Users Report DeepSeek V4.1-Flash-1017 Update Regressed on Terminal-Bench 4.0
teortaxesTex · x · 2026-10-26
User @gosrum found DeepSeek V4.1-Flash-1017 losing to the pre-update V4.1-Flash on Terminal-Bench 4.0; testing with both DSH and opencode harnesses ruled out evaluation framework issues, pointing to the model itself, with follow-up runs planned via the official API. @teortaxesTex adds that in his experience V4.1F is weaker than 1017 in every way that matters. Unconfirmed regression claim pending official response.
More from Models
- Dermatologist: a 10-year-old CNN beats frontier LLMs at skin cancer diagnosis, 62% vs 42% — RazRazcle · 2026-10-26
- Users complain ChatGPT is getting lazy, ignoring docs and memory unless explicitly told — sheeeeeeeeesh24 · 2026-10-26
- Together AI adds MiniMax M3.1 Flash post-training with SFT and RL support — togethercompute · 2026-10-26
- Reddit speculates Anthropic delays new model releases to reserve compute for internal use — arknightstranslate · 2026-10-26
- DeepSeek V4 and Kimi 3.1 Pro flub prompt injection help as commenter praises Claude's 'scary good' theory of mind — teortaxesTex · 2026-10-26
- User gets ominous Gemini policy-violation warning from Google with zero explanation — kingslyj · 2026-10-26