New Models 2026-08-10

Meta's new model is small enough to live on your own machine — and it is the first thing Superintelligence Labs has given away

Meta released Muse Glimmer on Monday, a 30-billion-parameter model built to run local agents on hardware people already own, under a permissive Apache 2.0 licence. It is a dense model with a dedicated vision encoder rather than a mixture-of-experts, distilled from the much larger Muse Spark, and it carries a 131,000-token context window. At full precision it needs more than 55 GB of memory; quantised to four bits it fits on a 24 GB consumer graphics card, and Meta reports about 233 tokens per second on an RTX 5090 with speculative decoding, roughly three times the unaccelerated rate. The published scores are strong for the size class — 94.7% on AIME 2026, 83.5 on GPQA Diamond, 76.0 on SWE-Bench Verified — and on agent-specific tests the gap is wide: 75.5 on MCP Atlas against 54.2 for Google's Gemma4-31B and 62.5 for Qwen3.6-27B. It is not a clean sweep. Qwen3.6-27B still edges it on SWE-Bench Verified at 77.2, and beats it on OSWorld-Verified and TerminalBench 2.1. Audio input and output are unsupported, and Meta's own card says the model should not be used by anyone under 18.

Why it mattersTwo things happened at once here. The first is technical: a model that scores in this range on tool use and multi-step tasks now runs on a single gaming GPU, offline, with no per-token bill and no data leaving the room. That changes who can build an agent — a school, a clinic, a small firm with one workstation — and it changes what an agent is allowed to touch, because nothing has to be uploaded to anyone. The second is political. This is the first open model from Meta Superintelligence Labs, and Mark Zuckerberg used the launch to attack OpenAI and Anthropic for keeping their best systems closed, and to press Washington to loosen the training-data rules he says put American labs behind. That argument has teeth right now for an uncomfortable reason: the leading open-weight models of the past year have mostly come out of China — Kimi K3, Qwen3.8-Max, DeepSeek V4-Flash — while the top US systems stayed locked. Meta is betting it can be the American exception, and it has picked the most permissive licence available to prove the point.

#Coding #AI Agents

✓ Verified · 5 sources

Read in the app — free, in 9 languages

Rumour: the anonymous model that just topped a coding benchmark, for free, is said to be Zhipu's unreleased flagship
2026-08-21 DeepSeek's cheap workhorse can now see — and on agent tasks that need eyes it says it is close to Anthropic's best
2026-08-21 Nvidia is paying $6 billion for the machine that builds a rival's models — and hiring 109 of the people who ran it
2026-08-21 Given four hours and a GPU to improve the way AI is trained, the best agent scored 0.25 out of 1 — and most never tried
2026-08-21 ChatGPT can now read your iMessages and send them — and the setting that lets it skip asking is the one OpenAI warns about
2026-08-21

Meta's new model is small enough to live on your own machine — and it is the first thing Superintelligence Labs has given away

Related stories