Modelpublic

Shockem/froggeric-terse-coder

sourceHugging Faceapache-2.0updated 1d agoView on Hugging Face

0likes

Model Card

froggeric-terse-coder — chat template for Qwen3.8 Terse-Coder

The recommended chat template for the Terse-Coder family (also bundled inside each model repo as chat_template.jinja).

Lineage: derived from froggeric/Qwen-Fixed-Chat-Templates — the house qwen3.8-froggeric-v5 template, tuned for terse-reasoning models. It is what our benchmarks and the daily driver run — serving the model without it (or with a scaffold that overrides the system prompt) changes behavior, most noticeably around tool-call discipline:

—Anti-rumination rules: after a tool error, do not repeat the same command and do not announce retries.
—reasoning_effort: high maps to the model's extended thinking tier; default thinking is on.
—Tool calls use the qwen3coder format (`<toolcall><function=...>), parsed natively by vLLM (qwen3codermcp`), SGLang, and TabbyAPI.
—Completed-turn <think> content is not replayed unless preserve_thinking is set.

Usage

vLLM:

bash

vllm serve Shockem/Qwen3.8-27b-Terse-Coder-NVFP4 \
  --chat-template chat_template.jinja \
  --speculative-config '{"method":"mtp","num_speculative_tokens":3}'

llama.cpp:

bash

llama-server -m Qwen3.8-27b-Terse-Coder.Q4_K_M.gguf \
  --chat-template chat_template.jinja -ngl 99

SGLang:

bash

sglang serve --model-path Shockem/Qwen3.8-27b-Terse-Coder-NVFP4 \
  --chat-template chat_template.jinja

Note: the think-budget ladder (caps per reasoning_effort level) is applied server-side in our deployment, not by the template itself — see the model cards for the measured setup. Licensed Apache 2.0.