← /editions
#60

Meta puts a 30B agentic model on your desk

// Meta Superintelligence Labs released Muse Glimmer, a 30B multimodal model for always-on local agents, under Apache 2.0. It runs on a single 24GB consumer GPU, leads its size class on agentic orchestration benchmarks, and relocates the entire agent threat model onto the deployer.

$ today_in_10_seconds

Meta Superintelligence Labs released Muse Glimmer, a 30B multimodal model built for always-on local agents — and put it under Apache 2.0, not a custom community license. It runs on a single 24GB consumer GPU or an M-series Mac, beats its size class on agentic orchestration benchmarks, and posts a 28.4% prompt-injection success rate. Both of those numbers matter.

$ lead_story

Meta returns to open source with Muse Glimmer — Apache 2.0, 30B, agent-optimized

Meta distilled Muse Glimmer from its flagship Muse Spark teacher model, quantized it to roughly 4-bit to drop memory needs from 55GB to under 20GB, and bolted on DFlash speculative decoding to get decode speeds fast enough to survive an agent loop — Meta claims 3.1x on an RTX 5090. Weights are on Hugging Face today with BF16, GGUF k-quants, ExecuTorch builds, and the drafter model included. The licensing is the real story: Apache 2.0 is a clean break from the Llama community license, meaning commercial use, modification, and redistribution with no acceptable-use addendum and no MAU threshold. For anyone building agents that need deep access to personal or regulated context, the calculus just changed — you can now run a competent tool-using model where the data already lives.

// why_it_matters: The cheapest agent token is the one you never send over the network, and the safest context is the context that never leaves the device.

read more →

$ also_worth_attention

  • Meta AI Releases Muse Glimmer: A 30B Open-Weights Agentic Model That Runs on One Consumer GPU

    The benchmark spread is more interesting than the headline. Glimmer leads Gemma4-31B and Qwen3.6-27B on MCP Atlas (75.5 vs 54.2 and 62.5), DeepSearch QA, Gaia2, and SWE-Bench Pro — but loses to Qwen on OSWorld-Verified and TerminalBench 2.1. Read that pattern carefully: it's strong at planning and orchestrating tool calls, weaker at actually driving a computer or a shell, which is precisely where "always-on desktop agent" work happens.

    link →
  • Meta releases Muse Glimmer, a 30B open agentic AI model that runs locally on PCs

    The hardware envelope is the product decision. Meta sized the whole stack — weights, KV cache, image encoder, and speculative drafter — to fit inside 24GB or 32GB of VRAM, which is a deliberate bet that the interesting agent deployments are the ones with no network dependency at all. Unsloth's dynamic quants reportedly get it down to 18GB, putting it within reach of hardware plenty of engineers already have under their desk.

    link →
  • The ecosystem shipped support on day zero

    Ollama, llama.cpp, LM Studio, MLX, and ExecuTorch all landed Glimmer support within hours, with AMD publishing tuning guidance for Ryzen AI Max+ and Radeon AI PRO systems the same morning. LM Studio can serve it over LAN behind an OpenAI- or Anthropic-compatible endpoint, which means existing agent frameworks point at it with a base-URL change. Coordinated day-zero tooling is now table stakes for an open-weights launch — a model without a same-day inference path is a paper.

    link →

$ bigger_picture

The trend behind the headlines

Local agents flip the threat model rather than removing it. Ship an agent onto a laptop with deep access to files, calendar, and messages, and the vendor-side safety layer, abuse monitoring, and audit trail all go away with the API call — a 28.4% AgentDojo injection rate means roughly one in four crafted attacks lands, on a model whose entire pitch is broad local access to personal context. Apache 2.0 gives you the freedom to deploy and, in the same clause, the whole of the liability; under the EU AI Act high-risk provisions now in force, you're the deployer of record. That's not an argument against local weights — it's an argument that the differentiator was never the benchmark. It's whether you can prove what your agent touched, on whose authority, and with what scope.

$ tool_of_the_day

Unsloth's Muse Glimmer guide

The fastest honest path from weights to a working local agent. Dynamic quants get the model into an 18GB envelope on Mac or GPU/CPU setups, and the same toolchain fine-tunes it, which matters if you want a narrow agent that reliably does one workflow instead of a generalist that occasionally does everything. Day-zero llama.cpp support means you're not waiting on a conversion script.

explore Unsloth's Muse Glimmer guide →

$ subscribe --free

Get the next edition in your inbox

// DAIly Agent covers agentic AI three times a week

// free, no spam, exit anytime