Meta's open-weight U-turn: Muse Glimmer targets Llama's old market

Meta has handed the open-weights market back to itself. On 10 August it published Muse Glimmer, a 30-billion-parameter model licensed under Apache 2.0, reversing the closed-model strategy it adopted earlier in 2026 when it retired the Llama line for the proprietary Muse Spark. The move puts a free, commercially usable model directly in the path of Qwen and Gemma, both of which Meta claims to beat on its own benchmark.

Muse Glimmer was distilled from Muse Spark 1.2, Meta's closed flagship launched just five days earlier, and built by Meta Superintelligence Labs. Full precision weights run to 59.6 GB, but a 4-bit version under 20 GB fits on a 24 or 32 GB consumer GPU or a Mac, with Meta reporting under 1% accuracy loss versus full precision. A bundled "DFlash" drafter for speculative decoding delivers reported speed-ups of 3.1x on an RTX 5090, 1.5x on an M4 Max and 1.8x on an M5 Max.

The model is pitched squarely at developers building local AI agents rather than chatbot users. It handles text and images through an 1.8-billion-parameter vision encoder, supports context windows above 131,072 tokens, and is trained specifically for multi-step tool use, function calling and recovering when an API call fails. On Meta's MCP Atlas benchmark it scores 75.5, ahead of Gemma 4 31B's 54.2 and Qwen 3.6 27B's 62.5, though its Artificial Analysis Intelligence Index score of 35 is comparatively modest.

Mark Zuckerberg confirmed the release directly, saying: "Today we're also opening the weights for Muse Glimmer, a great 30B parameter dense model that can run locally." He also said Meta will open-source Muse Spark 1.2 itself, extending the reversal beyond a single distilled model, according to gHacks.

Meta's retreat from Llama earlier this year looked like a bet that proprietary models were the better business. Muse Glimmer suggests that bet came with a cost: ceding the open-weights market to Qwen and Gemma was a gap competitors filled fast, and Meta wants it back. Giving away a distilled version of its new flagship, rather than a standalone open model, lets Meta capture developer mindshare without fully exposing Muse Spark's commercial edge.

The real signal is the promise to open Muse Spark 1.2 itself. If Meta follows through, it undercuts the pricing power of every closed-model vendor competing for agent-building developers, and forces Qwen and Gemma to defend share against a free model backed by benchmark claims on the exact agentic tasks enterprises are buying for.