Synoema on IoT

What this is

Synoema is a programming language designed for LLM code generation. The IoT track brings that design to the edge: a small model writes a Synoema rule, the toolchain compiles it to .wasm, and the artifact runs on a Raspberry Pi, an STM32, or an nRF5340 wearable — with the compiler proving it meets the contract the prompt declared.

This page reports what is measured as of 2026-04-18. No marketing claims, no roadmap teases. Numbers below all trace back to context/METRICS.md in the repository.

Three deployment tiers

Synoema programs can land on an IoT device three ways, ordered by "like a normal computer":

  Tier 2                       Tier 1                        Tier 0
  ──────                       ──────                        ──────
  Linux IoT                    Bare-MCU via wasm3            Native AOT to MCU
  (shipped)                    (shipped)                     (deferred)

  Synoema → Cranelift →        Synoema → WASM v2 →           Synoema → Cranelift →
  AArch64 / x64 ELF →          wasm3 interpreter on MCU →    Thumb2 ELF → flash
  scp or flash onto            firmware links wasm3          directly, no interpreter
  the device                   and loads the .wasm module

  Targets:                     Targets:                      Status:
  Raspberry Pi (aarch64)       Cortex-M0/M3/M4/M7            Blocked on Cranelift 0.104
  Jetson                       RP2040                        lacking an ARM32 backend.
  BeagleBone                   ESP32                         When upstream merges it,
  x86 edge gateway             any RTOS or bare metal        ~200 LOC to integrate.

Tier 2 is the native AOT path: sno build --native produces an ELF. Tier 1 is the WASM path: sno wasm produces a module, and a thin C host links wasm3 to execute it inside ~64 KiB flash and ~16–32 KiB RAM. Tier 0 is real bare-metal Thumb2 — we ship the CLI target name and the static preflight, but the backend is deferred pending upstream.

Six vertical MVPs, measured

Each vertical is 5 rules compiled end-to-end from a Synoema source. 30 rules total, 30/30 sno check pass rate, 29/30 WASM, 100% GBNF validation. Byte counts are from research/finetune/eval/results/2026-04-18-iot-platform-v1/wasm_sizes.json.

Measured

All numbers below are authoritative — they trace to context/METRICS.md and are regenerated by scripts/metrics.sh.

Try it yourself

The cloud_compile pipeline ingests an English prompt, runs it through a small fine-tuned Synoema model, compiles the result to .wasm, and returns the bytes + metadata. Mock mode uses deterministic fixtures so you can inspect the round trip without a model server running.

python3 lang/tools/llm/cloud_compile.py \
  --mock \
  --prompt "Trigger pump alarm when pressure exceeds 12.0 bar for 3 seconds"

Output is a JSON envelope: source, typechecker result, WASM byte count, contract summary. Swap --mock for --model <name> once you wire a local model with MCP.

Reference documentation for each tier:

• Tier 2 Linux AOT — docs/iot/linux-deploy.md (scp + systemd)
• Tier 1 wasm3 embedding — docs/iot/wasm3-embed.md (C host API)
• Bare-MCU host stub — docs/iot/bare-mcu-host.md
• Architecture — docs/iot/bare-mcu-architecture.md

What is not claimed

The IoT track is built honestly. These items are deferred and are not currently shipped:

• Tier 0 native AOT to Thumb2 (ARMv7-M). Blocked on Cranelift 0.104 not having an ARM32 backend. The CLI target name thumbv7m-bare, the heap-allocation preflight, and the SNO-AOT-TARGET-UNSUPPORTED diagnostic all ship today; the emitter is the ~200 LOC that lands once upstream merges.
• Real wasm3-on-MCU execution in CI. The build path is validated end-to-end, but we do not currently spin physical hardware inside the test gate.
• Three additional verticals (logistics, smart-grid, medical/regulated) are queued — Home, Industrial, Wearable, Automotive, Agriculture and Healthcare are measured today.
• Fresh cross-language competitor benchmarks. The comparison table in the eval bundle marks competitor cells n/a (no public data) where no credible source exists; we do not invent numbers.

When the upstream ARM32 backend lands, Stage 2c flips from "CLI surface only" to a working Tier 0. Track the trigger in openspec/changes/iot-runtime-bare-mcu/roadmap.md.

Pre-trained IoT models on HuggingFace

Three fine-tuned models for Synoema IoT rule generation — ready to use with llama.cpp, Ollama, or the transformers library. All models verified on 127-example unseen test set.

General Synoema coder model: delimitter/synoema-coder-3b-v2 — 72.1% run_pass on general Synoema programs.

Why Synoema for IoT