Shiftedx/Qwopus3.8-27B-Coder-Flash-MLX-Balanced

Qwopus3.8-27B Coder Flash · MLX Balanced

23.74 GB weights · 6/8 executable coding smoke · 64K/128K retrieval, tools and cached coding passed. Select Balanced for the higher local coding score, Fast for lower memory and faster decode.

4-bit MLP group 32; 14 low-sensitivity projections use group 64. Attention, embeddings, head and eight tail MLP layers use 8-bit group 64. Both retain 96 BF16 recurrent gating projections, publisher tokenizer/template, BF16 vision and native MTP. Exact precision recipe.

Local replay	64K	128K
Prompt tokens	61,441	126,976
Cold turn	192.87 s	579.09 s
Cached coding decode	78.4 tok/s	61.9 tok/s

M4 Max 40-core GPU / 64 GiB; Kapitsa 0.2.27 + DFlash2; one slot; disk cache off; greedy; thinking off. One run per context cell. The eight-task coding smoke missed weighted interval scheduling and TTL-cache semantics. Evaluation contract · Weight hashes.

hf download Shiftedx/Qwopus3.8-27B-Coder-Flash-MLX-Balanced --local-dir ./qwopus-balanced
hf download incoai/Qwen3.8-27B-DFlash2 --local-dir ./dflash2
kapitsa serve --private --local --model ./qwopus-balanced \
  --model-id qwopus-coder --context 131072 --slots 1 \
  --dflash ./dflash2 --disk-cache 0

Use --context 65536 for 64K. Send temperature=0, top_p=1, top_k=0 and chat_template_kwargs={"enable_thinking":false} to /v1/chat/completions. Prompt and output share the limit. DFlash2 is separate.

Structured JSON / vision. With the request schemas, both models passed 8/8 original cases twice, 12/12 held-out cases and 8/8 with native MTP3, graded as exact bare JSON. Apply the matching response_format from structured_output_profiles.json for OCR, chart, spatial-array or two-image replies. The schemas constrain structure without supplying answers. This request profile fixes the previously observed fences and nested arrays. Validation.

Kapitsa 0.2.31 + DFlash2	Balanced	Fast
Structured JSON decode	128.1 tok/s	165.6 tok/s
Code/prose decode	90.2 tok/s	108.0 tok/s

Median of three repetitions on M4 Max / 64 GiB. Native MTP depths 1–3 and DFlash2 each matched plain greedy text on 4/4 short probes. Use --dflash none --spec 3 for native MTP or --spec 0 for plain decoding. Earlier capability checks.

Experimental refusal-direction edit, strength 0.5 on attention outputs in layers 10–17. Refusal reduction and BF16 fidelity unverified. Text, still-image and short MTP checks only; long-context MTP untested.

Converted from pinned Apache-2.0 BF16 using streamed weights. Jackrong source · Kapitsa v0.2.27 · Other variant.