Meta releases Muse Spark Open: 816B parameters and the lead among US open-weight models

This week pulled American open models closer to the frontier. Meta has released the weights of Muse Spark Open: 816B parameters in total, 42B active per token, with a 1M-token context. According to Nathan Lambert of Interconnects, it is the strongest American open-weight model by a wide margin. Together with Reflection's first model, Beam, it has halved the gap to Chinese labs. Without these two releases, 20+ Chinese open models would be stronger than the best Western one.

Where Muse Spark Open sits in the hierarchy

According to the Artificial Analysis Intelligence Index, the model ranks above GLM 5.2 and close to DeepSeek V4.1 Flash. It does not reach GLM 5.3 (August 18) or Kimi K3 (July 16). For the market, that is enough: open LLM agents have been a practical working tool since roughly the GLM 5.2 level, and demand is growing. Inference providers are scaling quickly, an economy around open post-training is taking shape, and companies want to own their own "intelligence."

The model is natively multimodal and built for coding and agentic scenarios. This release received additional safety tuning so that it stays robust without system-level safeguards. Some observers concluded that the weights were "broken" during alignment, because they shipped after Muse Spark 1.3. Lambert thinks it is more likely a safety version of Muse Spark 1.2 that lost about one point on the index. That fits what Alexandr Wang said when he released the 30B agentic Muse Glimmer: Meta was preparing an open version of 1.2.

The license is the main brake

The main risk for adoption is a custom license that bundles almost every restriction critics have raised about other open models:

• products must display "Built with Muse," as with Llama before (almost nobody does this);
• companies with revenue above $20M per month across the whole business must sign a separate agreement with Meta;
• an acceptable use policy is mandatory, so by open-source standards the model is not open;
• the terms extend to model outputs and derivatives, which leaves synthetic data in a legal gray area;
• internal research and applications are permitted.

In practice, many will ignore the license, and Meta will probably pursue only flagrant violations. But as Lambert notes, for a lab that is winning its audience back, a complex non-standard license remains an unnecessary strategic risk. At a comparable quality-to-size ratio, Western open models draw noticeably more downloads than Chinese ones, and Meta has a chance here.

The closed frontier and the balance of power

A few days earlier, Meta unveiled the closed Muse Wonder and Muse Wonder Full. On classic benchmarks they trail only Anthropic's Opus 5.5 and are close to GPT-6-Astra and Gemini-4-Argon. Meta remains the only American company with models at both frontiers. Lambert credits the success to a flat structure: Wang has more than 200 direct reports, and releases ship consistently.

His ranking: Anthropic > OpenAI > Meta ≈ Gemini > xAI ≈ Chinese open-weight > American open-weight. He names Xiaomi V2.6 Pro the best open model overall. The field of American competitors has narrowed to five: Anthropic, OpenAI, Meta, Gemini and xAI. Microsoft could join if it starts releasing models. Third-tier players (Mistral, Nemotron, Reflection, Arcee) find it harder and harder to influence the frontier, because a flagship costs billions.

Key facts

• Muse Spark Open: 816B total / 42B active, 1M context, multimodal, coding and agents, safety tuning.
• Muse Wonder and Muse Wonder Full: Meta's closed models. The differences between them are undisclosed; they may be different pretrains.
• Muse Glimmer: a 30B agentic model.
• Beam: Reflection's first model.
• Mistral Large 4: 38 points on the index, weights not published.
• GLM 5.3 (August 18), Kimi K3 (July 16), DeepSeek V4.1 Flash, GLM 5.2: Chinese reference points.
• Xiaomi V2.6 Pro: the best open-weight model overall.
• Opus 5.5, GPT-6-Astra, Gemini-4-Argon: closed competitors to Muse Wonder.

What teams choosing a model should consider

An MoE architecture with 42B active parameters means per-token compute cost is close to that of a 40B model. But you have to store all 816B. By our estimate, in FP8 that is about 816 GB for the weights alone, which means a node of eight H200s with headroom for the KV-cache at long context. In 4-bit form, you will need about 410 GB. Self-hosting makes sense at high utilization; otherwise it is cheaper to use the model through an inference provider.

Legal teams should check the $20M-per-month threshold and the clause on derivatives in advance, especially if you distill the model or generate synthetic data. Those who need a clean open-source license are better served by alternatives with standard licenses. For those who want an American stack at acceptable quality, Muse Spark Open is now a real option, not a compromise.