LiquidAI's 1.7B LFM3 hits SOTA on Multi-IF, beating 100x-larger rivals including ERNIE 5.0
NielsRogge · x · 2026-11-11
LiquidAI's new LFM3-1.7B open model sets SOTA on Multi-IF, Meta's benchmark for multi-turn multilingual instruction following: 89.1, beating Baidu's closed ERNIE 5.0 (85.56) and Qwen3-VL 235B-A22B Thinking (79.1) despite being 100x+ smaller. Raises the question of how small a good assistant model can get.
More from Models
- OpenAI open-sources mcp-extensions to make plugins feel native to ChatGPT — SIGKITTEN · 2026-11-11
- Logan Kilpatrick questions AI leaderboard: why is only our model not tested in its native harness? — zacharynado · 2026-11-11
- Unverified hands-on claim: 'Gemini Carbon' is frontier-grade but hallucinates and rambles — zacharynado · 2026-11-11
- Top AI models still can't beat the 95.4% human score on word logic puzzles — ChrisGPT · 2026-11-11
- RISJ data: ChatGPT leads AI trust at +17, Grok lowest at -6 — _FelixSimon_ · 2026-11-11
- 7 testers put Gemini 4 Carbon through a Reach Test: 2 would always use it, 1 won't — every · 2026-11-11