# tokspur

> tokspur is a free local inference engine for coding and chat on Apple M5 Max. It runs the full Qwen3.8 Flash-Next model on your Mac and reuses earlier context between turns, 2.9× faster than llama.cpp and 3.1× faster than MLX-VLM in matched tests, with no per-token fee. The target machine has a 40-core GPU and 128 GB of unified memory.

tokspur is written in C++, Objective-C++ and Metal with a Rust host, and built with Apple's toolchain. Chat apps and coding agents in an editor or terminal call its OpenAI-compatible Chat Completions API, served on the loopback interface at http://127.0.0.1:18081/v1 by default. Each decoding step restores earlier context from saved state, lets a draft propose a short run of tokens, checks the whole run with the main model at once and keeps only what the main model agrees with.

The model is the full, unpruned Qwen3.8 Flash-Next with all 512 experts. The runtime budget is 96 GB for weights, caches and state, leaving 32 GB outside that budget for macOS, your editor and builds. The runtime carries the Apache 2.0 licence. Model weights have separate terms.

Source code: https://github.com/san-digital/tokspur (Apache 2.0). Signed macOS releases: https://github.com/san-digital/tokspur/releases/latest.

Results and specifications, measured on an M5 Max with cached prompts.

- Model parameters: 177B. Qwen3.8 Flash-Next, UD-Q4_K_XL. Rounded parameter count.
- Context: 65,536. Shared by input and output.
- Maximum output: 32,768. Per response, within the shared context limit.
- Peak speed: 148 tokens/sec. Complete request for code that passed type checking.
- Response time: 7.91 s. Time to receive the complete code.
- Speed vs llama.cpp: 2.9×. Identical code on 2 tests. llama.cpp without a drafter.
- Speed vs MLX-VLM: 3.1×. Same prompts and output length.
- Weight repacking: 0 weights changed. Packed weights match the originals byte for byte.

The comparison page also records 2 completed coding repairs.

## Pages

- [Fast, private AI coding](https://tokspur.com/): The full 177B Qwen3.8 Flash-Next on your Mac at up to 148 tokens/sec, 2.9× faster than llama.cpp. Private chat and coding with no token bills. Free, Apache 2.0.
- [Get started](https://tokspur.com/get-started/): Install tokspur on your Mac, adopt the model files, start the local server and send a first request from a chat app or coding agent.
- [The engine](https://tokspur.com/engine/): How tokspur runs natively on Apple M5 Max, reuses context between turns, repacks weights losslessly, verifies proposed tokens and offers an opt-in speed mode.
- [Documentation](https://tokspur.com/docs/): Find tokspur setup, model, API, configuration and troubleshooting guides.
- [Requirements](https://tokspur.com/docs/requirements/): The Apple Silicon hardware target, memory floor, system preparation and storage needs for tokspur.
- [Models](https://tokspur.com/docs/models/): The pinned Qwen3.8 Flash-Next build and draft, file verification, context limits and the separate model licences for a tokspur installation.
- [Integrations](https://tokspur.com/docs/integrations/): Connect a chat app, your editor or a coding agent to tokspur, with direct API settings and local gateway instructions for Claude Code and Codex.
- [Editor setup](https://tokspur.com/docs/integrations/editors/): Use tokspur in VS Code, JetBrains, Zed, Neovim and Emacs, or connect through Cline, Continue, Roo Code and Kilo Code.
- [Terminal agents](https://tokspur.com/docs/integrations/agents/): Connect OpenCode directly to tokspur, or run Claude Code and Codex through a local API gateway.
- [Local gateway](https://tokspur.com/docs/integrations/gateway/): Translate Messages and Responses requests to tokspur’s Chat Completions API with an optional local LiteLLM gateway.
- [Local API](https://tokspur.com/docs/api/): The tokspur local API: endpoint, chat completions, streaming, tool calls, authentication, cancellation and error responses.
- [Configuration](https://tokspur.com/docs/configuration/): The tokspur profile, private state paths, memory and context budgets, cached state and local server operation.
- [Troubleshooting](https://tokspur.com/docs/troubleshooting/): What each tokspur refusal means, from setup and model checks to memory, connection, output and shutdown problems, with safe diagnostics.
- [Results](https://tokspur.com/results/): Recorded speed, model capacity and coding results.
- [Code at speed](https://tokspur.com/results/comparison/): Spend less time waiting for code. Tokspur runs the full Qwen3.8 Flash-Next model locally, with no token bills.
- [Measurement method](https://tokspur.com/results/methodology/): How tokspur measures matched coding tasks: correctness grading, timing boundaries, repeated runs and reproducible reports.
- [Releases](https://tokspur.com/releases/): Which tokspur builds are available, with their requirements and setup guidance.
- [Build with us](https://tokspur.com/consulting/): Connect coding agents to your systems, tackle slow inference or build a private assistant. Work with San Digital on the software your team needs.
- [Privacy](https://tokspur.com/privacy/): The boundaries between tokspur inference, the calling app or agent and this static website, including local data and diagnostics.
- [Licence](https://tokspur.com/licence/): Read the tokspur runtime licence and distinguish its terms from model weights, font licences and the tokspur name and logo.