# Muse Spark Open (slim)

> Muse Spark Open is Meta's open-weight multimodal model for coding and agentic work, with 816B total and 42B active parameters and a 1M-token context, released on 15 October 2026 under the Muse Model Licence for self-hosting.

- Full: https://www.anchorterminal.com/tools/muse-spark-open.md (~9,000 tokens) · this version ~2,480 tokens · JSON https://www.anchorterminal.com/tools/muse-spark-open.json · canonical https://www.anchorterminal.com/tools/muse-spark-open
- Index: https://www.anchorterminal.com/llms.txt · API: https://www.anchorterminal.com/api/v1/index.json · Updated: 2026-10-15
- Licence: CC BY 4.0, credit Anchor Terminal (anchorterminal.com) and link this page when you pass anything on

**D · 53.4/100 · rank #760 of 1095 · #16 in Model APIs & inference · not agent-ready · confidence medium**


Assessment: Muse Spark Open runs on the owner's hardware in seven runtimes, with OpenAI-compatible servers in vLLM and SGLang, standard tool calls, a 1M-token context and Meta's own serving docs. Serving the 537 GB NVFP4 checkpoint takes four B300 GPUs, Meta's download is gated by manual review, and the licence adds attribution and a revenue threshold for model resellers.

## Facts

- Kind: Model API · vendor: Meta · category: Model APIs & inference · legal entity: Meta Platforms, Inc. · provenance 55/100
- Local only (HTTP)
- Auth: None · pricing: Free · x402: no · licence: Muse Model Licence (source-available weights, not an OSI licence), free of charge with an attribution notice, a separate agreement with Meta before Model as a Service by a licensee group above $20M gross revenue, and the Muse Spark Open Usage Policy
- Probe metrics: not measured yet (probes haven't run)
- Surface graded: The open weights as served by the owner, through the OpenAI-compatible servers in vLLM and SGLang that Meta documents, read from the model card, the licence, the Hugging Face metadata and Meta's docs on 15 October 2026. Nothing was downloaded or run. Hosted providers are listed below and not graded
- Model: 816B total and 42B active parameters (822,036,529,664 in the BF16 safetensors), mixture of experts with 256 experts and 8 per token, 62 layers, hidden size 8,192, vocabulary 202,048, plus a 1.8B-parameter perception encoder. Text, image, video and audio in, text out. Knowledge cutoff 4 January 2026 in the chat template
- Checkpoints: BF16 (about 1.65 TB, served at TP8 and EP8 on 8 × B300) and NVFP4 (about 537 GB, TP4 and EP4 on 4 × B300) from Meta. FP8 and MXFP4 from AMD, MXFP4 and INT4 from Intel, GGUF from ggml-org and Unsloth, per Meta's developer post
- Hardware: Meta's docs give 8 × B300 for BF16 and 4 × B300 for NVFP4, with about 2 TB of VRAM budgeted for BF16. The card says 700 GB for NVFP4. MLX serves NVFP4 on two 512 GB or four 256 GB Mac Studio M5 Ultra machines linked by Thunderbolt 5
- Context and reasoning: 1,048,576 tokens. Reasoning is always on, with `reasoning_effort` low, medium, high or xhigh (strengths 8, 32, 256, 512), default xhigh. Recommended sampling temperature 1.0 and top_p 1.0
- Licence: Muse Model Licence, version Muse Spark Open, 15 October 2026. Free of charge to use, copy, modify, fine-tune, distil, merge, distribute, sublicense and sell. Conditions are the notice and Usage Policy in every copy, "Built with Muse Spark Open" shown in the product or its documentation, and a separate written agreement before Model as a Service when the licensee and affiliates exceed $20M gross revenue in 12 months. Internal use, research and products that use the model as a component are exempt from that clause. Meta claims no ownership of outputs. No governing-law or termination clause
- Usage Policy: Exhibit A of the licence, dated 15 October 2026. Not intended for anyone under 18. Rules out, among others, military, warfare, nuclear, espionage and ITAR-controlled uses, operation of critical infrastructure, transportation technologies or heavy machinery, malicious code, unlicensed professional practice, inferring sensitive personal data without the right to, and presenting outputs as human-generated
- Access: Both Meta repositories on Hugging Face are gated with manual review. The form asks for full legal name, date of birth, country, affiliation and job title, records IP location, and shares the details under the Meta Privacy Policy
- Runtimes: SGLang and vLLM (OpenAI-compatible, `muse` reasoning and tool-call parsers, Docker images `lmsysorg/sglang:dev-muse-spark-open` and `vllm/vllm-openai:muse-spark-open`), TensorRT-LLM from the `feat/musespark-open` branch, NVIDIA Dynamo, Transformers 5.19 or later, llama.cpp, Unsloth (OpenAI- and Anthropic-compatible) and MLX. Draft models for speculative decoding from RadixArk and Inferact
- Hosted by others (not graded): Baseten, Fireworks AI, Modal, Nebius Token Factory and Together AI, each with a context of about 1M tokens. Groq in limited preview with organisation approval and a 131,072-token context. OpenRouter routes to Together, Nebius, Fireworks and Baseten at $1.25 input and $4.25 output per million tokens (checked 2026-10-15). Also in Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models catalogues
- Meta Model API: Does not serve Muse Spark Open. Meta's models page says 'You don't call it over Model API' and lists its API endpoints as none. See the meta-model-api listing for the hosted Muse models and muse-code for Meta's coding agent, which the launch post says Meta tested with this model on two Mac Studios
- Meta's figures: Vendor-reported at xhigh effort. AA Intelligence Index 39, MCP-Atlas 89.2, JobBench 61.8, Toolathlon Verified 72.8, OSWorld Verified 80.4, Terminal-Bench 2.1 76.4, DeepSWE 1.1 60, SWE-Atlas-QnA 58, SciCode 58.5, GPQA Diamond 89.0, HLE with tools 61.0. Sources mix Artificial Analysis, Scale AI and Meta's own runs
- Meta's safety figures: Vendor-reported. StrongREJECT v2 (an internal benchmark) attack success 0.5. FORTRESS average risk score 5.8 with an over-refusal score of 17.4, the highest over-refusal in the table. 99.2 per cent of cyber jailbreaks refused on an internal benchmark
- Adoption: On 15 October 2026 the BF16 repository showed 165 likes and 402 downloads and the NVFP4 repository 68 likes and 331 downloads
- Age: Released on 15 October 2026, the day of this check. Both repositories carry a creation date of 15 September 2026
- Support: Discussions on the Hugging Face page and museoss@meta.com, per the model card. No security reporting route named on the card or in the docs

- Scores: Reliability 42, Performance pending, Schema & documentation 61, Agent ergonomics 69, Security & auth 44, Payments & pricing 50, Task success pending, Maintenance & community 60, Transparency & trust 53 · total over the 7 assessed categories
- Why: Reliability, Read with the local-software lines, since the owner serves the weights. · Schema & documentation, Read for a model the owner serves. · Agent ergonomics, Model reading of the checklist (tool use, structured output, caching, context, throughput, defaults and SDKs, errors). · Security & auth, Read as software the owner runs. · Payments & pricing, Self-hosted rule. · Maintenance & community, Read for an open-weight model released today. · Transparency & trust, Weights are downloadable under the Muse Model Licence, which is clear and broad in its grant but not an OSI licence.
- Sources: 24, open questions: 11, both in the full twin
- Capabilities: inference.llm, inference.open-weights, inference.local
- JSON: https://www.anchorterminal.com/api/v1/tools/muse-spark-open.json
- Verify (for the vendor): the badge `https://www.anchorterminal.com/badges/muse-spark-open.svg` or a link to https://www.anchorterminal.com/tools/muse-spark-open from a page on meta.ai or one of its subdomains, then `POST https://www.anchorterminal.com/api/v1/verify` `{"slug", "url"}` or `verify_listing` at /mcp; re-checked weekly, no effect on the grade. Snippets in the full twin.

## Before you call it

1. 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
2. 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
3. Stop only on `<|eot|>` and end of text. Stopping on `<|eom|>` cuts off tool calls and final answers
4. Stream long generations and leave `max_tokens` room for reasoning, or `content` can come back empty with `finish_reason` `length`
5. 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`

## Connect

```bash
hf download meta-models/Muse-Spark-Open-816B-A42B-NVFP4 --local-dir ./muse-spark-open-nvfp4

export VLLM_USE_RUST_FRONTEND=1
vllm serve meta-models/Muse-Spark-Open-816B-A42B-NVFP4 \
    --tensor-parallel-size 4 \
    --enable-expert-parallel \
    --max-model-len 1048576 \
    --reasoning-parser muse \
    --tool-call-parser muse \
    --enable-auto-tool-choice
```

```bash
curl http://localhost:8000/v1/chat/completions \
  -H 'Content-Type: application/json' \
  -d '{
    "model": "meta-models/Muse-Spark-Open-816B-A42B-NVFP4",
    "messages": [{"role": "user", "content": "Refactor this module."}],
    "reasoning_effort": "xhigh"
  }'
```

## Similar tools

| Tool | Grade | Score | Shared capabilities | Slim |
| --- | --- | --- | --- | --- |
| Docker Model Runner | C | 57.1 | inference.local, inference.open-weights, inference.llm | https://www.anchorterminal.com/tools/docker-model-runner.min.md |
| Ollama | C | 56.3 | inference.local, inference.open-weights, inference.llm | https://www.anchorterminal.com/tools/ollama.min.md |
| GroqCloud | BB | 75.6 | inference.llm, inference.open-weights | https://www.anchorterminal.com/tools/groq.min.md |
| Mistral AI API | BB | 71.2 | inference.llm, inference.open-weights | https://www.anchorterminal.com/tools/mistral-api.min.md |
| Cloudflare Workers AI | B | 68.3 | inference.llm, inference.open-weights | https://www.anchorterminal.com/tools/cloudflare-workers-ai.min.md |


## Panel reviews (0, desk reviews from public material, no calls made)