This repository is a corrected fork of yunmorning/broken-model.

All other files (config.json, generation_config.json, model weights, vocabulary files) are identical to the original.

The tokenizer_config.json in the original repository was missing the chat_template field entirely.

chat_template is a Jinja2 template that an inference server uses to convert an OpenAI-style messages array:

into the raw token sequence the model was trained on:

<|im_start|>system
You are a helpful assistant.<|im_end|>
<|im_start|>user
What is 2 + 2?<|im_end|>
<|im_start|>assistant

Without this template, any /chat/completions API server (vLLM, TGI, FriendliAI Engine, etc.) either raises a KeyError / TemplateNotFound error at request time, or falls back to a generic template that produces token sequences the model was never trained to handle — resulting in garbage output or a hard failure.

The chat_template value was taken verbatim from the canonical Qwen/Qwen3-8B release on Hugging Face. This is the correct template for the ChatML format Qwen3 was trained with, including support for tool calls and the optional enable_thinking reasoning mode.

No other values were modified.

reasoning_effort is an OpenAI API extension for models with budget-controlled extended thinking — where the depth of the model's internal reasoning chain is regulated at the server level (e.g., o1, o3, Claude 3.7 Sonnet extended thinking).

Two distinct issues prevent it from working here:

Issue 1 — Prerequisite: missing chat_template
Qwen3's thinking mode is toggled via the enable_thinking variable inside chat_template. Since chat_template was absent, the server could never render the thinking-enabled prompt format at all. Even if reasoning_effort were correctly mapped, thinking would remain off.

Issue 2 — No server-side mapping from reasoning_effort to enable_thinking
Even with the template restored, reasoning_effort does not automatically activate thinking. The inference server must explicitly read reasoning_effort from the incoming request and pass enable_thinking=True/False as a template variable when rendering the prompt. Without this mapping, the parameter is silently ignored.

Issue 3 — No budget-level granularity in Qwen3-8B
Qwen3's thinking mode is a binary switch (enable_thinking: true/false). The reasoning_effort values "low", "medium", "high" are meaningless unless the model was post-trained with reasoning budget forcing — a technique where target thinking length is conditioned on during RLVR/GRPO training so the model can produce calibrated reasoning depth at inference time. Qwen3-8B does not have this training.

Steps 1–3 can be addressed at the infrastructure level without touching model weights. Step 4 requires a new training run and produces a distinct model variant.