Meta Releases Muse Glimmer, 30B Parameter Open License Agent Running on 24 GB GPU
Distilled model from Muse Spark arrives under Apache 2.0 on Hugging Face and Ollama. It's Meta's answer to Anthropic and OpenAI's dominance in agentic code.
Meta has released Muse Glimmer, a 30 billion parameter agentic model with open weights under the Apache 2.0 license, on Monday, August 10. It is the first model from Meta Superintelligence Labs designed from the ground up for autonomous workflows, with distribution occurring simultaneously on Hugging Face and Ollama, signaling a bet on rapid adoption by developers.
Glimmer runs on a consumer-grade GPU with 24 GB of VRAM using 4-bit quantization and the DFlash variant of speculative decoding. In this configuration, Meta disclosed in its research blog that a consumer card delivers up to 20,000 tokens per second. Running an agent at this speed without calling an external provider was, until August, a privilege of those with their own H100 cluster.
What the Benchmarks Show
In the SWE-Bench Pro, a test measuring real bug resolution in production codebases, Glimmer scored 51.2 and took the top spot in the public ranking. In SWE-Bench Verified, it reached 76.0, compared to 77.2 for Qwen3.6-27B, released by Alibaba in July. The margin is small and favors the Chinese model, but Glimmer outperforms in GAIA2, MCP-Atlas, DeepSearch QA, tau-Bench, IFBench, and AIME, all tests requiring multiple tool calls and error recovery.
Architecture choice matters. The model is dense, not a mixture-of-experts, which simplifies deployment in smaller containers and reduces latency of the initial response. According to Meta's research page, post-instruction training was oriented toward three behaviors: reliably calling tools, chaining reasoning over long horizons, and diagnosing tool call failures rather than simply aborting.
The Commercial Move Behind the Open Gesture
Meta keeps the larger model of the same family, Muse Spark, under wraps, charging $4.25 per million output tokens via paid API. By delivering Glimmer for free, it executes the classic commoditizer’s calculation: reducing the competitor's margin, gaining telemetry and developer mindshare, and reserving pricing for those needing top-tier performance. This is the same playbook Meta has used with Llama since 2023, now applied to a category where Anthropic and OpenAI still charge a premium for output tokens.
The dual distribution on Hugging Face and Ollama is nearly as significant as the model itself. Engineering teams requiring strict compliance, particularly in financial and healthcare services, gain the option to run the agent entirely within their corporate perimeter, without traffic to external APIs.
What Changes in the Architectural Radar
For European teams, Mistral loses immediate advantage. The Mistral Large 3, currently the most adopted open model among French and German companies governed by data sovereignty, performs well in French and Portuguese but falls behind Glimmer in agentic capabilities. The question that architects in Paris and Berlin will bring to the table in the coming weeks is whether the cost of staying with Mistral for regulatory reasons still justifies the performance difference.
In China, Alibaba's Qwen maintains a marginal lead in pure code benchmarks but loses ground in tools. This signals that the competition for open models has shifted from parameters to orchestration. DeepSeek and Kimi, the other two major Chinese labs, need to respond with updates to tool use in the next release to avoid being sidelined as suppliers in on-prem deployment in the West.
In Brazil, the change has a direct impact on banks and consulting firms under Central Bank regulation. Institutions operating under resolution 4.658 maintain internal restrictions on sending data to external APIs in pipelines that touch client information. Running Glimmer on a local GPU addresses compliance without sacrificing agentic capability, and the same applies to international consultancies operating internal automation labs on sensitive client data.
Compute Pricing is Now the Ceiling
The announcement sends a clear message to hyperscalers. AWS Bedrock and Azure AI Foundry charge a premium for hosting open models with managed fine-tuning. If Meta continues to release agentic generations with open weights and competitive performance, the economics of managed inference will resemble the economics of managed databases in the 2010s: prices drop, margins compress, and revenues shift to software surrounding inference. For enterprise clients, it is no longer straightforward that the cheapest route for running agents in production goes through a U.S. public cloud.