Meta released Muse Spark Open on October 15, 2026, an open-weight AI model the company calls “the most capable US open-weight model to date.” The model carries 816 billion total parameters with roughly 42 billion active per token, a mixture-of-experts design that lets it scale up capacity without forcing every query through the full parameter count. It ships with a context window of up to 1 million tokens and accepts text, images, audio, and video as input, producing text as output.
The release lands at a moment when open-weight models from China, including Kimi and GLM-family systems, have been closing the gap with US labs on cost and coding benchmarks. Meta’s framing, repeated across its own channels and in day-one coverage from Hugging Face, positions Muse Spark Open as a direct answer to that pressure: a model built specifically for coding and long-horizon agentic work rather than general chat.
What Meta Actually Shipped on October 15
The headline figure is scale. At 816 billion total parameters, Muse Spark Open sits near the top of the publicly disclosed open-weight field. But the number that matters more for anyone planning to run it is the 42 billion active parameters per token. That gap between total and active capacity is the entire point of a mixture-of-experts architecture: the model routes each token to a small subset of specialized sub-networks rather than computing through every parameter every time.
According to the model documentation hosted on OpenRouter, Muse Spark Open uses 256 experts per layer and selects 8 of them for any given token. Weights are published in two formats, BF16 for full-precision serving and NVFP4 for lower-memory deployment, giving hosting providers flexibility in how they trade off cost against output quality.
Context length is the other headline number. OpenRouter lists the window at 1,048,576 tokens, enough to hold a sizable codebase, a long document set, or hours of transcribed audio in a single pass without chunking. Meta pairs that long context with native multimodal input, meaning a developer can hand the model a screenshot, a short video clip, or an audio recording alongside code and get a single coherent response back.
Why Meta Built This for Code, Not Chat
Meta’s own announcement, posted through its AI at Meta account, frames the release squarely around software work. The company wrote that Muse Spark Open “excels at long-horizon tasks, complex software engineering, and multi-step reasoning,” a description that reads more like a pitch to engineering teams than to casual chatbot users. That framing matters because it signals where Meta expects the model to actually get used: inside agent loops that read a repository, plan a change, write code, run tests, and fix what breaks, over and over, without a human steering every step.
Hardware partners picked up on the same framing almost immediately. Intel, which confirmed day-zero support for the model, described Muse Spark Open as “a coding-optimized open-weight model” built to deliver “higher first-attempt accuracy, reliable tool calling and deep codebase understanding across the repositories developers use.” That kind of fast hardware-vendor endorsement, landing within hours of the model’s release, suggests Meta coordinated the launch with its infrastructure partners well before going public, rather than dropping the weights cold.
The open-source inference engine vLLM echoed the same positioning in its own technical writeup, describing the release as “an open-weight sparse mixture-of-experts (MoE) model, built for long-horizon agentic coding workflows.” When three independent parties, a chip maker, an inference framework, and Meta itself, converge on identical language within a day of launch, it is a reasonable signal the coding and agent framing is not just marketing. It is the design brief the model was actually built against.
Day-One Hosting: Seven Platforms, One Price Floor
Meta did not just publish weights and walk away. The company lined up API access across seven hosting platforms on day one: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. That is an unusually wide simultaneous rollout for an open-weight release, and it means developers get a real choice of provider from the first hour rather than waiting weeks for third-party support to catch up.
Pricing is consistent across most of the listed providers. OpenRouter, which aggregates several of the hosts and auto-routes requests among them, lists the model at $1.25 per million input tokens, $4.25 per million output tokens, and $0.15 per million cached-read tokens. That cached-read rate is worth paying attention to: for agentic workloads that repeatedly reuse large chunks of context (a codebase, a long system prompt, a running conversation history) cache pricing can end up mattering more than the headline input rate.
| Provider | Day-one access | Published rate (per OpenRouter listing) |
|---|---|---|
| Baseten | Yes | $1.25/M input, $4.25/M output, $0.15/M cached |
| Fireworks AI | Yes | $1.25/M input, $4.25/M output, $0.15/M cached |
| Groq | Yes | Not publicly listed on OpenRouter at launch |
| Modal | Yes | Not publicly listed on OpenRouter at launch |
| Nebius | Yes | $1.25/M input, $4.25/M output (cache rate not shown) |
| OpenRouter | Yes (aggregator) | $1.25/M input, $4.25/M output, $0.15/M cached |
| Together AI | Yes | $1.25/M input, $4.25/M output, $0.15/M cached |
That pricing undercuts most frontier closed-model APIs by a wide margin, though it sits above the rock-bottom rates some pure open-weight communities have pushed smaller models to. For a model this size, the per-token price is arguably more about Meta subsidizing adoption than reflecting the raw compute cost of serving an 816-billion-parameter system across seven providers simultaneously.
The Muse Model License: Open-Weight, Not Unconditionally Open Source
Muse Spark Open ships under a license Meta calls the Muse Model License, distinct from the Llama community licenses the company has used for prior releases. Meta has not published exhaustive commercial terms alongside the launch, and outside legal review of the license text is still catching up as of this writing. Companies planning to fine-tune, redistribute, or embed the model inside a commercial product should read the license directly rather than assume it behaves like a fully permissive open-source release. “Open-weight” describes how the model is distributed, not automatically how it can be used.
That distinction has tripped up developers before with other large-model releases, where downloadable weights came bundled with field-of-use restrictions, user-count thresholds, or attribution requirements that only became obvious once legal teams read the fine print. Given how aggressively Meta is pushing hosted access through seven partners, the practical path for most developers will be renting inference through one of those providers rather than self-hosting an 816-billion-parameter model, which sidesteps a lot of the license complexity around redistribution anyway.
How Muse Spark Open Fits the Open-Weight Field
Meta’s claim that this is “the most capable US open-weight model to date” is a specific, bounded statement: capable among US-developed open-weight models, not a claim of beating every model on every benchmark globally. Independent, verified head-to-head benchmark comparisons against Llama’s prior generations, DeepSeek, Qwen, Kimi K2, GLM, or Mistral Large were not available at launch. What is available is scale and strategy, and on those two axes the positioning is clear.
Other large open-weight MoE releases in recent months give some sense of scale. Mistral’s Large 4 “Le Chonk” launched at 675 billion total parameters with 41 billion active, a configuration close in spirit to Muse Spark Open’s 816B/42B split. Tencent’s Hy4 model landed at 770 billion parameters and edged out GLM-5.3 on internal comparisons. Moonshot’s Kimi line has leaned hard into coding, with Kimi K3 posting a 61-Elo-point lead over GPT-6.1 Sol on a coding leaderboard. And Reflection AI’s Beam, built at roughly 501 billion parameters, has pitched itself as a lower-compute alternative rather than a scale play.
| Model | Developer | Reported scale | Primary positioning |
|---|---|---|---|
| Muse Spark Open | Meta | 816B total / 42B active | Coding, long-horizon agents, multimodal input |
| Large 4 “Le Chonk” | Mistral AI | 675B total / 41B active | General-purpose MoE, European open-weight push |
| Hy4 | Tencent | 770B total | Reasoning, narrowly ahead of GLM-5.3 internally |
| Kimi K3 | Moonshot AI | Not disclosed in full | Coding leaderboard performance |
| Beam | Reflection AI | ~501B total | Lower-compute efficiency claim |
What stands out from that table is not that Muse Spark Open is dramatically larger than everything else. It sits roughly in the same neighborhood as Mistral and Tencent’s recent flagships. The differentiator Meta is actually selling is the 1-million-token context window combined with native audio and video input, a combination few of the other open-weight releases in this list have matched at launch.
A Geopolitical Subtext: Racing China’s Open-Weight Push
Coverage of the launch has leaned into a framing that goes beyond specs. The Straits Times reported that Meta unveiled the model explicitly to rival Chinese competitors, a framing that lines up with a broader pattern this year of US labs citing competition from Chinese open-weight releases as justification for opening up their own flagship systems. Reports on China’s open-weight momentum, including recent industry newsletters tracking the gap between US and Chinese open releases, have noted that Chinese labs have been shipping competitive open-weight models faster and at lower declared training cost than many US counterparts.
That dynamic gives Muse Spark Open a strategic weight beyond its technical specs. If Meta genuinely wants developers building agentic coding tools to default to a US-origin open-weight model rather than a Chinese one, giving away 816 billion parameters’ worth of capability at $1.25 per million input tokens is a meaningful way to make that case. Whether it works depends less on the launch headlines than on whether third-party benchmarks over the following weeks back up the “most capable” claim once independent testers get access.
From Llama to Muse: Meta’s Shifting Open-Model Strategy
Muse Spark Open did not appear out of nowhere. Meta built its entire open-model identity on the Llama line, starting with Llama 1’s research-focused release and scaling through Llama 3’s widely adopted developer ecosystem. Llama 4 pushed that family toward larger multimodal mixture-of-experts systems, and Meta’s Behemoth project represented an attempt at a genuinely frontier-scale model, though it was never broadly distributed the way earlier Llama releases were.
Muse Spark Open represents a branch away from the Llama name entirely. According to reports tracking Meta’s broader Muse model family, the company has run several internal variants under names like Wonder and Full as closed systems before converting the Spark variant into this public, open-weight release. That naming split (keeping Llama as one lineage and building a separate Muse lineage aimed specifically at agentic coding) suggests Meta sees enough daylight between general-purpose chat models and specialized coding agents to justify running two distinct model families rather than folding everything into one.
The license change reinforces that read. Moving from the Llama community license to a new Muse Model License is not a cosmetic choice. It gives Meta room to set different commercial terms for its coding-agent line than it has historically set for Llama, without having to renegotiate the terms of an already widely deployed license.
Hardware Vendors Move Fast on Day-Zero Support
One of the more telling signals from launch day was how quickly hardware and infrastructure vendors confirmed compatibility. Intel’s newsroom post went up alongside the model’s own launch window, not days later, describing the model in terms nearly identical to Meta’s own pitch. The inference engine vLLM published a same-day technical blog describing the model’s MoE routing and serving considerations, documentation that normally takes engineering teams at least several days to produce after a model drops cold.
The Washington Times reported on claims that the model does not require a traditional data center to run, a framing that lines up with secondary reporting (not Meta’s own official technical specifications) suggesting the model can run across two paired Mac Studio desktops. That claim has circulated widely since launch, but it has not been confirmed through Meta’s own published benchmarks or hardware requirements documentation, and readers should treat it as an interesting but unverified data point rather than an official spec.
curl https://openrouter.ai/api/v1/chat/completions \
-H "Authorization: Bearer $OPENROUTER_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "meta/muse-spark-open",
"messages": [{"role": "user", "content": "Refactor this function for readability"}]
}'That kind of simple API call is the realistic entry point for most teams wanting to try the model this week, rather than attempting to self-host 816 billion parameters on local hardware regardless of how it is partitioned across NVFP4 checkpoints.
What Developers Can and Cannot Verify Right Now
It is worth being precise about what is established fact versus what is still an open claim less than 48 hours after launch. Confirmed: the parameter count, the context window, the modality support, the MoE routing structure, the checkpoint formats, and the list of seven hosting partners. Not yet independently confirmed: training data composition, total training cost, standardized third-party benchmark scores, minimum hardware requirements for self-hosting, and the full text of commercial licensing terms beyond the license’s name.
That gap matters for anyone deciding whether to build on top of Muse Spark Open this week versus waiting for independent evaluation. Early coverage from outlets like Baseten’s model library focuses on serving configuration and deployment mechanics rather than benchmark scores, which is consistent with a launch where hosting infrastructure was ready well ahead of independent third-party testing.
Market and Industry Reaction So Far
The reaction in the first 48 hours has centered on three threads: the scale of the release, the breadth of day-one hosting, and the licensing terms. Independent analyst commentary on stock-price impact, benchmark leadership, or broad commercial adoption had not materialized in verifiable form as of this writing, and most early coverage has repeated Meta’s own specifications rather than offering independent testing results. That is a normal pattern for a model released less than two days ago. It also means the “most capable US open-weight model” claim remains Meta’s own characterization until outside benchmarking firms publish comparative numbers.
Documentation from Hugging Face’s Transformers library already describes the model architecture in neutral technical terms, confirming that “Muse Spark Open is a multimodal mixture-of-experts (MoE) model developed by Meta that accepts text, images, audio, and video as input and generates text.” That kind of fast library integration (landing in official Transformers documentation within a day of release) is itself a signal of how much coordination Meta put into the rollout before the public announcement.
Context: Anthropic and the Open-Weight Risk Debate
Meta’s push into open-weight coding agents arrives against a backdrop of ongoing debate over whether releasing powerful model weights publicly creates safety risk. An Anthropic executive recently warned about the risks of open-weight AI while stopping short of calling for a ban, and the UK’s AI Security Institute has been running evaluations across more than 55 models, including open-weight releases, to assess breach and misuse risk. Muse Spark Open’s launch does not settle that debate, but it does add one more large, capable, publicly downloadable model to the pool regulators and safety researchers are tracking.
Meta has not detailed specific safety evaluations run against Muse Spark Open ahead of release in its public announcement materials, which is consistent with the sparse disclosure pattern around training data and benchmarks noted earlier in this piece. Companies planning to deploy it in production agentic workflows should expect their own red-teaming to carry more weight than anything in Meta’s launch materials for now.
Five Predictions for the Weeks Ahead
- Independent benchmark results (SWE-bench-style coding evaluations, long-context retrieval tests) will start appearing within one to two weeks, and they will likely show Muse Spark Open competitive with but not dominant over Mistral’s Large 4 and Tencent’s Hy4 on narrow tasks.
- Expect at least one additional hosting provider beyond the seven named at launch to add support within 30 days, following the pattern of rapid provider expansion seen with prior large open-weight drops.
- Scrutiny of the Muse Model License will intensify as legal and developer-relations teams at mid-size AI companies publish their own interpretations, particularly around redistribution and fine-tuning rights.
- Chinese labs will likely respond with their own coding-focused releases or updated benchmark claims within the quarter, continuing the back-and-forth dynamic that has defined open-weight competition through 2026.
- The Mac Studio self-hosting claim will either get formally confirmed with Meta-published hardware specifications or quietly fade from coverage as unverified, with no clear in-between outcome given how specific and testable the claim is.
Why This Matters Beyond the Spec Sheet
Strip away the parameter counts and Muse Spark Open’s real significance is about distribution strategy. Meta is not just publishing weights on Hugging Face and hoping developers find them. It coordinated seven hosting partners, got a chip maker to publish same-day support documentation, and landed in an official inference library’s documentation within 24 hours. That is closer to a coordinated platform launch than a traditional research-paper-style model drop, and it signals Meta treating this release as core infrastructure strategy rather than a side research project.
For developers building agentic coding tools, the practical takeaway is straightforward: there is now a free-to-download, commercially hostable, 1-million-token-context model purpose-built for long-horizon coding work, priced competitively across seven providers. Whether it displaces GPT-6-class closed models or Chinese open-weight rivals in actual production pipelines will depend on benchmarks nobody has published yet, not on the scale of Tuesday’s announcement.
Frequently Asked Questions
What is Muse Spark Open?
Muse Spark Open is an open-weight, multimodal mixture-of-experts AI model released by Meta on October 15, 2026. It has roughly 816 billion total parameters, with about 42 billion active per token, and accepts text, image, audio, and video input while producing text output.
How much does it cost to use Muse Spark Open?
Through OpenRouter, the model is priced at $1.25 per million input tokens, $4.25 per million output tokens, and $0.15 per million cached-read tokens. Rates are similar across Baseten, Fireworks AI, Together AI, and Nebius.
Where can I access Muse Spark Open?
Meta announced day-one API access through seven platforms: Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter, and Together AI. Weights are also downloadable for self-hosting through Hugging Face in BF16 and NVFP4 formats.
Can Muse Spark Open really run on two Mac Studio desktops?
That claim has circulated in secondary reporting but has not been confirmed through Meta’s own official technical specifications. It should be treated as an unverified report rather than a confirmed hardware requirement.
Is Muse Spark Open better than DeepSeek, Qwen, or GLM?
There is no independently verified head-to-head benchmark comparison available yet. Meta’s claim is limited to calling it the most capable US open-weight model, not a claim of beating every global competitor on every benchmark.
What license does Muse Spark Open use?
It ships under a new Muse Model License, separate from the Llama community license Meta has used for prior model families. Full commercial terms beyond the license name had not been exhaustively published as of this article.
What is Muse Spark Open designed to be used for?
Meta built it specifically for coding and agentic work, including long-horizon tasks, complex software engineering, and multi-step reasoning, rather than general-purpose chat.
How does Muse Spark Open’s mixture-of-experts design work?
The model uses 256 experts per layer and routes each token to 8 of them, which is how it reaches 816 billion total parameters while only activating about 42 billion per token during inference.