Releases·Tool Release
Unsloth GPT-OSS chat template bug drops answers from replayed history; fixed template adds preserve_thinking
A LocalLLaMA user released an updated fixed Jinja chat template for GPT-OSS after discovering that Unsloth's widely-used template drops the model's answer in multi-turn conversations when a message contains both reasoning and content. The bug explains reported incoherence in GPT-OSS 20B; the author's new template fixes the rendering bug and adds a preserve_thinking option for faster multi-turn inference via prefix caching.
76 engagement1 sourceSat, Sep 26, 2026, 08:35 PM
AI Overview
- Unsloth's GPT-OSS chat template renders only reasoning, dropping the answer, when replayed messages contain both thinking and content — OpenAI's reference template lacks the bug
- GPT-OSS 20B could go fully off-track from the bug; 120B recovers conversation direction from reasoning traces alone
- The fixed template (huggingface.co/arbv/gpt-oss-fixed-jinja-template) adds preserve_thinking, enabling prefix caching for faster multi-turn inference at higher token usage
Related
Releases·Model Release
Qwen 3.8 adds prompt-steered reasoning effort, but official template has bugs
Qwen released its first 3.8 model, introducing prompt-steered reasoning effort via a `reasoning_effort` parameter (xhigh, medium, low). However, the official Jinja chat template has serious issues: it crashes when thinking is disabled, poisons multi-turn chat history, and breaks tool calling with JSON string arguments. A community fix addresses these problems for Qwen 3.5, 3.6, and 3.8.
Qwen released its first 3.8 model, introducing prompt-steered reasoning effort via a `reasoning_effort` parameter (xhigh, medium, low). However, the official Jinja chat template has serious issues: it crashes when thinking is disabled, poisons multi-turn chat history, and breaks tool calling with JSON string arguments. A community fix addresses these problems for Qwen 3.5, 3.6, and 3.8.
Updated Sun, Aug 16, 2026, 01:42 PM — A second community template fixes issues in both the original and the first improved template.
012476 engagement4 sourcesredditThu, Aug 13, 2026, 08:22 PM
Social·Community
Community reports chat template bugs breaking reasoning in Laguna-S-2.1 and DeepSeek-V4
Users report that chat template bugs in llama.cpp and Ollama break reasoning behavior in Laguna-S-2.1 and DeepSeek-V4, causing models to skip thinking or emit empty tags. The community has identified fixes, including updated chat templates and GGUFs, which restore proper reasoning and tool calling.
Users report that chat template bugs in llama.cpp and Ollama break reasoning behavior in Laguna-S-2.1 and DeepSeek-V4, causing models to skip thinking or emit empty tags. The community has identified fixes, including updated chat templates and GGUFs, which restore proper reasoning and tool calling.
Updated Fri, Jul 31, 2026, 08:27 PM — Poolside released updated GGUFs and chat template fixing reasoning and tool calling in Laguna-S-2.1.
113290 engagement6 sourcesredditWed, Jul 22, 2026, 07:56 AM
News·Incident
Fine-tuning Qwen3-8B with standard chat template can erase thinking mode without affecting loss or evaluations
During fine-tuning of Qwen3-8B, one build lost its thinking mode while the other retained it, yet training metrics and evaluations showed no difference. The cause was the chat template: Qwen3's standard template renders non-thinking assistant turns with an empty <think> block, which when repeated in answer-only training data can train the model to always output an empty thinking field. This highlights a subtle data leakage issue that can silently disable thinking capabilities.
During fine-tuning of Qwen3-8B, one build lost its thinking mode while the other retained it, yet training metrics and evaluations showed no difference. The cause was the chat template: Qwen3's standard template renders non-thinking assistant turns with an empty <think> block, which when repeated in answer-only training data can train the model to always output an empty thinking field. This highlights a subtle data leakage issue that can silently disable thinking capabilities.
0 engagement1 sourceredditSun, Jul 26, 2026, 12:45 AM
Social·Community
Reddit user shares jinja template prompting rules that curb overthinking in Qwen3 8B with preserve_thinking=false
A Reddit user reports success configuring Qwen3 8B in llama.cpp with preserve_thinking=false, extracting and removing reasoning_content via their harness to mimic Claude-style context-efficient reasoning summaries. The setup requires a non-default, tricky configuration across harness and backend, limiting its usefulness to experienced users.
A Reddit user reports success configuring Qwen3 8B in llama.cpp with preserve_thinking=false, extracting and removing reasoning_content via their harness to mimic Claude-style context-efficient reasoning summaries. The setup requires a non-default, tricky configuration across harness and backend, limiting its usefulness to experienced users.
23 engagement1 sourceredditSat, Sep 5, 2026, 12:50 AM
News·Company News
DeepSeek nears $500M ARR, plans IPO as valuation hits $71B
DeepSeek is approaching $500M in annual recurring revenue and considering an IPO, with a valuation of $71B. The company joins OpenAI and Anthropic as major AI startups eyeing public markets.
DeepSeek is approaching $500M in annual recurring revenue and considering an IPO, with a valuation of $71B. The company joins OpenAI and Anthropic as major AI startups eyeing public markets.
60 engagement2 sourcesredditWed, Jul 15, 2026, 10:07 AM