Free to Download Was Never Free to Run. Meta Says 24GB Changes That.

Two weeks ago, Moonshot published the weights for Kimi K3, and the takeaway from my post was one line: free to download is not free to run. Moonshot's own blog recommended running that model on supernode setups with 64 or more accelerators, so "open" meant open to inspect, not open to run at home.

Today Meta released a model aimed at the other end of that sentence.

What shipped

Muse Glimmer is a 30-billion-parameter open-weight model from Meta Superintelligence Labs, released today under the Apache 2.0 license. Per Meta's model card on Hugging Face, the full count is 29.6 billion parameters, including a roughly 1.8 billion parameter vision encoder. It takes text and images in, produces text, supports more than 100 languages, and lists a context window of 131,072 tokens.

Per Meta's research blog, Glimmer was trained by distillation from Muse Spark, Meta's larger proprietary model. Its pre-training ran on Muse Spark's outputs, then Meta added longer-context, agent-heavy data and finished with supervised fine-tuning plus reinforcement learning.

Two context points. Per VentureBeat, this is Meta's first fully open release since April, when it replaced its open-weight Llama family with the proprietary Muse Spark. And Apache 2.0 is more permissive than the old Llama license ever was, since Llama's license carried restrictions like a cutoff for companies above 700 million monthly users. Also per VentureBeat, what is open here is the weights, the quantized variants, and the vision encoder. Meta has not released the training data or the training code.

The laptop claim has a memory column

At full precision, a 30-billion-parameter model needs over 55GB of memory, per Meta's blog, which is more than any consumer GPU offers. So Meta shipped two roughly 4-bit versions.

• K-Quant-17GB targets machines with 24GB of VRAM, like an RTX 3090 or RTX 4090, with an average accuracy loss of 1.0 percent across 15 benchmarks.
• K-Quant-Dynamic targets 32GB machines like the RTX 5090, with a 0.2 percent loss.

Those degradation numbers are Meta's own measurements, not independent testing.

The Mac version of the story, per VentureBeat: Apple Silicon's unified memory plays the role of VRAM, so a MacBook Pro or Mac Studio with 32GB or more can hold the full stack. Meta ran its speed tests on M4 Max and M5 Max MacBook Pros.

When Alexandr Wang, Meta's chief AI officer, wrote on X that the model "can run on 24GB of VRAM without losing agentic reliability," he was describing the K-Quant-17GB configuration above. The part that tends to drop out of the "runs on a laptop" headline is the other one: per VentureBeat, a typical 8GB or 16GB laptop is still out of reach.

Meta also ships a small companion model called DFlash, which proposes blocks of 16 tokens at a time that the main model verifies in parallel. Per VentureBeat, that raises generation speed on an RTX 5090 from 74.9 tokens per second to 233.4. On an M5 Max it goes from 26.6 to 50.2. On an M4 Max, 23.7 to 37.8.

What Meta's own benchmark table says

Glimmer lands in a size class that already has two open models in it: Google's Gemma 4 31B and Alibaba's Qwen3.6-27B. The comparison table in Meta's materials is Meta's own.

Per that table, Glimmer leads both of them on several agentic tests: MCP Atlas at 75.5, DeepSearch QA at 74.6, and SWE-Bench Pro at 51.2, against 36.9 for Gemma 4 and 50.2 for Qwen.

Qwen leads on OSWorld-Verified, 75.6 against Glimmer's 65.9, and on TerminalBench 2.1, 60.7 against 51.7, plus most of the multimodal tests. On SWE-Bench Verified, Glimmer scored 76.0 against Qwen's 77.2.

On the safety side, per VentureBeat, Meta evaluated Glimmer under its Advanced AI Scaling Framework and determined it does not meet the framework's definition of Frontier AI. Meta's Preparedness Team assessed it at Moderate or lower risk across chemical and biological, cyber, and loss-of-control categories. On a prompt-injection test called Siren AgentDojo, 28.4 percent of attacks succeeded against Glimmer, against 25.6 percent for Gemma and 40.3 percent for Qwen.

The essay, and the money behind it

The model landed alongside a 6,500-word essay from Mark Zuckerberg called "The Future is for Everyone," published Monday on Meta's newsroom. Per CNBC, The Verge, and Fortune, the essay frames Meta's open-weight push as an answer to Chinese labs like DeepSeek, Moonshot, and Alibaba, and argues that Washington should remove hurdles for American open models rather than ban foreign ones.

"Foreign labs currently hold several advantages here since American labs have to comply with many additional restrictions on training data," Zuckerberg wrote. "US policy must reduce this additional friction if we want American open source models to lead over time."

He also wrote that he does not believe restricting access to foreign open source models is an effective solution. And in what read as a shot at the closed-model labs, he wrote: "The notion that AI is so dangerous that the only safe path is an extreme concentration of power seems inherently problematic."

Per CNBC, Zuckerberg said in an Instagram video on Monday that Meta will also open the weights for Muse Spark 1.2, its current flagship, soon. CNBC reported that Meta shares were up 2.1 percent in premarket trading Monday and are down around 10 percent this year, as investors scrutinize a capital expenditure forecast of up to $145 billion for this year.

What it looks like from the agent side

The model card names two agent frameworks specifically: OpenClaw and Hermes Agent. Hermes is what I run my own agents in, so that line caught my eye. Per Ollama's blog, Glimmer is available on Ollama on day one, and the documented command for pointing Hermes at it is one line: ollama launch hermes --model muse-glimmer:30b-mlx. Per VentureBeat, support is rolling out this week through LM Studio, vLLM, SGLang, Together AI, Fireworks AI, and OpenRouter, with llama.cpp, MLX, and ExecuTorch integrations landing in the coming days per Meta's blog.

There is no Meta API price attached to the downloadable model. The cost is whatever hardware you run it on, or whatever a third-party host charges.

Two weeks ago, the stated floor for running a serious open agent model at home was a supernode with 64 accelerators. Today it is a 24GB graphics card or a 32GB Mac, with a published spec sheet and an Apache 2.0 license file next to it. That is what changed on August 10.