no instructions. A single instruction can take an entire loop as its operand, cores can call functions on one another through fabric messages, and event-based control flow replaces polling and branching.
Tungsten carries an architectural model that maps high-level code onto access-pattern registers, event-driven tasks, fabric calls and loop-bearing instructions. In the announcement’s Jacobi stencil example, each source line represents a loop over an entire block and compiles to one vector instruction, without descriptor setup or task wiring. The notation uses <+- to accumulate into a destination, similar to +=.
The compiler also allocates access patterns, loop definitions, fabric sockets and microthreads across the whole program, using cost models and optimization passes to swap those resources in and out during execution. Cerebras compares the approach with CPU register allocation and spilling. With those decisions owned by the compiler, later restructuring becomes an edit to the high-level code rather than a manual rewrite.
Paint Separates Computation From Wafer Layout
The Cerebras Labs documentation describes Tungsten as one half of a two-language programming model. Tungsten defines what each processing tile computes and the data it sends or receives. Its companion language, Paint, describes where those programs run and how the tiles connect across the wafer.
Paint 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 allows a developer to resize or rearrange a program’s deployment without rewriting the numerical kernel. The languages compile into a machine image rather than generating Cerebras Software Language (CSL) source.
Aviary Spans Tutorials and Scientific Applications
Tungsten and Aviary are available through Cerebras Labs. The application portfolio ranges from compact heat-equation and Gray–Scott reaction-diffusion stencils to particle transport, sparse solves, fast Fourier transforms, molecular dynamics, plasma simulation and high-order hydrodynamics. Its Gray–Scott tutorial is about 50 lines of Tungsten, giving new users a smaller entry point alongside the full scientific codes. The molecular-dynamics applications cover short-range Lennard–Jones interactions and tabulated embedded-atom potentials for metal alloys.
Cerebras reports that SNAP, a standard proxy for deterministic nuclear radiation transport, reaches 1.6 PFLOP/s on one CS-4, ahead of 10,800 GPUs on the El Capitan supercomputer, and that HPCG, a sparse-workload supercomputing benchmark, reaches 1.04 PFLOP/s on a single system.
Many of the applications grew out of work with Sandia, Lawrence Livermore and Los Alamos national laboratories under NNSA’s Advanced Memory Technology program. The molecular-dynamics codes build on the collaboration’s 2024 Gordon Bell Prize finalist work, but are separate codes; that earlier result used an early version of Tungsten.
Researchers Can Start in a Simulator Without Cerebras Hardware
Cerebras Labs packages the compiler, simulator, tutorial, and application portfolio as the Labs SDK, currently labeled a research preview. Researchers can request the SDK and build and simulate the examples without access to a physical Wafer-Scale Engine.
Aviary’s application pages include source files, explanations of algorithms and data exchange, and tests that build and simulate the programs and compare their results with NumPy references. This makes the portfolio a starting point for understanding and adapting scientific codes, rather than simply a collection of performance demonstrations. Simulator runs help check program behavior; they do not establish measured throughput on a physical wafer.
The development tools include Vim and Visual Studio Code integrations for Tungsten and Paint, with syntax highlighting and Unicode-symbol completion in VS Code. Build tools, pipeline traces, memory inspection, and per-tile debugging help developers investigate incorrect results, communication problems, and execution behavior. Together, these tools give scientists and coding agents a way to iterate locally before moving a workload onto Cerebras hardware.
Agent-Written Plasma Code Demonstrates the Workflow
The particle-in-cell application provides the release’s central agentic-coding example. Cerebras says one researcher and one coding agent produced it in about two weeks through a chat interface, with the agent writing every line. 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 tests, figures and documentation while finding and fixing its own bugs.
Particles in the simulation move under electromagnetic fields and migrate between cores, while an implicit solver computes fields on a decomposed mesh from their positions and velocities. The modeled problem is an electron beam entering an argon plasma. Cerebras reports that the full run advanced 60 million particles at 7,200 time steps per second, with throughput of 430 billion particle updates per second on one system.
Cerebras says the Tungsten reference is complete enough to support this agent-led workflow. Michael James, Cerebras co-founder and chief architect of advanced technologies, still encourages users to learn the language themselves: “Someone has to keep the agents on track.”
Cerebras says the same Tungsten code runs on one wafer or across a cluster; its global tsunami simulation used 64 wafers within an AI deployment, and such deployments scale to 2,048 systems.
Cerebras lists a design optimizer as forthcoming rather than part of the released portfolio. It would couple an AI model to a physics simulation acting as its world model, with both running on the wafer, allowing the AI model to test designs against that simulation.
Tungsten’s practical significance is the programming model around the reported throughput. The key test is whether researchers and agents can move from Aviary’s examples to their own codes without returning to intrinsics and hand-routed fabric. That capability would determine how broadly Cerebras’s specialized hardware can support scientific computing and AI-guided design.












