Work and code faser. Zero oken bills.

3.1× vs MLX-VLM 2.9× vs llama.cpp 148 tokens/sec peak

The full 177B Qwen3.8 Flash-Next for chat, writing and code, faster than any other Mac engine. Nothing leaves your machine, and nothing is metered.

Start coding fasterSee the benchmarks

Up o 148 okens/sec

Fast responses from the full Qwen3.8 Flash-Next model, running on your Mac.

tokspur vs llama.cpp

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
See the comparison

Code that passes.

Tokspur carried Rust and Swift repairs from failing test to passing build, clearing every visible test and hidden check along the way.

Start coding See the completed repairs

Your ools, upgraded

Bring local AI to the chat app, editor and coding agent you already use. Choose your setup.

Buil for coding

Work through features and fixes in the editor you already use. Tokspur keeps the model running locally and ready for your next change.

  • Keep work private

    Send files, diffs and conversations to a model on your own Mac. Your code stays under your control.

    Privacy

  • Code without metering

    Ask for another approach, run another edit, keep iterating. Tokspur has no subscription or per-token fee.

    Free and open source

  • Carry work forward

    Tokspur reuses the context your agent has already sent, so each turn can pick up where you left off.

    How it works
  1. A file sheet slides off a stack towards the model block, which carries the t.

    Read

    Your agent sends the files and tool output the task needs.

  2. A vermilion line settles into a gap in a code sheet in front of the model block.

    Edit

    The model replies, and your agent applies the change.

  3. A vermilion marker runs along a row of test blocks while the model block waits.

    Test

    Your agent runs the tests while the model stays loaded.

  4. The top slab of a context stack slides out and back while the slabs below stay.

    Again

    The next turn reuses everything that still matches.

tokspur serve

Token verification

context
let total = items.iter()Restored from cache
draft
.map(|i| i.cost).sum()Proposed continuation
verify
.map(|i| iKept
.costFirst miss
).sum()Discarded
output
let total = items.iter().map(|i| i.priceFull model supplies .price
  1. Reuse. What the model has already read comes back from memory.
  2. Propose. A small draft model guesses the next few tokens.
  3. Verify. The full model checks every guess at once.
  4. Commit. Right guesses are kept, and the full model writes its own token where the first guess goes wrong.

All 512 expers

Run the whole Qwen3.8 Flash-Next model, with every expert and layer. Tokspur fits it onto a 128 GB M5 Max using UD-Q4_K_XL quantisation.

128 GB, with a 96 GB budget for the model and its working memory. That leaves 32 GB outside the runtime budget for macOS, your editor and builds.
Model detailsMemory settings

Sar building

Install Tokspur, connect your coding agent and put your Mac to work.

Start coding fasterSee the benchmarks