MAX — a math-native language for neural runtimes.

MAX is a programming language whose syntax is mathematics — not as a library but as the language itself. Definitions are written with ≡, returns with ↩, ranges with ⟳ i ∈ 0..N, conditions with ⊃ / ⊅, matrix multiply with @, Hadamard product with ⊙, Frobenius norm with ‖·‖_F.

It is designed for the workloads the lab actually runs — neural-network training and inference, ML inference kernels, matrix-state dynamics, agent runtimes, OS-level substrates. The compiler emits C, compiled by a standard C compiler, and CUDA or HIP source, compiled by NVIDIA's or AMD's compiler, with cross paths for Apple Metal, ARM and RISC-V bare-metal, and Verilog accelerators for a Zynq-7020 FPGA, verified in simulation.

It is also small. The bare-metal MAX-OS kernel image for the Raspberry Pi is a tiny fraction of a typical Python + PyTorch + CUDA-runtime install, which runs into multiple gigabytes before a single model loads. There is no interpreter and no package manager; MAX programs link the C library and, for GPU work, the vendor runtime.

Mainstream stacks bury the math inside library calls. The math the lab works with — matrix-state evolution, Genesis dynamics, neural ODE / closed-form continuous-time networks, oscillator coupling — is far cleaner when the operators are part of the language. MAX programs read like the paper they came from.

Owning the compiler also lets the lab control reproducibility: the compiler itself is a byte-identical fixed point, checked in CI.

MAX compiles itself in three stages: seed → stage 1 → stage 2 → stage 3. The compiled outputs of stage 2 and stage 3 are byte-identical (sha-256 equality is checked in CI). That fixed-point invariant means any code-generation change that breaks it is rejected.

The standard library covers the math the lab leans on — DCT, RSVD, dense linear algebra, statistics and probability, BPE / SentencePiece tokenization, autograd primitives, optimizer steps (Adam / SGD), oscillator dynamics, closed-form continuous-time (CfC) networks, safetensors I/O, GGUF loader, paged KV-cache, FlashAttention-2, RoPE, RMSNorm, SwiGLU, quantization (Q3/Q4/Q5/Q6 K-block, W4A16), and cryptography builtins implemented in C in the MAX runtime (TLS 1.3, Ed25519, ChaCha20-Poly1305, Argon2id, BLAKE2b). Every module header cites the source paper or internal theorem it realizes, and the language keeps its own theorem registry.

A matrix multiply written in MAX is compiled by MAX to CUDA C. In our tests, built with NVIDIA's nvcc for a Jetson Orin GPU and run under Linux, it matched a CPU reference bit-exactly (see the hardware silicon page). No cuBLAS or cuDNN is in the path, but NVIDIA's compiler, the CUDA runtime and the Linux GPU driver are. The same principle carries to the FPGA work, where a hand-written RTL accelerator is verified in simulation to be byte-identical to its MAX/C software counterpart before it is trusted to replace it in the fabric.

• Qwen3-32B from-scratch GPU forward passwith a full GPU backward pass; in our tests on the lab's local NVIDIA GB10 the layer backward matched a CPU reference (rel-L2 < 5e-6).
• Layer-streaming forward — runs a 64-layer 32B-parameter model by holding one layer at a time (≈4.4 GB peak RSS in our tests).
• Native HTTP server for chat APIs (OpenAI-compatible), no Python in the serving path.
• Falcon-H1-7B (base model) with a confidence trigger — the Sadalsuud detector — designed to fetch from Wikipedia or refuse when confidence is low.
• OS substrate work — MAX OS, with userspace, ext2-rw and an EL0 loader in its QEMU harness, designed to run bare-metal on real boards.

The compiler engineering — syntax, bootstrap, backends, stdlib API — is openly documented. The math behind specific theorems wired into the language (Genesis dynamics, Sadalsuud routing, substrate decompositions) stays inside the lab. Engineering partners receive the engineering. Co-development partners receive more under NDA.