What Actually Shipped
A 30B dense architecture: every parameter activates per token, with none of the routing overhead of a mixture-of-experts model—Meta's stated tradeoff is long-context coherence and predictable latency. The context window is 131K, and the model handles tool calls and image input. At full precision it needs more than 55GB of memory; the official 4-bit build comes in under 20GB, still leaving room for the KV cache, an image encoder, and a speculative-decoding drafter inside a 24GB or 32GB card. Meta uses DFlash speculative decoding, reporting a 3.1x speedup on an RTX 5090, 1.8x on an M5 Max, and 1.5x on an M4 Max. The weights are on Hugging Face, with support already in Ollama, LM Studio, llama.cpp, MLX, ExecuTorch, vLLM, and SGLang. Training ran in three phases: logit distillation from Muse Spark 1.2, then longer-context and agent-heavy data, and finally a mix of supervised fine-tuning, on-policy distillation, and reinforcement learning.
A Swing Back From the Closed Flagship
In April, Meta released Muse Spark, its first closed flagship, widely read as the end of the open path Llama had built. This time the company is not only publishing Glimmer—Zuckerberg said the weights for the stronger Muse Spark 1.2 will be opened too, and he published a long essay alongside the release arguing that AI capability should be distributed rather than concentrated in a handful of companies. Apache 2.0 is the detail that matters most here: commercial use, modification, and redistribution are all permitted, without the usage restrictions attached to the Llama line.
What to Watch
The benchmark numbers are all vendor-run so far. Meta compares Glimmer against Gemma4-31B and Qwen3.6-27B on DeepSearch QA, MCP-Atlas, and SWE-Bench, and no third-party reproduction has appeared yet. One safety number is worth tracking: on the Siren AgentDojo prompt-injection test, Glimmer shows a 28.4% attack success rate—between Gemma's 25.6% and Qwen's 40.3%—while posting the highest utility score of the three at 94.2. The Register's read was blunter: at 30 billion parameters, this does not do much to reclaim a lead in open weights.
Why It Matters
For developers who want agents running locally, the combination is a practical one: an open enough license, a model that fits on one card, and tool calling and multimodal input included—so the selection question shifts from capability to deployment cost. But an attack success rate near 30% means that once a local agent touches your filesystem or the network, sandboxing and human approval steps still are not optional. Getting it running and being able to leave it running are two different things.
via: Meta AI Research's announcement; NVIDIA's developer blog on local deployment; VentureBeat; The Register; verified 2026-08-12