DeepSeek V4.1 Tool Calling Allegedly Regresses, Breaks 100-Step Agent Workflow
MaziyarPanahi · x · 2026-09-11
A developer reports DeepSeek-V4.1-Flash via OpenRouter shows a big drop in tool calling accuracy, breaking his 100+ step clinical agent workflow, while the older V4-Flash-0731 completes it fine. He questions whether the issue lies with the model, the provider, or the serving setup — unverified, but suggesting a possible regression in the new version.
Related event: DeepSeek-V4.1-Flash Regression Breaks 100-Step Agent Workflows(2 posts)→
More from coding & agent
- Anaconda demos declare-to-cloud workflow: dependencies, browser dev, containers, cloud runs — anacondainc · 2026-09-11
- Claude Code v2.1.268 ships gateway billing sync, fixes HTTP 400 on third-party endpoints — ashwin-ant · 2026-09-11
- Modeling a fraud-approval agent as a POMDP: approve, verify, or escalate under uncertainty — deepikaau · 2026-09-11
- Ruff shaves 8 more bytes off Expression enum, cutting retained memory by >1% — charliermarsh · 2026-09-11
- Developer Admits He Was Wrong: Astra Agent Does CAD in Onshape via Chrome, Design-to-Print Workflow — carlosdponx · 2026-09-11
- NVIDIA releases Sol-Pi: four mechanisms to cut token costs of Pi coding agents — Thrumpwart · 2026-09-11