Code a speed

Spend less time waiting for code. Tokspur runs the full Qwen3.8 Flash-Next model locally, with no token bills.

A ruled track, moving marker, stopwatch and waiting token blocks.

3.1× vs MLX-VLM

More code, less waiting. The full Qwen3.8 Flash-Next model, running locally.

More bits retain finer model detail. A draft helper proposes tokens for the main model to check.
  • tokspur
    147.6 tokens/sec
    Main model
    4-bit mix
    Draft helper
    8-bit draftsDrafter options
  • MLX-VLM
    47.2 tokens/sec
    Main model
    4-bit mix
    Draft helper
    Off

2.9× vs llama.cpp

Identical code, delivered sooner. Both code tests passed type checking.

More bits retain finer model detail. A draft helper proposes tokens for the main model to check.
  • tokspur
    140.7 tokens/sec
    Main model
    4-bit mix
    Draft helper
    8-bit draftsDrafter options
  • llama.cpp
    48.5 tokens/sec
    Main model
    4-bit mix
    Draft helper
    Off

Choose your drafer

Native MTP

The default 8-bit helper, built into Tokspur with 2- or 3-token lookahead.

DFlash2

A separate draft model that proposes up to 7 tokens at once.

Prompt lookup

Reuses matching code and text already in context, alongside native MTP.

Ahead of every engine we have found

Recorded and published Qwen3.8 Flash-Next rates on M5 Max, with Tokspur at the top. Each row names the engine it runs on and its setup.

M5 Max, 40-core GPU, 128 GB. 7th November 2026.More bits retain finer model detail. A draft helper proposes tokens for the main model to check.
  • tokspurOwn Metal engine
    148.0 tokens/sec
    Main model
    4-bit mix
    Draft helper
    8-bit draftsDrafter options

    Cached prompts, UD-Q4_K_XL with Q8 drafts. Complete request time.

    Benchmark
  • MTPLX 2.11.3, publishedMLX
    125.8 tokens/sec
    Main model
    4-bit mix
    Draft helper
    16-bit built-in

    Optimized Speed, depth 3, cached prompts and maximum fans.

    Source
  • Strata, Q2_0 with MTPOwn Metal engine
    122.8 tokens/sec
    Main model
    2-bit
    Draft helper
    2-bit experts

    Q2_0 with fresh context. Generation time only.

    Source
  • Splash 1.4.0, localOwn Metal engine
    112.5 tokens/sec
    Main model
    4-bit mix
    Draft helper
    On

    UD-Q4_K_XL with drafts and fresh code prompts. Generation time only.

  • oMLX, adaptive MTPMLX
    111.0 tokens/sec
    Main model
    3.5-bit mix
    Draft helper
    Built-in MTP

    oQ3.5e with fresh context. Server timing.

    Source
  • Lily, Q4 with MTPOwn Metal engine
    107.0 tokens/sec
    Main model
    4-bit
    Draft helper
    Built-in MTP

    Q4 with 2 drafts. Median of 3 requests.

    Source
  • mlx-serve, opt-in TokenV3MLX
    106.5 tokens/sec
    Main model
    Mixed precision
    Draft helper
    MTP on

    Mixed pack with KV8 and depth 3. Modified sampling.

    Source
  • MFQ, V3-XXS with MTPOwn Metal engine
    105.4 tokens/sec
    Main model
    Mixed precision
    Draft helper
    Built-in MTP

    V3-XXS quantisation with Python code prompts.

    Source
  • mlx-serve, default verifierMLX
    92.5 tokens/sec
    Main model
    Mixed precision
    Draft helper
    MTP on

    Mixed pack with KV8 and depth 3. Default sampling.

    Source
  • llama.cpp with MTPllama.cpp (ggml)
    59.3 tokens/sec
    Main model
    4-bit mix
    Draft helper
    8-bit sidecar

    UD-IQ4_XS with 2 drafts and fresh context.

    Source
  • Ollama, plain GGUFllama.cpp (ggml)
    44.5 tokens/sec
    Main model
    4-bit mix
    Draft helper
    Off

    UD-Q4_K_XL without drafts. Median of 3 runs.

  • llama.cpp, no MTPllama.cpp (ggml)
    44.4 tokens/sec
    Main model
    4-bit mix
    Draft helper
    Off

    UD-IQ4_XS without drafts, with fresh context.

    Source

Code ha passes

Tokspur carried Rust and Swift repairs from failing test to passing build, clearing every visible test and hidden check along the way. Bring it into your workflow with OpenCode.

Completed repairs, 2nd November 2026. M5 Max, 40-core GPU, 128 GB.
RepairSetupResult
Rust repairtokspur defaultPass
Foundation Swift repairtokspur with NAXPass

Sar coding

Put Qwen to work in your editor. Your code stays on your Mac, with no per-token bill.

Get startedConnect your tools
Test setup

M5 Max, 40-core GPU, 128 GB. Recorded on 7th November 2026 with cached prompts.

MLX-VLM comparison

Same prompts and output length. Tokspur used UD-Q4_K_XL with Q8 drafts. MLX-VLM 0.7.12 used oQ4e with a checkpoint scale fix and no drafter.

llama.cpp comparison

Complete request time across 2 code tests, with the same UD-Q4_K_XL target weights and prompts. Tokspur used Q8 drafts. llama.cpp v0.9.0 ran without a drafter.

Published MLX-VLM reference

Stock MLX-VLM with fresh context, measuring generation time.

M5 Max MacBook Pro, 128 GB. 7th November 2026.More bits retain finer model detail. A draft helper proposes tokens for the main model to check.
  • Stock MLX-VLM 0.7.8, no MTPMLX
    56.4 tokens/sec
    Main model
    4-bit
    Draft helper
    Off

    4-bit MLX with fresh context. Generation time only.

    Source
Measurement method