Small model 1017 loses evals to Haiku but "sometimes has frontier-level ideas," argues researcher
teortaxesTex · x · 2026-11-05
AI commentator teortaxesTex defends the small model codenamed V4.1F-1017: he accepts that evals like AA show it losing to Haiku or Luna — "that's fair. On those evals." But he argues it is a vastly superior model, since Haiku or Luna cannot produce better ideas than Opus or Astra, while 1017 sometimes can.
His core point: benchmark scores and real capability can diverge — this is a small model with frontier ambitions, occasionally producing top-tier creative ideas despite its size.
More from Models
- Claude Max user shows $100/mo plan with only 24% of weekly limit used on Opus — prasenx · 2026-11-05
- ThursdAI weekly: Gemini 4 Carbon takes #1, Anthropic files S-1, sandbox escapes — thursdai_pod · 2026-11-05
- ThursdAI weekly: Gemini 4 Carbon takes #1, Anthropic files S-1, sandbox escapes — altryne · 2026-11-05
- Ricursive releases Ripple-65B Preview, claiming US open-weight SOTA on agentic coding — DZhang50 · 2026-11-05
- Liquid AI ships LFM 3 models (1.7B/2B) running locally on iPhone Duo via MLX — helloiamleonie · 2026-11-05
- particle-1 tested on real agent work: near-Opus quality for under $2 a day — heyneighbor · 2026-11-05