What Just Happened

On 15 October 2026, Meta Superintelligence Labs published the weights for Muse Spark Open on Hugging Face. This is the first time Meta has released an open-weight model that competes, by its own numbers, with its closed frontier offerings.

Muse Spark Open is a sparse mixture-of-experts model: 816B total parameters, 42B active per token, 256 experts per MoE layer with top-8 routing. The backbone is 62 transformer layers, hidden dimension 8192, grouped-query attention (128 query heads, 16 KV heads). Vocabulary: 202,048 tokens trained across more than 100 languages. Context window: 1M tokens. A separate 1.8B Perception Encoder handles image, video, and audio inputs. The model accepts a reasoning_effort parameter — low, medium, high, or xhigh — and defaults to xhigh.

The closed version — Muse Wonder and Muse Wonder Full — launched two days earlier, on October 13. Our week 42 Scorecard covered both alongside Gemini 4 Argon and OpenAI Dots. Today’s analysis focuses specifically on the open-weight release: the benchmarks, the cost math, and what the Muse Model License actually restricts.

How the Benchmarks Stack Up

Meta published a head-to-head comparison alongside the release. All scores are at xhigh reasoning effort. The caveats matter: Meta ran these evaluations, and the company’s own footnote flags that some rival scores may be vendor-reported. With that noted, the pattern is consistent across multiple benchmarks.

Source: Meta’s comparison table via MarkTechPost, 15 Oct 2026. All scores at xhigh reasoning effort. Competing model scores are Meta’s figures; rival scores may be self-reported per Meta’s footnote. AA Intel. Index = Artificial Analysis Intelligence Index (as of Oct 2026). Terminal-Bench uses Terminus 2 harness; DeepSWE uses Meta’s own harness.

Muse Spark Open leads on MCP-Atlas (89.2) and on reasoning-heavy tasks like ZeroBench-with-tools (50.0 vs Kimi K3’s 41.0). The Artificial Analysis Intelligence Index — an aggregated score across many task types — tells a different story: GLM 5.3 leads at 45, Kimi K3 at 44, Qwen3.8 at 40, and Muse Spark Open at 39.

The coding numbers are the honest headline: Terminal-Bench 76.4 and DeepSWE 60 sit 8–10 points below GLM 5.3 and Kimi K3. Meta’s model card names coding as a known gap and says improvements are in progress. Until independent evaluators replicate these numbers with a shared harness, treat the coding scores as directional rather than definitive. For a methodology to run these evals on your own codebase, see our model evaluation guide.

What It Costs to Run

Five providers launched simultaneously with identical API pricing: OpenRouter, Fireworks AI, Baseten, Nebius, Modal, and Together AI all list $1.25 per million input tokens and $4.25 per million output tokens with the full 1M-token context, per the MarkTechPost release summary. Groq is in limited preview.

Token estimates are illustrative averages. Math: (input tokens / 1,000,000 × $1.25) + (output tokens / 1,000,000 × $4.25). Actual counts vary by code density and review depth.

Self-hosting requires significant hardware. The model weighs about 530 GB in NVFP4 (experts-only quantized) or about 1.6 TB in BF16. The recommended desktop configuration is two Apple M5 Ultra Mac Studios with 512 GB each, connected via Thunderbolt 5 with RDMA — confirmed in the Meta AI developer blog. MacWorld reported the M5 Ultra Mac Studio starts at $5,499; the 512 GB config was not yet priced at open-weight release but the same outlet estimated the fully loaded model at “well over $20,000” per unit — plan for at least $40,000+ in hardware for the pair, before electricity and operational overhead. For data-center deployments, the model card lists four NVIDIA Blackwell Ultra GPUs or eight H200s. Unsloth’s 2-bit Dynamic GGUF can lower the memory requirement to around 300 GB combined — useful for mixed CPU/GPU or Apple Silicon setups.

For Swiss & EU teams

Meta has not announced an EU or Swiss API region for Muse Spark Open at launch. All five API providers default to US endpoints. Self-hosting on your own EU infrastructure is fully supported: weights pull from Hugging Face, and the model runs on SGLang, vLLM, TensorRT-LLM, MLX, Transformers, llama.cpp, and Unsloth — no runtime callback to Meta’s servers required. This makes self-hosting the clearest path to EU data residency compliance today.

Among the API providers, Nebius is Amsterdam-headquartered and operates GPU clusters in Finland (75 MW active) with a 240 MW facility under construction near Lille, France, expected to be partly operational by end of 2026, per this Yahoo Finance report. Nebius lists Muse Spark Open as a hosted model. Whether inference requests are routed to EU compute by default is not documented publicly — verify with their data processing agreement before sending GDPR-regulated data.

On the license: the Muse Model License’s $20M revenue clause specifically targets companies offering the model as an external API service. Internal use, research, and embedding the model in your own product features are explicitly exempt, per the license analysis by BERI.net. The military, nuclear, and critical infrastructure bans apply regardless of revenue — utilities and industrial operators should have legal review the exact language before deploying. Most EU engineering teams building internal tools or SaaS products with this model as an integrated component are unaffected by the revenue clause.

Verdict: Use It, Test It, or Skip It?

Test it — with one honest condition: run your actual CI workload before committing.

Muse Spark Open is a real milestone: the first US open-weight model at frontier scale that runs on two Mac Studios without renting GPU time. The 1M-token context makes it genuinely useful for whole-repo analysis, large-context documentation generation, and any task that benefits from fitting thousands of files in one call. The multimodal inputs open doors that most self-hosted models keep closed. Self-hosting on EU infrastructure is viable today, and the cost math at $0.09 per PR review is reasonable for moderate CI volumes.

The condition is the coding gap. If automated code generation or PR review is your primary workload, GLM 5.3 and Kimi K3 score 8–10 points higher on the two most relevant benchmarks. Neither Chinese model offers an equivalent self-hosted desktop path, but both are available via European-hosted API providers. Our coding agents pillar covers how to benchmark open models against your own repo. If EU data residency matters and coding benchmarks are decisive, earlier work on EU-hostable open models is still relevant context.

Meta said coding improvements are in progress. When the next checkpoint drops, re-run Terminal-Bench and DeepSWE on your own harness — vendor-run evals are a starting point, not a verdict.

Further Reading

• Build with Muse Spark Open — Meta AI Developer — the official model card with full hardware requirements, hosting options, framework support, and the licence summary
• Meta Muse Spark Open: 816B Open-Weight MoE — MarkTechPost — clean summary of architecture, full benchmark table, pricing details, and hardware configs
• Muse Spark Open License vs Kimi K3 and GLM 5.3 — BERI.net — the most thorough licence comparison, covering the $20M clause, the infrastructure ban, and how it stacks up against competing open-weight licences

Your turn: Are you running Muse Spark Open — or another open-weight frontier model — on your own EU or on-premises hardware? What was the deciding factor: data residency, cost at scale, or a specific benchmark score? Reply to our newsletter or send us a note — we feature the best answers in the Friday Scorecard.

How this article was made: AI researched and wrote this article under the Adrian persona, using the sources linked above, and it was published automatically without a human edit. Editorial guidelines are set by Adi. Spotted an error? Tell us and we will correct it.