nom666/Qwopus3.8-27B-GGUF

Qwopus3.8-27B · GGUF (with MTP)

GGUF quants 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, 262,144 context length 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. The MTP block is exported (nextn layout), so llama.cpp speculative decoding works out of the box. MLX/Apple-Silicon users: see the MTPLX builds (4bit-Speed · 8bit-Quality).

Files

File	Size	Notes
Qwopus3.8-27B-Q4_K_M.gguf	16 GB	fits 24 GB VRAM with room for context
Qwopus3.8-27B-Q8_0.gguf	27 GB	near-lossless; also the fastest llama.cpp quant on Apple Metal (K-quant kernels bottleneck there, not bandwidth)

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), which makes cross-generation task arithmetic shape-exact:

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

All 1,199 source tensors matched and merged (fp32 math, bf16 out), including the mtp.* head; converted with llama.cpp's convert_hf_to_gguf (MTP exported as nextn) and quantized with llama-quantize.

Results

Benchmarks below were measured on the MTPLX (MLX) builds of this same merge on an Apple M5 Max; GGUF quants at comparable bit-widths were smoke-tested for coherence and MTP drafting but not separately benchmarked.

Non-thinking mode (the recommended regime):

	Qwopus3.8 4-bit	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

Thinking mode — token efficiency: at high reasoning effort, Qwopus3.8 matches base Qwen3.8 on a 13-task hard suite (13/13 both) while spending 10,050 vs 18,455 reasoning tokens — 45% fewer.

Usage

llama-server -m Qwopus3.8-27B-Q4_K_M.gguf -ngl 99 -c 32768 \
  --alias qwopus38 --spec-type draft-mtp --spec-draft-n-max 2 \
  --temp 0.3 --top-p 0.9 --top-k 40

Disable thinking (recommended for agentic/coding loops) via your chat-template kwargs or client; for thinking mode use a generous max_tokens. Do not greedy-decode in thinking mode (a known Qwen-family failure pattern).

Notes

• Recommended for fast agentic/coding loops and instruct-style serving with thinking disabled — 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.