Muse Spark Open

# Fix list: Muse Spark Open

From Anchor Terminal's listing at https://www.anchorterminal.com/tools/muse-spark-open, the October 2026 research run, assessed 15 October 2026. Grade D, 53.4 out of 100.

This is everything the published grade says the listing lacks, the biggest possible gain to the total first. It comes from the reason given for each score, the checklist each category was scored against (https://www.anchorterminal.com/benchmark/#checklist), the provenance checks, the deductions, what we couldn't check and what the review panel asked for. A fix counts at the next check, once it's public.

For a coding agent working on Muse Spark Open: work through the items below in the product, its docs and its public pages. Each category gives the reason for its score, with the points each checklist item earned, and the checklist itself, so the gap is the items that earned less than their points. Change the product, not the wording, and keep a note of what you changed and where it's published.

## 1. Reliability, 42 out of 100, up to 11.6 more on the total

Why it scored 42: Read with the local-software lines, since the owner serves the weights. Meta's weights come from two Hugging Face repositories gated with manual review. The docs state Python 3.10 or later, Transformers 5.19 or later, Docker images for SGLang and vLLM and the GPUs for each checkpoint (4 × B300 for NVFP4, 8 × B300 for BF16), but SGLang's image is a dev tag and TensorRT-LLM's support sits on a feature branch (13 of 20). No CI or test suite of Meta's for the model was found. github.com/meta-models answered 403 to our fetcher, so its repositories are unchecked (0 of 25). Three Hugging Face discussions on release day, a setup question and two general posts, none a crash report and none answered by Meta when read. A day is too short to judge (15 of 25). No changelog, version tags or revision notes, and the commit history needs gated access (2 of 15). Released as a final checkpoint with no preview label from Meta, though runtime support ships as launch-specific images and a branch, and Groq lists it as a preview (12 of 15). 42.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-reliability):

Hosted APIs, MCP servers, models and platforms.

- 20, a public status page with component history (Statuspage, Instatus, BetterStack or the vendor's own).
- 0 to 30, the incident record for the last 90 days on that page. 30 for a clean record or trivial incidents only, 20 for minor incidents only, 10 for one major outage (an hour or more of a core API down, or errors across the board), 0 for several. 5 when there's no history we could read, and the note says so.
- 15, rate limits documented with numbers.
- 15, documented 429 or overload handling (Retry-After, backoff guidance), and idempotency keys or safe-retry guidance where writes are involved.
- 10, an SLA published for any paid tier.
- 10, the surface agents use is generally available, not beta or preview.

Local packages, SDKs, frameworks and stdio MCP servers.

- 20, installs from an official package with supported runtimes stated.
- 25, a public CI and test suite, passing on the default branch.
- 0 to 25, open crash or regression issues relative to activity (25 for few and handled, 0 for many, old and unanswered).
- 15, semver discipline and breaking changes called out in a changelog.
- 15, version 1.0 or later, or declared stable.

Protocols are read from their reference implementations, the public facilitators or servers, spec stability and test vectors.

## 2. Security & auth, 44 out of 100, up to 9.8 more on the total

Why it scored 44: Read as software the owner runs. No account or key to run it. Meta's weights need manual approval on Hugging Face and an access token, and server authentication is the owner's to add (12 of 30). An agent model that calls whatever tools the operator gives it. The card recommends human-in-the-loop confirmation for irreversible actions and Meta reports agentic-risk evaluations, but nothing in the model or template enforces an approval (8 of 20). The chat template neutralises special-token sequences and media placeholders in caller text and wraps tool results in `<tool_output>`, and the card advises protecting internal assistant messages and layering defences against adversarial input. No indirect prompt-injection results are published (10 of 15). No call log of its own. Visibility depends on the serving stack (5 of 15). Weights are safetensors that load through Transformers' own classes from 5.19, and Meta published cyber, chemical, biological and misalignment evaluations. No security.txt on dev.meta.ai or research.meta.ai, and no security route on the card or in the docs, only museoss@meta.com and the Hugging Face discussions (9 of 20). 44.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-security):

- 0 to 30, the credential model. 30 for OAuth 2.1 with scopes, or scoped and revocable keys with rotation. 20 for plain revocable API keys. 10 for one all-powerful key. 10 off when a secret can travel in a URL query string as a documented option.
- 0 to 20, read-only or least-privilege modes, and confirmation or approval for destructive actions.
- 0 to 15, prompt-injection posture where the tool returns untrusted content (documented mitigations or guidance). A tool that returns no untrusted content gets 10.
- 0 to 15, audit logs or per-call visibility for the operator.
- 0 to 20, a security programme. security.txt or a disclosure policy, a bug bounty, SOC 2 or ISO 27001, advisories handled in public.

Models are read for retention, whether API data trains models (and whether that's off by default), zero-retention options and certifications. Frameworks for telemetry defaults, approval hooks, guardrails and sandboxing.

## 3. Schema & documentation, 61 out of 100, up to 6.3 more on the total

Why it scored 61: Read for a model the owner serves. No specification of its own. The contract is the Jinja chat template, which renders OpenAI-format tools with `additionalProperties: false` added and rejects bad values, plus the OpenAI-compatible servers in vLLM and SGLang with the `muse` parsers (10 of 25). dev.meta.ai/llms.txt lists eight Muse Spark Open pages, each with a Markdown copy (10). The card lists seven intended uses and six limitations, the docs say when to pick BF16 or NVFP4 and which reasoning strength suits which task, and the Usage Policy states what is out of scope (17 of 20). `reasoning_effort` is an enum of four values and anything else returns an error, and tools take JSON Schema. Structured output is documented only for TensorRT-LLM's `response_format`, not in the prompting guide or for vLLM and SGLang (11 of 15). curl and Python examples for every runtime and hosted provider, the rendered prompt for each case, and the template's own error messages, but no error reference for a served model (11 of 15). One release, with no changelog or version scheme beyond the licence's version line, Muse Spark Open (2 of 15). 61.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-schema):

APIs and MCP servers.

- 25, a machine-readable contract (a public OpenAPI file or similar; for MCP, typed JSON Schema inputs on every tool).
- 10, llms.txt or Markdown docs served for agents.
- 0 to 20, descriptions that say what a tool is for, when to use it and when not to, read from the tool definitions in the source or the API reference.
- 0 to 15, typed inputs with enums, constraints and required fields, and no free-form JSON blobs.
- 0 to 15, examples and documented error responses.
- 15, versioning and a public changelog.

Models are read from the API reference, the OpenAPI file, llms.txt, the structured-output and tool-use docs and the model cards. Frameworks from docs a model can follow, typed interfaces, examples and the API reference.

## 4. Payments & pricing, 50 out of 100, up to 6.3 more on the total

Why it scored 50: Self-hosted rule. Nothing to buy from Meta, and the licence grants use free of charge. The Model as a Service clause asks for a separate agreement only from licensee groups above $20M gross revenue in 12 months (20). Free to run with no card (20). Meta's own weights need a manually reviewed form with legal name and date of birth, so an agent can't get them alone. The ungated GGUF quantisations from ggml-org and Unsloth that Meta's docs point to download without approval, so part of the line (10 of 20). No payment protocol (0). Hosted copies at $1.25 input and $4.25 output per million tokens need each provider's own account and are not graded. 50.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-payments):

The published rubric, also on the [x402 page](https://www.anchorterminal.com/x402/).

- 40, a machine payment protocol (x402, MPP or L402) on the tool's own endpoints. 10 to 30 when it covers only some endpoints or only goes through a third party, and the note says which.
- 20, per-call or per-unit pricing published without a login. 10 for public plan-only pricing, 0 for "contact sales" or prices behind a login.
- 20, a free tier or trial that doesn't need a card.
- 20, autonomous onboarding, meaning an agent can get access without a person signing up in a browser (keyless use, x402, a programmatic key API).

Payment platforms and agent wallets rarely charge for their own API over a machine protocol, so the first line has steps for them, and the highest one that applies counts. 40 when x402, MPP or L402 runs on all their own endpoints, 30 when it runs on part of their own API, 25 when their merchants can accept one, 20 for running a facilitator, 15 for paying as a buyer, and 0 when the only protocol is their own. Merchant acceptance sits above a facilitator because the platform's own customers can charge agents through it, while a facilitator settles for sellers who wire up the protocol themselves. The counter-argument (a facilitator does more for the protocol as a whole) has a point. Each note says which step applied.

Open-source software you run yourself is scored on its hosted or paid option if it has one. A free, self-hosted package with nothing to buy gets 20, 20 and 20 for the last three lines, and 0 to 40 for the first only if it ships a payment protocol.

## 5. Agent ergonomics, 69 out of 100, up to 5 more on the total

Why it scored 69: Model reading of the checklist (tool use, structured output, caching, context, throughput, defaults and SDKs, errors). OpenAI-format tools and parallel calls come back as `tool_calls`, and the template supports namespaced tools and deferred loading through a `tool_search` tool. A forced `tool_choice` is not documented (17 of 20). JSON mode and JSON Schema come from each runtime's guided decoding, which Meta's docs show for TensorRT-LLM only (8 of 15). Prefix caching is the runtime's. Meta's SGLang command turns on hierarchical caching, and Dynamo routes each request to the worker holding its cached context (8 of 10). 1,048,576-token context, with sliding-window attention on 46 of 62 layers holding KV memory down (15 of 15). Draft models for speculative decoding in SGLang and vLLM, and no batch mode of its own (7 of 10). Any OpenAI SDK works against a local server, but reasoning can't be turned off, defaults to the most expensive level, and needs streaming, a `max_tokens` budget for reasoning and the right stop tokens (6 of 15). The template raises named errors for an unsupported reasoning effort, tool arguments passed as a string and a tool message with no name. Server errors are the runtime's (8 of 15). 69.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-ergonomics):

- 0 to 25, context cost. For MCP, the number and size of the tool definitions (25 for ten or fewer compact tools, 15 for 11 to 30, 5 for more than 30, plus up to 10 back for toolsets, dynamic loading or read-only subsets). For APIs, whether responses can be sized (field selection, limits, summaries).
- 20, pagination, filtering and output-size controls.
- 20, actionable, documented error responses, codes and messages an agent can recover from.
- 20, idempotency or safe retries, and for MCP the `readOnlyHint` and `destructiveHint` annotations.
- 15, sensible defaults, few required parameters, and official SDKs in at least two languages.

Models are read for tool use, structured output, prompt caching, context length, batch and SDKs. Frameworks for how much code and how many defaults a tool-calling agent with MCP needs.

## 6. Transparency & trust, 53 out of 100, up to 4.1 more on the total

Made of editorial 50, provenance 55.

Why it scored 53: Weights are downloadable under the Muse Model Licence, which is clear and broad in its grant but not an OSI licence. It adds a display requirement, a separate agreement before Model as a Service by a licensee group above $20M gross revenue, and a Usage Policy with field-of-use limits (military, nuclear, espionage, critical infrastructure, heavy machinery) and an age limit of 18. It has no regional exclusion and no governing-law clause. Training code and data aren't released (18 of 30). Inputs stay on the owner's hardware, and Meta's docs say self-hosting keeps 'every prompt and response' inside the owner's infrastructure. The access form's details go to Meta under its general privacy policy, which we didn't read, and the training data is described in outline only (20 of 30). No deprecation policy, dated notices or versioning found (0 of 20). The Get the model page says everything runs offline once the weights land, and safetensors weights carry no code. No other telemetry statement, and Meta's download form records name, date of birth, country and IP location (12 of 20). 50.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-transparency):

- 0 to 30, source availability and licence clarity. 30 for open source under an OSI licence, 15 for closed with clear terms, 0 for unclear terms.
- 0 to 30, data handling and retention statements that agree with each other (privacy policy, DPA, retention periods, subprocessors).
- 0 to 20, a deprecation policy or notices with dates.
- 0 to 20, telemetry disclosed with an opt-out (local software), or subprocessors and data locations disclosed (hosted).

The other half of Transparency and trust is the provenance score, computed from checked facts (below). The category score is the mean of the two.

Provenance checks not met in full (half of this category, computed from checked facts):

- Domain age: meta.ai, registered 2017-12-16 (8 years) (11 of 15)
- Status page: not found (0 of 10)
- Changelog: not found (0 of 10)
- security.txt: not found (0 of 10)

## 7. Maintenance & community, 60 out of 100, up to 3.5 more on the total

Why it scored 60: Read for an open-weight model released today. Released on 15 October 2026 (30). One release, BF16 and NVFP4 on the same day, with no history of revisions, successors or notice periods, and Meta Model API has no deprecation policy either (6 of 20). Three Hugging Face discussions opened on release day had no reply from Meta when read. Feedback goes to the discussions or museoss@meta.com, and Meta's docs and developer blog carry dated posts (8 of 25). Day-one support in vLLM, SGLang, TensorRT-LLM, Transformers, llama.cpp, MLX and Unsloth with Meta's own recipes, though SGLang's is a dev image and TensorRT-LLM's a feature branch (12 of 15). Current dependencies (Transformers 5.19), but no CI visible to us (4 of 10). 60.

The checklist (https://www.anchorterminal.com/benchmark/#checklist-maintenance):

- 0 to 30, time since the last release, or the last published model or API change for a closed service. 30 within 30 days, 20 within 90, 10 within 180, 0 older.
- 20, at least three releases or dated changelog entries in the last 90 days.
- 0 to 25, responsiveness. Issues and pull requests answered on GitHub (the open issues and how recent the replies are). For closed services, a public changelog and a support or community channel that answers, 0 to 15.
- 15, presence in the official MCP registry under a verified namespace (MCP servers), or current official SDKs (APIs and models).
- 10, package health, current dependencies and CI.

Models are read for deprecation notice periods and model churn rather than release counts.

## What we couldn't check

What we couldn't read counted as absent. Publishing it on a page a plain HTTP fetch can read (not only in a browser) lets the next check count it.

- unchecked: config.json, generation_config.json, USE_POLICY.md and the commit history of both Meta repositories, which answered 401 without gated access. The Usage Policy was read as Exhibit A of LICENSE.md, which is public
- unchecked: github.com/meta-models, which answered 403 to our fetcher, so any CI, issues or reference code there were not read
- unchecked: the Meta Privacy Policy that governs the access form, because www.facebook.com's robots.txt asks readers like ours not to fetch it
- WebFetch refused the Hugging Face model card URL as outside its allowed set. The card, licence, repository and collection metadata, discussions, evaluation PDF, dev.meta.ai docs and the providers' pages were read with curl. Only the research post came through WebFetch
- The full evaluation report is described as 'available soon' in the launch post and was not published on 15 October 2026. Benchmarks here are Meta's own, from the card, the scorecard image and the methodology PDF
- How long manual approval takes, and whether it is refused by country or for trade-compliance reasons. The card mentions trade compliance laws, and the licence has no regional exclusion
- Which hosted Muse Spark version these weights match. Meta's docs call it the open-weight release of Muse Spark without naming a version, and OpenRouter prices it the same as Muse Spark 1.3 on Meta Model API
- Hardware figures differ between pages. The card says a GPU with at least 700GB of VRAM for NVFP4 and 2TB for BF16, the docs say 4 × B300 and 8 × B300, the developer post says four Blackwell Ultra GPUs or eight H200s for NVFP4, and the smallest Unsloth GGUF is given as about 300 GB in one place and 256 GB in another
- 'The most capable US open-weight model to date' and 'the first model in its class to run on your desktop' are Meta's claims. The card's own table puts GLM 5.3, Kimi K3 and Qwen3.8-2.4T-A95B above it on the AA Intelligence Index and doesn't say which compared models are American. No deduction taken
- Hosted providers were read for availability, context and price only, not graded. Groq is a limited preview needing organisation approval, with a 131,072-token context
- Muse Glimmer, Meta's 30B open-weight model on Hugging Face since 9 August 2026, has no listing and is a candidate for one

## Weaknesses

- The NVFP4 checkpoint (about 537 GB) serves on four B300 GPUs and the BF16 one (about 1.65 TB) on eight
- Meta's Hugging Face repositories need manual approval with legal name and date of birth, which the Get the model page doesn't mention
- Products must display "Built with Muse Spark Open", and a company group with over $20M revenue needs Meta's agreement to offer it as Model as a Service
- The Usage Policy rules out military, nuclear, espionage, critical infrastructure and heavy machinery uses, and users under 18
- Released on 15 October 2026, with no changelog, deprecation policy or CI found, and the full evaluation report not yet published

## What costs an agent a turn today

The notes we give agents before they call it. Each one is a workaround an agent shouldn't need.

- Start vLLM or SGLang with `--reasoning-parser muse --tool-call-parser muse` (vLLM also `--enable-auto-tool-choice`), or tool calls come back as raw text
- Set `reasoning_effort` to `low` or `medium` for short tasks. The default is `xhigh`, reasoning can't be turned off, and other values return an error
- Stop only on `<|eot|>` and end of text. Stopping on `<|eom|>` cuts off tool calls and final answers
- Stream long generations and leave `max_tokens` room for reasoning, or `content` can come back empty with `finish_reason` `length`
- Don't call Meta Model API for this model. Use a self-hosted server or a hosted provider's model string, such as `meta-models/muse-spark-open`

## When it's done

Send what changed and where it's published as a dispute (https://www.anchorterminal.com/builders/#disputes, or `POST https://www.anchorterminal.com/api/v1/contact` with `"kind": "dispute"`). Disputes are answered in public, and the listing is checked again by the same checklist. Paying for an audit or a listing claim changes nothing here.