DeepSeek-V3.2 GGUFs "eat" think tags

// 45d agoNEWS

DeepSeek-V3.2 GGUFs "eat" think tags

Users running Unsloth's DeepSeek-V3.2 GGUF models on llama-server report missing opening <think> tags, which breaks reasoning UI features in tools like Open WebUI. The issue is caused by the chat template prepending the tag to the assistant's response within the prompt, effectively omitting it from the generated output stream.

// ANALYSIS

The missing tag bug is a classic chat template mismatch that highlights the friction between raw GGUF quants and complex reasoning models.

–The <think> tag is included as a postfix in the prompt template, meaning the model starts generating content after the tag has already been "consumed."
–Frontends like Open WebUI rely on the literal presence of the <think> tag in the stream to trigger collapsed reasoning blocks; without it, the raw "thought" text leaks into the main UI.
–The immediate fix is using the --jinja flag in llama-server to ensure the internal engine correctly handles the reasoning field.
–This recurrence of a known R1-era bug suggests that quantization pipelines for newer DeepSeek versions are still struggling with template consistency across different inference engines.

// TAGS

deepseek-v3-2llama-cppggufunslothreasoningllm

DISCOVERED

45d ago

2026-04-20

PUBLISHED

45d ago

2026-04-20

RELEVANCE

8/ 10

AUTHOR

Winter_Engineer2163

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

UPDATE30m ago

Grok Imagine 1.5 adds dual-video generation

A user highlighted a helpful feature in xAI's Grok Imagine 1.5 video generator, where the tool sometimes outputs two different videos simultaneously. This dual-video generation lets users compare options side-by-side and select the best result for their needs.

NEWS35m ago

DeepSeek tops OpenRouter rankings four weeks straight

OpenRouter has reported that DeepSeek has sustained the number one position in its token share rankings for four weeks in a row. This milestone indicates a substantial and steady volume of developer traffic routing to DeepSeek models, highlighting their growing popularity and integration within the AI application ecosystem.

VIDEO42m ago

Google establishes Gemini as the model-layer foundation of a vertically integrated AI stack.

Google's flagship AI model family, Gemini, serves as the critical model-layer foundation of a vertically integrated AI stack that spans from custom TPU hardware to developer APIs and consumer-facing workflows. In the video, The PrimeTime analyzes Google's strategy to dominate the AI ecosystem by playing at all layers—hardware, models, developer tools, and consumer applications. This integration allows Google to optimize performance and cost efficiency, positioning them to capture maximum value across the AI value chain.