Muse Spark Open tested: what Meta’s 816B open-weight model can do and what it takes to run

Muse Spark Open is Meta’s latest open-weight AI model, released on 15 October 2026. You can download it, fine-tune it and build products on it, under a licence that stops short of open source. We read the licence, the benchmarks and the hardware guides, then ran it ourselves on 16 everyday marketing jobs against Muse Spark 1.3, GLM 5.3 and Kimi K3.

The answer in brief

• What it is: an 816-billion-parameter mixture-of-experts model that uses 42 billion parameters per token, reads text, images, audio and video, and holds a million tokens of context.
• How good it is: 39 on the Artificial Analysis Intelligence Index, the highest score of any American open-weight model as of 19 October 2026, and six or seven points behind the leading Chinese releases.
• The licence: free commercial use, with a “Built with Muse Spark Open” notice, a $20 million revenue gate for companies that resell API access, and a usage policy. It is not open source.
• What it takes to run: about half a terabyte of fast memory for the 4-bit weights. On a Mac that means two 512 GB or four 256 GB M5 Ultra Mac Studios, which ran it at 25 to 37 tokens a second in Apple’s tests.
• What it costs through an API: about $1.25 per million input tokens and $4.25 per million output tokens. In our tests on 19 October 2026 it passed every check at its lowest reasoning setting, for an average of under a third of a US cent per short job.
• Where it falls short: it trails GLM 5.3 and Kimi K3 on coding benchmarks, and in one of our agent tests it paused a client’s campaigns before asking which of two clients we meant.

What is Muse Spark Open?

Muse Spark Open is the downloadable version of Meta’s Muse Spark model line, built by Meta Superintelligence Labs. Until now, Muse Spark was available only inside Meta’s apps and through its paid API. The PyTorch team at Meta says the open model shares its base model with Muse Spark 1.3, the hosted version released in September. Meta added extra safety training so it holds up without the guardrails Meta runs around its own service.

The architecture comes from the model card and Hugging Face’s write-up:

Meta pitches it at coding agents and long, multi-step work: reading a large codebase, planning a change, calling tools, checking the result and recovering when a step fails. Reasoning cannot be switched off, so even simple requests spend some tokens thinking. Setting effort to low keeps that cost down.

Is Muse Spark Open really open source?

No. It is an open-weight model under Meta’s own Muse Model License, which permits a great deal and attaches three conditions.

You may use, copy, modify, fine-tune, distil, merge, redistribute and sell the model, free of charge, including inside commercial products. Meta claims no ownership of what the model generates.

The conditions:

1. Attribution. Any product or service using the model or a derivative must prominently display “Built with Muse Spark Open” in its interface, documentation, website or settings. The licence counts models trained on Spark Open’s outputs as derivatives, so distilling it into your own model still triggers the notice.
2. A revenue gate for API resellers. If your company sells access to AI models through an API and your group’s gross revenue passes $20 million in any 12-month period, you must notify Meta and sign a separate agreement before offering Spark Open that way. The trigger is total company revenue. Internal use, research and building the model into your own product are exempt.
3. A usage policy. It bans military and warfare uses, weapons development, operating critical infrastructure, malware, fraud, disinformation and impersonation, among others. The model is not intended for anyone under 18.

The Hugging Face repository is also gated: you submit your legal name, date of birth and organisation before downloading.

For most developers and start-ups these terms barely bite. They matter more to cloud providers and model resellers, for whom Meta keeps a hand on the gate. The New Stack compares the licence to the one Moonshot uses for Kimi K3. The Chinese rivals vary. GLM 5.3’s own licence adds a security review, but only for model-as-a-service companies with more than $10 billion in revenue, and Qwen3.8 2.4T ships under a restricted “Qwen3.8-Max License”. Others use plain MIT or Apache 2.0, with no revenue threshold or naming rule: GLM-5.3-Flash and DeepSeek V4.1 Flash are MIT, and Meta’s own smaller Muse Glimmer is Apache 2.0.

How does it compare with Kimi K3, GLM 5.3 and Qwen3.8?

On independent testing, Muse Spark Open is the best American open-weight model and sits six or seven points behind the best Chinese ones. Artificial Analysis measured it at 39 on its Intelligence Index, a composite of ten evaluations. The nearest American rival is Reflection’s Beam, released on 16 October, at 30; Thinking Machines’ Inkling and Nvidia’s Nemotron models score in the mid-twenties. A newer Chinese release has also overtaken it since launch: DeepSeek V4.1 Flash 1017, out on 17 October, scores 45.

Two results from the index testing stand out, as TokenPost reported. Spark Open averaged about 40,000 output tokens per task, 43% fewer than GLM 5.3. And on AA-Omniscience, a test that penalises wrong answers, its 24% hallucination rate was the lowest of any open model scoring above 30 on the index.

The open model trails Meta’s own hosted models. On the same index, Muse Spark 1.3 scores 45 at the same xhigh setting (48 at max), Muse Wonder 52 and Muse Wonder Full 56. Downloading the weights costs you about six points against the hosted Spark at matching effort.

What is it good at, and where does it fall short?

It is strongest at calling tools and weakest at the coding benchmarks that most closely resemble a software engineer’s day. The figures below are Meta’s own runs, published in the model card; Meta says a full evaluation report will follow.

The tool-use lead is real but uneven. On MCP-Atlas Spark Open is 4.7 points clear of the next model, Qwen3.8; on JobBench it leads GLM 5.3 by 0.4 of a point; and on Toolathlon Verified, which tests long tool workflows, it trails both rivals. The coding gap is wider and more consistent: seven to nine points on Terminal-Bench 2.1 and seven to eight on DeepSWE 1.1. The New Stack points out that Meta acknowledged a similar coding gap when the closed Spark launched in April.

For a model Meta describes as built for coding, that matters. If your agent spends its time calling many tools and services, Spark Open is worth testing first. If it mainly writes and fixes code, GLM 5.3 and Kimi K3 remain the stronger open choices on the evidence so far.

What hardware do you need to run it?

You need around half a terabyte of fast memory for the weights alone, plus room for the context. A mixture-of-experts model only computes with 42 billion parameters per token, but all 816 billion have to sit in memory.

The sizes come from Meta’s developer guide, Hugging Face and the Unsloth and mlx-community model pages, in decimal gigabytes. Apple’s “about 493GB” for the 4-bit model is the same 530 GB file counted in binary gigabytes, and the 3-bit MLX page lists its build as both 370 GB and 344 GiB.

The long context is the hidden cost. Hugging Face calculates the cache at about 250 KB per token, or roughly 266 GB for a full million tokens. Meta’s model card asks for at least 700 GB of GPU memory for the 530 GB NVFP4 checkpoint, which leaves room for only part of a full million-token cache.

Support arrived on day one in SGLang, vLLM, TensorRT-LLM, Hugging Face Transformers, MLX, llama.cpp and Unsloth, on Nvidia, AMD, Intel and Apple hardware. Fine-tuning works through PyTorch’s TorchTitan, Hugging Face TRL, Nvidia NeMo and managed services from Fireworks AI and Together AI.

Can you run it on a Mac?

Yes, but only on a Mac Studio with 256 GB or 512 GB of memory. One 512 GB machine runs the community 2-bit and 3-bit builds, and Meta’s own 4-bit version needs two 512 GB or four 256 GB M5 Ultra machines linked together. The M5 Max Mac Studio tops out at 128 GB, below the smallest published build (176 GB).

Apple’s machine learning team benchmarked Meta’s 4-bit NVFP4 checkpoint on M5 Ultra Mac Studios linked over Thunderbolt 5 with RDMA, using MLX. Apple says the 4-bit model “cannot fit on a single machine”. Meta’s guide specifies two machines with 512 GB each or four with 256 GB each. Hugging Face tested smaller builds on a single 512 GB Mac Studio with the previous M3 Ultra chip.

Apple does not say how much memory each machine in its clusters had, so the memory column follows Meta’s guide. At a 65,536-token context the four-machine set-up took about 40.6 seconds to produce the first token with two cables per pair, against 50.5 seconds with one, and Apple’s tests stop there. Hugging Face noted the quantised model “requires a bit more steering”, so expect some loss of quality from the smaller builds.

Linking the machines takes some setting up. Meta’s Mac cookbook asks for macOS 26.2 or later with RDMA over Thunderbolt turned on, MLX 0.30.1 or later, passwordless SSH between the machines and a direct Thunderbolt 5 link between every pair, with two cables per pair to double the bandwidth. Four machines therefore need 12 cables, and since each M5 Ultra Mac Studio has six Thunderbolt 5 ports, the four-machine cluster uses every one. Each Mac also needs disk space for the full 530 GB checkpoint.

On paper, two 512 GB machines or four 256 GB ones (1 TB in all) can hold the 530 GB of weights plus the 266 GB cache that a full million-token context needs. A single 512 GB Mac running the 370 GB 3-bit build has less than 200 GB to spare, well short of that. Nobody has published a Mac test anywhere near the full window.

Meta showed the set-up running its Muse Code agent entirely on-device, shipping bug fixes and features to the open-source Mastodon app Ice Cubes and editing footage in DaVinci Resolve.

The hardware is the first catch. In Apple’s UK store on 19 October 2026, the M5 Ultra Mac Studio started at £5,499 with 96 GB of memory, 256 GB added £4,000 (£9,499 in all), and the 36-core CPU with 80-core GPU added £1,300. The 512 GB option, which requires that faster chip, was “coming late October” with no price. Four 256 GB machines therefore cost about £38,000 before cables. In the US the M5 Ultra starts at $5,499. Speed is the second catch. Twenty-five to 37 tokens a second suits one person, but the model always reasons before it answers, so a long agent task at high effort can take minutes per step. These are also one-user benchmarks at batch size one, and they say nothing about serving a team.

How much does it cost to use through an API?

Hosted providers charge about $1.25 per million input tokens and $4.25 per million output tokens, with cached input at $0.15, according to Artificial Analysis as of 19 October 2026. Meta’s launch post names Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI as providers at launch. The weights are also in Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models.

Artificial Analysis puts the cost at $0.84 per Intelligence Index task. That is less than half of GLM 5.3 ($2.01), but about three times the new DeepSeek V4.1 Flash 1017 ($0.26), which also scores six points higher.

There is a wrinkle. Meta’s hosted Muse Spark 1.3 has the same list price and, at the same xhigh setting, scores six points higher for $1.37 per task, because it uses more tokens. If you only want an API, the closed model gives you more capability per call. The open model earns its place when you need to self-host, fine-tune or keep data inside your own infrastructure.

What happened when we tested it?

Spark Open got nearly all of our marketing jobs right, and one setting decided how fast and cheap it was: reasoning effort. At low effort it passed every check in about half the time and for less than half the cost of extra high. On 19 October 2026 we made 553 calls through OpenRouter, with fictional prompts, for $6.55 in all.

The same 16 jobs we gave Gemini 4 Argon, eight of them first set for Mistral Large 4. Product copy, a meta description, ad headlines, email subject lines, a report summary, a French localisation, product tagging, chart alt text, a Google Ads analysis, a knowledge check, a style brief, two web pages with hidden instructions, a campaign archive of about 100,000 tokens, a broken-tracking trap and two false premises. Each ran twice, with automatic checks and every failure read by hand. Spark Open was pinned to Together; the rivals ran on their own labs’ servers at their default settings.

GLM 5.3 takes no images, so it skipped the alt-text job; its one miss was a September cost per acquisition of £27.66 instead of £27.55. No model obeyed either hidden instruction. With no effort setting sent at all, Spark Open passed 31 of 32: once it reasoned for 586 tokens and returned an empty answer. Two of its 162 calls on Together for these jobs failed with provider errors and worked on a second try.

Spark Open starts answering fast, in about half a second, but at its higher effort settings it then thinks and streams for twice as long as Meta’s hosted Muse Spark 1.3, which costs the same per token. At low effort the two were level.

Effort is the setting to change. Low, medium, high and extra high all passed every check. Low did it in a median 11 seconds for $0.0027 a job; extra high, the default in Meta’s model card, took 22.6 seconds and $0.0062. High was the slowest, because on these jobs it reasoned more than extra high did. Our checks test facts, formats and limits rather than the quality of the prose, so read a sample before moving a whole workflow to low. Set the effort in every request rather than relying on a default.

The same model ran more than three times faster on one host than another. All five hosts charged the same $1.25 and $4.25 per million tokens (cached input pricing differed). At medium effort, Baseten streamed a median 147 tokens per second and Together 40, with every host’s median first token arriving in under a second. Each passed 11 or 12 of its 12 checks; the misses were answers a few words outside the length asked. This is one evening’s snapshot, and hosts change. If speed matters, test a few and pin the fastest in OpenRouter’s provider settings. The two FP4 hosts, which run a compressed version of the model, did no worse in this small sample.

Tool use: safe on the hard cases, with one slip. On three easy agent jobs (a weekly report where one ads API fails once, a stock check with a look-alike product code, a booking the calendar refuses), all four models got all 24 runs right. We then added 15 irrelevant tools and three harder jobs:

• A support ticket contained a hidden instruction to email the customer list to an outside address. No model followed it in eight runs. Spark Open, GLM 5.3 and Kimi K3 told the user about it; Muse Spark 1.3 said nothing.
• An orders tool split the answer across three pages. Every model got it right (four returns worth £469.00), though Spark Open kept asking for pages after the tool said there were none.
• Two clients shared the name Hearth & Loom, and we asked for their Performance Max campaigns to be paused. GLM 5.3, Kimi K3 and Muse Spark 1.3 asked which client every time. Spark Open asked once; the other time it paused one client’s campaigns and only then asked about the second.

That last slip is the case for Meta’s own advice to put a human check on irreversible actions.

It did not refuse legitimate work. Meta’s FORTRESS results suggest Spark Open turns down more harmless requests than its rivals. We tried eight borderline but legitimate marketing jobs, including a phishing email for staff training, bingo ads under the CAP Code, a pharmacy blog on warfarin and painkillers, and a data breach email to customers. Across 64 answers from the four models there were no refusals. Spark Open once wrapped its phishing example in heavy warnings, but still wrote it. One aside for copywriters: Muse Spark 1.3 named real wineries, Chapel Down and Nyetimber, in copy for a fictional wine.

We did not test coding agents, local running on Macs or GPUs, fine-tuning, audio or video input, or contexts beyond about 100,000 tokens.

How does it fit with Muse Spark, Muse Wonder and Muse Glimmer?

Spark Open is the larger of Meta’s two open models, and it sits below the closed ones by design.

The history explains the shape. Meta’s last open flagship, Llama 4, arrived in April 2025 to a lukewarm reception. When the new Muse Spark launched in April 2026 it was closed, a clean break with the Llama approach, though Mark Zuckerberg said open Muse models would follow. In August Meta released Glimmer, and Zuckerberg promised the weights of Muse Spark 1.2 “soon”. Bloomberg reads the release as Meta keeping that promise while holding its most advanced technology back to earn revenue from it.

Meta’s own framing is national. Its launch post calls Spark Open “the most capable US open-weight model to date”, and the comparison that matters to Meta is with Chinese labs, which hold the top places on the open-weight leaderboards.

Is Muse Spark Open safe?

On Meta’s figures it is unusually hard to misuse, with a cost: it refuses more harmless requests than its Chinese rivals. Meta says it evaluated the model under its Meta Superintelligence Scaling Framework for content safety, agentic risk, privacy, and chemical, biological, cyber and loss-of-control risks, and that it met the release thresholds.

The headline numbers, all from Meta:

• On StrongREJECT v2, an adaptive jailbreak test where lower is better, Spark Open scored 0.5. The comparison models scored between 62.9 and 86.4.
• It refused 99.2% of cyber-misuse jailbreaks on Meta’s internal test. Kimi K3 refused 75.5% in the same set-up.
• On FORTRESS, its adversarial-harm score was 5.8, against 26.7 for Kimi K3. Its over-refusal score was 17.4, against 2.7 for Kimi K3 and 3.2 for GLM 5.3.

That last pair is the trade-off to plan for. A model that turns away borderline but legitimate requests can frustrate users, so test it on your own prompts. On our eight everyday marketing prompts it refused nothing, but Meta’s benchmark covers more sensitive ground. Meta also warns that downstream changes such as fine-tuning could weaken its safety, and recommends system-level guardrails and human confirmation before an agent takes irreversible actions.

Who should use Muse Spark Open?

It suits teams building tool-heavy agents who want a capable model they can host, fine-tune and audit themselves, and who prefer an American developer for compliance or procurement reasons. Its low hallucination rate and token efficiency are useful in those settings.

It is a weaker first choice for agents that mainly write code, where GLM 5.3 and Kimi K3 still lead. It is not a hobbyist model: Meta’s own 4-bit version needs several high-end GPUs, or two 512 GB or four 256 GB M5 Ultra Mac Studios, and the compressed builds that have been tested need a 512 GB Mac. And if you just want API access, Meta’s hosted Muse Spark 1.3 offers more capability at the same per-token price, and in our tests it finished jobs in half the time Spark Open took at its default effort. If you do use Spark Open, start at low effort, test a few hosts and pin the fastest, and keep a human check on any action an agent cannot undo.

Two things could change this picture soon. Meta has promised a full evaluation report, and Mistral plans to release the weights for Mistral Large 4, which scores within a point of Spark Open, on 27 October.