Qwen Batch API silently drops document sections with finish_reason=stop, tests show
yiliang114 · ghdev · 2026-09-27
A developer batching translation of nine Qwen Code Markdown docs via /batch-api (DashScope Batch, qwen3.7-plus, maxtokens 16384) found batch collect reported all 9 delivered, yet three files were missing sections.
Key comparisons:
- headless.md: 42 headings/30 code blocks in source → only 20/17 in Batch output, with 6 missing links; the identical request body via realtime Chat Completions preserved all 42/30.
- sdk-typescript.md: 28/17 source → 9/2 via Batch; realtime output complete.
Evidence locating the fault:
- The submitted JSONL contained the full source, and delivered files matched the raw provider response byte-for-byte — truncation happens in the provider's Batch generation path.
- Responses returned HTTP 200 with finishreason=stop (not length), using only 2,623 and 751 output tokens, far below the 16,384 limit.
- All nine per-item body hashes matched, ruling out request-side differences.
- Rerunning the same two docs in a fresh Batch job reproduced the exact same 20/42 and 9/28 truncation; a separate retry that returned stoptimeout was correctly marked failed.
Conclusion: the Batch path silently truncates document structure independent of token limits, while realtime is unaffected.
More from coding & agent
- The Vanishing Apprentice: How AI Is Reshaping the Junior Developer Role — ArtificialOther · 2026-09-28
- Higgsfield ships 11 production skills that leave Claude with editable project files — xiaohu · 2026-09-28
- AI-generated 7-minute SQLite repo explainer stuns with coherent code walkthrough — deedydas · 2026-09-28
- Is Agentic scores how AI-agent-ready your website is, via a single npx command — seanwbren · 2026-09-28
- SolidBot moves real steel: post-processed robot programs now heading into TCP and accuracy tests — MatthewChang · 2026-09-28
- Grok Bot and Muse too dumb for business agents, says engineer comparing Claude Code — jdjohnson · 2026-09-28