nom666/Qwopus3.8-27B-MTPLX-4bit-Speed

Qwopus3.8-27B · MTPLX 4-bit Speed

The speed-focused MTPLX build of Qwopus3.8-27B — a task-vector transplant that carries the Qwopus reasoning distillation onto the Qwen3.8-27B base. Qwopus3.8 keeps Qwen3.8's architecture, context length (262,144) and native MTP speculative head, and layers the Opus-style distilled reasoning trait on top: materially fewer tokens for the same result, and a stronger non-thinking (instruct) mode — which is exactly the regime fast agentic and coding loops run in. Higher-precision sibling: Qwopus3.8-27B-MTPLX-8bit-Quality.

How it was made

Qwen3.6-27B and Qwen3.8-27B share an identical qwen3_5 skeleton (64 layers, 5120 hidden, 248K vocab, hybrid GDN + full attention, 1-layer MTP head). That makes cross-generation task arithmetic shape-exact:

Qwopus3.8 = Qwen3.8 + (Qwopus3.6-27B-v2 − Qwen3.6)

• All 1,199 tensors matched and merged (fp32 math, bf16 out) — including the mtp.* speculative head and the vision tower; zero skipped, zero shape mismatches.
• Forged with MTPLX: body 4-bit (group 32, affine) with 8-bit overrides on embeddings, lm_head, linear_attn.out_proj, and the last-8-layer MLPs; MTP head kept bf16.
• The delta-merged MTP head verifies above the donor artifacts: acceptance 97.5% / 85.6% at draft depths 1–2.

Results (vs Qwen3.8-27B, same harness, same quant tier, Apple M5 Max)

Non-thinking mode — the recommended regime for this build:

	Qwopus3.8 4-bit (this build)	Qwopus3.8 8-bit	Qwen3.8 8-bit
38-task instruct/code/math suite	37/38	36/38	35/38
Instruction-following subset	4/5	3/5	2/5
Hard-task suite (executed code + math + logic)	—	8/13 @ 2,889 tok	8/13 @ 3,384 tok
Decode @ d0 (turbo)	48.6 tok/s	49.6 tok/s	~49 tok/s
Decode @ 32K depth	45.4 tok/s	—	—

Non-thinking Qwopus3.8 answers with ~15% fewer tokens at equal hard-task success, and holds instruction-following notably better than the base model does with thinking disabled.

Thinking mode — token efficiency: at high reasoning effort, Qwopus3.8 matches the base on a 13-task hard suite (13/13 both) while spending 10,050 vs 18,455 reasoning tokens — 45% fewer. The efficiency trait holds on every jointly-solved task of a harder adversarial set (e.g. 11.6K vs 18.0K tokens on a combinatorial coding task).

Usage

brew install youssofal/mtplx/mtplx
mtplx pull nom666/Qwopus3.8-27B-MTPLX-4bit-Speed
mtplx quickstart --model nom666/Qwopus3.8-27B-MTPLX-4bit-Speed \
  --reasoning off --profile turbo \
  --temperature 0.3 --top-p 0.9 --top-k 40

For thinking mode, use --reasoning on --reasoning-effort low (or medium) with a generous max_tokens. Do not greedy-decode in thinking mode (a known Qwen-family failure pattern; temperature 0 makes reasoning termination deterministic-worst-case).

Notes

• Recommended for fast agentic/coding loops and instruct-style serving with reasoning off — that is where this merge measurably beats its base.
• With thinking enabled, Qwopus3.8 solves hard tasks with ~45% fewer reasoning tokens, though base Qwen3.8 remains the stronger choice for heavy thinking-mode workloads — it is more consistent at concluding very long reasoning inside tight token budgets (Qwopus3.8 can occasionally deliberate past a tight budget on some prompts) and keeps an edge on the hardest reasoning tasks.
• This is a weight-space merge, not a fine-tune: no gradient training was performed on Qwen3.8.
• All credit for the Qwopus distillation to Jackrong (Qwopus3.6-27B-v2); base model Qwen3.8-27B by the Qwen team; MTPLX runtime by Youssofal.