Good morning — it's Wednesday, November 4, 2026. Mistral Large 4 is open — and its best score belongs to its own agent. From the bench, LFM3-Omni-2B hears, sees, and answers back. In the drops: Yantr's rootless app store and an engine that streams giant models off your NVMe drive. Don't keep us a secret: forward this to a fellow home labber, or share the web version.
Share this issue: X WhatsApp Reddit Facebook LinkedIn Text Email
On October 29, Mistral released the weights of Mistral Large 4: about one trillion parameters, 49 billion active per token, a one-million-token context, trained on 3,800 NVIDIA Grace Blackwell GPUs in Europe.
The headline claim (vendor): on Terminal-Bench 4.0, a 66-task terminal test, the released model improved by 13 points to 41% over its preview, leading open-weight models with 30% fewer output tokens.
The catch: that 41% was measured inside Mistral Vibe, Mistral's own coding agent. In a minimal, generic agent, the gain disappears: Artificial Analysis measures 25% (independent) in mini-swe-agent, Vals AI 26.26% (independent) — level with the preview in the same harness. The own-agent gap is 16 points here; for the other twelve models measured it runs from minus 4 to plus 7, median zero.
Nobody is serving a trillion parameters from a home rack — the card's example spreads it across eight GPUs. But labbers pick models from these charts: a coding score belongs to the model plus its agent, and the download does not include the agent. The license is not plain MIT either — free to use and commercialize, but hosting it for others needs a separate license. When a lab hands you a benchmark, ask which agent was driving.
LFM3-Omni-2B, from Liquid AI, is trending on Hugging Face, downloads still in the low five figures. The family's first omni model: text, images, video, and audio go in, text or speech comes out — one small model doing the work of a separate transcription, vision, language, and speech model. A non-reasoning model, built for fast perception, not coding or agent work.
The 2B beats models three times its size on the omni tests — and loses two measures the same tables print. Higher is better, except word error rate:
Daily-Omni: LFM3-Omni-2B 70.09, Qwen2.5-Omni-7B 61.32 — Omni ahead by 8.77
WorldSense: LFM3-Omni-2B 47.10, Qwen2.5-Omni-7B 45.37 — Omni ahead by 1.73
AV-SpeakerBench: LFM3-Omni-2B 46.79, Qwen2.5-Omni-7B 43.24 — Omni ahead by 3.55
OmniBench overall: LFM3-Omni-2B 46.32, Qwen2.5-Omni-7B 49.65 — the 7B ahead by 3.33
English ASR: LFM3-Omni-2B 6.36, Phi-4-multimodal 5.14 — Phi-4 ahead by 1.22
Every score above was measured by Liquid AI (vendor); nothing independent has replicated them yet.
Setup notes: a GGUF build runs in llama.cpp, speech server included; the full checkpoint serves via SGLang or Transformers with trust_remote_code. Difficulty: light setup.
An app store that leaves your OS alone. Yantr puts 170-plus self-hosted apps on the Linux you already run, each in a rootless Podman container under your own account. The catch is in its own FAQ: personal use is free, commercial use is not, under the PolyForm Noncommercial License.
GitHub find: colibri streams the giant models off your disk. Colibri is a C engine for very large mixture-of-experts models: the dense part stays in RAM, a GPU holds the hottest experts, and the rest stream from NVMe as the router asks, through a cache that learns your work. No GPU required, 47,352 stars, Apache-2.0.
Tuesday, 11 p.m. Your monitoring digest shows a successful SSH login to your media server at 3 a.m. from an address in another country. The box has one port forwarded so a friend can reach it, and the friend's account on it is years old.
Your call:
A. Block that address at the router and audit the logs properly tomorrow
B. Change every password on the box, kill the session, and keep it running while you watch it
C. Pull the box off the network, preserve the logs, and treat it as compromised until you have checked for new accounts and persistence
100 Hero Points for the right call. Your score is on the outcome page, ready to share.
The first principle is that you must not fool yourself — and you are the easiest person to fool.
— Richard P. Feynman
Reading this because a fellow home labber forwarded it? Get your own copy every morning — subscribe free, and pass it on — that is how this grows.
How we source: facts come from primary sources opened for this issue. Benchmarks are labeled (vendor), (independent), or (community), never blended. Deal prices are checked the day they run. This is not buying advice.
Corrections: none today.
Reader-supported: paid subscriptions, labeled sponsors, and affiliate links on some gear and deal items (we may earn a commission). Coverage is never paid for — sponsors and affiliate programs don't pick stories.