Qwen3.8-Scoped-Terse-Chat-Template

Qwen3.8 Scoped-Terse Chat Template

中文说明

An operational chat template for Qwen3.8-27B agent workloads. It keeps the official Qwen message, vision, and tool-call formats while changing two defaults that matter in long-running local agents:

1. Historical reasoning is scoped to the active user/tool round by default.
2. low reasoning effort is tuned for concise completion instead of full-history reconstruction.

This repository contains no model weights.

Status: v0.1.0-beta. Validated on SGLang with Qwen3.8-27B NVFP4, DFlash2 speculative decoding, OpenCode, and 2× DGX Spark TP2. Treat the results below as an operational regression test, not a model-quality benchmark.

Why this exists

The official Qwen3.8 template preserves reasoning from all historical assistant messages by default. That can improve continuity and KV-cache reuse, but very long agent sessions may also reactivate completed tasks. Separately, reasoning_effort is a soft instruction rather than a hard token budget: an ambiguous follow-up can spend the entire output allowance on reasoning and return no final text.

This template provides a middle ground:

• completed-turn reasoning is stripped by default;
• reasoning generated after the latest real user query remains available across tool calls;
• explicit preserve_thinking=true still restores full-history behavior;
• minimal → low, high/max → xhigh aliases improve harness compatibility;
• low focuses on the latest request, asks a concise clarifying question when the request is ambiguous, targets a short reasoning pass, and requests non-empty final text unless a tool call is required;
• image/video placeholders and Qwen XML tool calls are unchanged.

SGLang usage

Download chat_template.jinja, then launch SGLang with:

python -m sglang.launch_server \
  --model-path Qwen/Qwen3.8-27B \
  --chat-template ./chat_template.jinja \
  --reasoning-parser qwen3 \
  --tool-call-parser qwen3_coder \
  --enable-multimodal \
  --host 0.0.0.0 \
  --port 30000

The template is also suitable for separately quantized Qwen3.8-27B checkpoints when their tokenizer and chat protocol remain compatible with the official model.

Request defaults

from openai import OpenAI

client = OpenAI(base_url="http://localhost:30000/v1", api_key="EMPTY")
response = client.chat.completions.create(
    model="qwen3.8-27b",
    messages=[{"role": "user", "content": "Explain the failure briefly."}],
    reasoning_effort="low",
    max_tokens=16384,
    extra_body={
        "chat_template_kwargs": {
            "preserve_thinking": False,
            "enable_thinking": True,
        }
    },
)

preserve_thinking=False is already the template default. Passing it explicitly is useful when a gateway may inject its own defaults.

To restore the official full-history behavior for a request:

{"chat_template_kwargs": {"preserve_thinking": true}}

OpenCode

See examples/opencode.json. The important fields are:

{
  "reasoning": true,
  "interleaved": {"field": "reasoning_content"},
  "options": {"reasoningEffort": "low"},
  "limit": {"context": 262144, "output": 16384}
}

reasoning_content must be declared as interleaved reasoning so it survives assistant/tool round trips without being merged into visible text.

Supported reasoning levels

Unknown values fail closed with a template error.

Regression evidence

A real-world replay used a long OpenCode agent session and the same ambiguous follow-up. The context sizes differ slightly because the second run was a later fork of the session; this is a regression replay, not a controlled quality eval.

Additional live checks covered structured tool calls, a tool-result round trip, image input, preserve_thinking on/off, and all reasoning-level aliases. See benchmarks/dgx-spark-regression.json.

Tests

python -m pip install jinja2
python tests/test_chat_template.py chat_template.jinja

The test suite checks:

• all reasoning-level aliases;
• scoped history by default;
• explicit full-history preservation;
• current-round tool reasoning retention;
• structured tool serialization;
• multimodal placeholders;
• fail-closed handling of unknown reasoning levels.

Limitations

• The 1,024-token target in low is instruction-level steering, not a hard runtime budget. Configure a reasonable total output limit and handle finish_reason=length.
• This template does not delete visible conversation messages or tool results. Long sessions still need compaction or a fresh task boundary.
• Quantization, speculative decoding, sampling, and runtime versions can change behavior. Re-run the included tests and your own agent evals before production.
• No claim is made that this template improves benchmark quality.

Provenance and credits

Derived from the official Qwen3.8-27B chat template under Apache-2.0. Operational ideas were informed by the Qwen and SGLang documentation and community work on long-running Qwen agents, including the Qwen Sharp Chat Templates project. The scoped-terse wording and compatibility tests in this repository are maintained independently.

License

Apache License 2.0. See LICENSE.