Local-first AI inference

One Rust GPU core.
Everywhere.

Run large language models on-device with a single Rust core compiled for macOS, iOS, Android, Linux, Windows, and Web. Exposes a C ABI to a full TypeScript agent framework.

bash — combs-cli

The Stack

Compute lives exclusively in L0 — every layer above is thin orchestration over the same highly-optimized engine.

L3 — Platform Shells

UI / Wrappers

Svelte 5 UI template · Android (JNI/Kotlin) · iOS (Swift) · Web Worker transport

L2 — Deno / TypeScript Framework

Engine/Js

@combs/core · @combs/graph · @combs/agents · @combs/runtime · @combs/flows

L1 — C ABI / JSON-FFI

combs-ffi

combs_engine_create · combs_chat_completion_stream · OpenAI-compatible server

L0 — Rust GPU Core

Engine/Core

Burn 0.21 · CubeCL · wgpu (Metal/Vulkan/DX12/WebGPU) · Paged KV cache · full sampler · GGUF/SafeTensors

Memory & Execution

Engineered for minimal overhead. Features a custom paged KV cache (page-16 arenas, LIFO free-list, prefix-safe popn), chunked prefill, and quantized linear layers with weights packed directly in VRAM.

KV Cache Chunked Prefill Metal/Vulkan/DX12

Model Source Adapters

Llama-family architectures validated with SmolLM2. Load SafeTensors (HF layout) or GGUF v3 directly via a single ModelSource adapter trait. combs run --model anything.gguf just works.

  • GGUF — F32 / F16 / Q8_0 / Q4_0
  • SafeTensors (HF layout)
  • 🚧 GGUF — K-quants (Planned)

L2 TypeScript Framework

A complete Deno ecosystem built on top of the Combs C ABI, abstracting infrastructure while keeping native performance.

@combs/core

Combs.init("smollm2-135m") — presets, model cache, device planner, FFI / Remote / Worker engines.

@combs/graph

LangGraph-equivalent: StateGraph, channels, Pregel runner, checkpointers (Deno KV/SQLite), human-in-the-loop.

@combs/agents

Tools & ToolNode, ReAct agents, structured output, memory, MCP client, skills loader.

@combs/runtime

Agent HTTP/WS servers, orchestrator, agent pool, KV task queue, session stores.

@combs/flows

createWorkflow (steps/loops/checks), createRoleplayChat, addMemory.

@combs/telemetry

Scoped logging, OpenTelemetry-shaped spans, metrics observability.

UI Scaffolding via combs chew

Interactive CLI tool that stamps out a ready-to-build Svelte 5 app complete with Ed25519 auth, fine-grained network/storage permissions, and realtime network monitors.

$ combs chew chat-ui my-app --yes

Performance

Measured on Apple Silicon (wgpu → Metal) running SmolLM2-135M-Instruct.

Decode
~35 tok/s
Prefill (1261-tok)
~660 tok/s
TTFT
~200 ms
FFI Streaming
~31 tok/s

Platform Matrix

cargo xtask bundle produces dist artifacts for every target.

Target Artifact Status
macOS (arm64 / x86_64) dylib ✓ Built & tested
iOS (arm64) .a ✓ Builds
Android (arm64) .so ✓ Builds (NDK)
Linux / Windows ✓ cargo check clean
Web (wasm32) 🚧 skeleton