Qwen3.6-35B-A3B fails closing CoT token

// 45d agoMODEL RELEASE

Qwen3.6-35B-A3B fails closing CoT token

Alibaba's new sparse MoE model occasionally outputs the multi-token string </thinking> instead of the dedicated </think> closing token during reasoning. This minor regression breaks API adapters and coding harnesses that rely on precise token detection to separate internal reasoning from final output.

// ANALYSIS

This "infinite thinking" bug highlights the fragility of CoT-enabled models when paired with strict regex-based output parsers. The mismatch between the model's intended vocabulary and its generated output suggests training or quantization edge cases, with observed issues occurring across context lengths from 16k to 128k. While quantization like IQ4_NL may exacerbate the behavior, workarounds involve manual Jinja template adjustments or the use of specific reasoning parsers like the vLLM qwen3 implementation.

// TAGS

qwen3.6-35b-a3bqwenllmreasoningai-codingopen-weightsquantization

DISCOVERED

45d ago

2026-04-19

PUBLISHED

45d ago

2026-04-19

RELEVANCE

8/ 10

AUTHOR

Confident_Ideal_5385

// KEEP READING

More AI developer news from the feed

EXPLORE FULL FEED

UPDATE12m ago

Anthropic's Claude expands its agentic capabilities with long-term memory, autonomous design, and complex multi-step automation workflows.

Anthropic's AI assistant, Claude, has received significant updates enabling it to autonomously design and write end-to-end email campaigns and execute complex multi-step automation workflows. Crucially, these workflows are supported by Claude's persistent memory capabilities, which allow the agent to retain context, user preferences, and detailed requirements across distinct sessions and tasks without forgetting critical details or losing continuity.

LAUNCH21m ago

Hermes Desktop drops with native GUI, simple setup

The new Hermes Desktop app is a native, cross-platform GUI that simplifies setting up and interacting with Nous Research's Hermes Agent on macOS, Windows, and Linux. Built natively to reduce latency, the app integrates memory management, user profiles, and agent skills into a unified workspace.

INFRA24m ago

DigitalOcean partners to cut inference costs

DigitalOcean has partnered with vLLM developer Inferact to bring prefix-aware routing and prefix caching to its Serverless Inference platform. By tracking vLLM instance KV cache events, the Inference Gateway routes requests to pods with matching warm cache blocks, boosting hit rates from 25% to over 75%.