Plain-spoken Qwen3.8-27B fine-tune: 42% less slop, 73% fewer em dashes, same coding-agent pass rate. MTP speculative decoding built in.

Applications

Claude Code ollama launch claude --model RefinedNeuro/qwen3.8-27b-noslop

OpenCode ollama launch opencode --model RefinedNeuro/qwen3.8-27b-noslop

Hermes Agent ollama launch hermes --model RefinedNeuro/qwen3.8-27b-noslop

OpenClaw ollama launch openclaw --model RefinedNeuro/qwen3.8-27b-noslop

Qwen3.8-27B-noslop

The plain-spoken 27B. A fine-tune of Qwen3.8-27B that cut the base model’s slop score by 42 percent and its em dashes by 73 percent on held-out writing prompts, with coding-agent pass rates identical to the base.

Full-precision model: RefinedNeuro/Qwen3.8-27b-noslop
GGUF files and full methodology: RefinedNeuro/Qwen3.8-27b-noslop-gguf

Quick start

ollama run refinedneuro/qwen3.8-27b-noslop

Thinking is on by default. Turn it off per request with "think": false in the API, or /set nothink in the CLI.

Tag	Size	Best for
`q4_k_m` / `latest`	16.8 GB	The sweet spot for a 24 GB card with long context
`q3_k_xl`	13.1 GB	Per-layer mixed precision, best quality per byte near 3 bit
`q3_k_m`	13.5 GB	Stock llama.cpp 3-bit
`q2_k_xl`	9.8 GB	Smallest file, per-layer mixed precision. The one tested in coding agents
`q2_k_m`	10.9 GB	Stock llama.cpp 2-bit

What’s set up

Chat format, thinking and tool calls use Ollama’s built-in Qwen3.8 renderer and Qwen3.5 parser, the same as the official qwen3.8 model.
Multi-token prediction: every tag keeps the model’s MTP head, and draft_num_predict 4 is set for speculative decoding. In llama.cpp tests, 46 to 47 percent of draft tokens were accepted.
Sampling defaults (recommended for plain answers): temperature 0.7, top_p 0.8, top_k 20, min_p 0, repeat_penalty 1.
Importance matrix built on this model, from a mix of our prompt bank, the original model’s answers and human-written stories.

Measured, not promised

Coding agents: Q2_K_XL served through llama-server passed the hidden-test tasks in pi (23 of 24) and Claude Code (22 of 24), the same as an identically quantized base.
Style: em dashes are reduced by roughly three quarters on writing prompts, not eliminated, and the reduction does not carry over to tool-using agent sessions.

Limitations

Text only in these Ollama tags. The vision tower is unchanged from Qwen3.8-27B; for images, use the GGUF with the mmproj file from unsloth/Qwen3.8-27B-GGUF in llama.cpp.
Low-bit quantization costs quality; Q2 tags are for tight memory budgets. Perplexity and long-context quality were not measured per file.
Thinking-mode behaviour was not evaluated.

License

Apache 2.0