Observed Signal · May 18, 2026 · Technical Issue · Source: DEV Community · Impact: 2/5 · Sentiment: Negative
Qwen 3.6 enable_thinking broke agent JSON parsing
A developer field report describes a Qwen 3.6 behaviour where the model's default reasoning mode inserts <think>…</think> chain-of-thought blocks before the requested output, which broke JSON parsing in agent/tool loops. The root cause is a chat-template flag: apply_chat_template injects reasoning markers by default; passing enable_thinking=False to tokenizer.apply_chat_template removes the reasoning preamble and restores clean structured outputs. The flag must be applied at template-render time (not to model.generate() or tokenizer load). The post recommends disabling thinking for machine-parsed paths and keeping it enabled for human-facing chat to preserve answer quality.
Developer-facing behavioural change in a widely used LLM (Qwen 3.6) can silently break agentic pipelines that parse structured outputs; relevant to teams building agent/tool integrations and pipeline stability.
Track OpenAI Signals & Market Shifts in Real-Time
Polaris7 autonomous intelligence agents track regulatory filings, primary sources, executive changes, and deal flow 24/7. Create your free Explorer workspace to monitor these entities.
Key Takeaways & Evidence Grounding
- Qwen 3.6 defaults to a reasoning mode that inserts <think>…</think> chain-of-thought blocks before the user-facing answer.
- The remedy is to call tokenizer.apply_chat_template(..., enable_thinking=False) so the chat template does not inject <think> markers.
- The enable_thinking flag only takes effect when passed to apply_chat_template; passing it to model.generate(), tokenizer load, or generate wrappers has no effect.
- Earlier Qwen versions (e.g., 2.5) did not include the enable_thinking behaviour; the feature is new in Qwen 3.6 MoE.
Connected Companies & Entities
1 Entity mappedOntology Mapping & Concepts
Related Market Signals & Shifts
Recent verified developments and strategic activity across this market segment.
Qwen3.8-27B multimodal language model guide
This article is a technical guide to Qwen3.8-27B, a 27-billion-parameter dense causal language model with an integrated vision encoder. The model supports native image and hour-scale video understanding, a native 262,144-token context window (extendable to 1,000,000 tokens), and a default "thinking mode" that emits explicit reasoning chains. Qwen3.8-27B is released in Hugging Face Transformers format, compatible with inference frameworks such as vLLM, SGLang, and TokenSpeed, and is distributed under the Apache 2.0 license. Recommended sampling and reasoning parameters, typical hardware expectations for 27B dense models, benchmark results across coding, multimodal, math/vision, and document tasks, and limitations (token cost, latency, and hardware needs) are documented.
Measured Context Window Reveals Why AI Agent Deteriorated
A June 17, 2026 DEV Community post by Rapls describes diagnosing an AI coding agent that seemed to get 'dumber' mid-session. Instead of immediately disabling connected MCP tools, the author inspected a per-category breakdown of the model's context window. Measurement showed conversation history was the largest consumer of tokens (roughly a fifth of the window), while connected MCP tool definitions were a small slice in their setup. The author concludes that long session history accumulation — not always visible tooling overhead — commonly drives quality drift. Practical mitigations include scoping sessions, summarizing and carrying forward concise summaries or locked decision blocks, re-grounding against source files, and measuring token allocation before removing tools.
GPT-5.4 Arrives: ChatGPT Reveals Reasoning, Controls Apps
OpenAI is rolling out GPT-5.4 across ChatGPT, the API, and Codex, introducing GPT-5.4 Thinking and GPT-5.4 Pro for more complex tasks. The update presents a reasoning-first interface, showing users the planned solution path and enabling intervention before final answers. GPT-5.4 Thinking will replace GPT-5.2 Thinking for Plus, Team, and Pro users, with GPT-5.2 remaining as a Legacy option until June 5, 2026. OpenAI reports reliability gains, citing a 33% reduction in incorrect statements versus GPT-5.2 and an 18% decrease in errors in complete answers. In GDPval benchmarks across 44 professions, GPT-5.4 meets or surpasses industry experts in 83% of cases. The release also includes an Excel Add-in enabling natural-language creation, analysis, and updating of tables. Overall, OpenAI emphasizes embedding AI more deeply into real-world workflows and enabling agents to operate software across environments.
Track Real-Time Market Signals & Shifts
Set up custom watchlists to receive automated, evidence-grounded executive digests whenever material signals or shifts occur across your tracked landscape.