Resuls

Recorded speed, model capacity and coding results.

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

Capaciy

Run Qwen3.8 Flash-Next on a 128 GB M5 Max with all 512 experts.

  • 177Blogical model parameters, rounded
  • 65,536tokens of input and output context
  • 32,768maximum output tokens per response

UD-Q4_K_XL quantisation. Input and output share the context limit.

Speed

Recorded on 7th November 2026 with cached prompts on a 40-core M5 Max with 128 GB of memory.

  • 148 tokens/secover the complete request
  • 7.91 sto receive the complete code

The generated code passed type checking. Compare with llama.cpp.

Resuls

Results and specifications
MeasureTestResult
Model parametersQwen3.8 Flash-Next, UD-Q4_K_XL. Rounded parameter count.177B
ContextShared by input and output.65,536
Maximum outputPer response, within the shared context limit.32,768
Peak speedComplete request for code that passed type checking.148 tokens/sec
Response timeTime to receive the complete code.7.91 s
Speed vs llama.cppIdentical code on 2 tests. llama.cpp without a drafter.2.9×
Speed vs MLX-VLMSame prompts and output length.3.1×
Weight repackingPacked weights match the originals byte for byte.0 weights changed

Mehods and seup

Read the measurement method and the hardware target. Available builds are listed under Releases.