Summerwise-27B is a fine-tune of Qwen3.8-27B by Pleochroic. It is not an official release from the Qwen team.

Summerwise-27B

Summerwise is Pleochroic's model for Chinese literary writing. It is built on Qwen3.8-27B and keeps everything that makes the base model good at long-form work, including its native thinking mode, while writing better prose, fiction, poetry and non-fiction.

We saw real potential in Qwen3.8 for literary writing, and wanted to see how much further careful fine-tuning could take it.

Highlights

Summerwise-27B features the following enhancements over Qwen3.8-27B:

• Literary Writing: Higher scores on open-ended literary writing across all four genres tested (prose, fiction, poetry and non-fiction), with the largest gains in poetry and non-fiction.
• Substance: The gains come from content, such as emotional depth, concrete detail and command of language.
• Full Native Thinking: Qwen3.8's plan-draft-revise thinking is preserved, and reasoning length is unchanged.
• Drop-in Compatibility: Same chat template, sampling parameters and serving setup as Qwen3.8-27B.

Model Overview

• Type: Causal Language Model
• Base Model: Qwen/Qwen3.8-27B
• Number of Parameters: 27B
• Languages: Chinese (primary), English
• Context Length: same as the base model
• Precision: BF16
• License: Apache-2.0

Benchmark Results

Writing Performance

	Summerwise-27B	Qwen3.8-27B
Overall
Literary writing HoWToBench	7.34	7.24
By dimension
Content emotion, detail, language, structure, logic	7.22	7.06
Format adherence to the requested form	7.14	7.14
Impression overall reading impression	8.52	8.43
By genre
Prose 67 prompts	7.33	7.27
Fiction 34 prompts	7.02	6.95
Poetry 30 prompts	7.81	7.65
Non-fiction 29 prompts	7.24	7.10

1. HoWToBench (ACL 2026), open-ended literary subset: 160 Chinese writing prompts, scored on a 10-point scale with the official Tree-of-Writing rubric.
2. Both models were evaluated by us under identical settings: thinking mode, temperature 1.0, top_p 0.95, top_k 20. Each model answered every prompt four times and the scores are averaged. The judge model is DeepSeek v4.1 Flash.
3. Overall difference: +0.10, 95% confidence interval +0.04 to +0.16 (paired by prompt). The criterion for calling it an improvement was fixed before the data were seen.
4. The better result in each row is shown in bold.

Quickstart

Summerwise-27B is served the same way as Qwen3.8-27B. Thinking mode is on by default.

Serving Summerwise

vllm serve Pleochroic-Ltd/Summerwise-27B \
  --reasoning-parser qwen3 --language-model-only \
  --mamba-ssm-cache-dtype float32 --max-model-len 32768

API Usage

from openai import OpenAI

client = OpenAI(base_url="http://localhost:8000/v1", api_key="EMPTY")

response = client.chat.completions.create(
    model="Pleochroic-Ltd/Summerwise-27B",
    messages=[{"role": "user", "content": "写一写夏日将尽的傍晚。"}],
    temperature=1.0,
    top_p=0.95,
    extra_body={"top_k": 20},
    max_tokens=16000,
)
print(response.choices[0].message.content)

Best Practices

1. Sampling Parameters: Use the Qwen3.8 thinking-mode defaults: temperature=1.0, top_p=0.95, top_k=20. All results above were obtained with these settings.
2. Adequate Output Length: Summerwise drafts in its thinking before it writes, so thinking on a writing prompt is long. Leave generous room for output tokens and context.

License

Summerwise-27B is released under the Apache-2.0 license, the same as Qwen3.8-27B. We thank the Qwen team for open-sourcing the base model and the authors of HoWToBench.

Contact: contact@pleochroic.co.uk