Llama.app adds Liquid AI LFM3, Mistral Large 4 and coding-agent model Flint
mervenoyann · x · 2026-11-03
The llama.app team (from the llama.cpp team and Hugging Face) rounded up last week's notable model drops, all runnable in its free, offline app:
- Liquid AI LFM3 family: small on-device multimodal models — LFM3-1.7B (16-language text, runs on CPU), LFM3-Omni-2B (see/watch/hear/talk), LFM3-Audio-2B, and d1-2B, a decision model that outputs yes/no, choices and answer scores over text, image and video.
- Mistral Large 4: a 1T-param (52B active) frontier model with 1M context and image input; the 2-bit quant fits in 512 GB of unified memory, and it's said to be strong at cybersecurity and agentic coding.
- Recursive Flint: built for coding agents — 2.5B active of 39.8B, 74.2% on SWE-bench Verified, 262K context, can investigate codebases, fix bugs and write tests.
More from Infra
- $640 homelab with 40 cores and 1TB RAM hits 63 concurrent VMs — haydendevs · 2026-11-03
- Homelab builder asks: any non-Mac alternative with 400GBps+ unified memory bandwidth for local LLMs? — AcreMakeover · 2026-11-03
- Nvidia tells data center clients to cap China exposure at 51%, per FT — rohanpaul_ai · 2026-11-03
- Supabase's Multigres enters private alpha; head explains HA-first and custom pooler choices — dshukertjr · 2026-11-03
- Break-even math: 4.8B tokens/month is the crossover for owning an 8-GPU H200 node — recentheartbroken · 2026-11-03
- Dev unifies RTX 20/30/40 tensor core FP32-accumulation unlock into one cross-platform tool — MindAllik · 2026-11-03