SSyncDev

Muse Spark Open

SyncDev model directory · Profile updated Nov 17, 2026

Compare with another model

Muse Spark Open is Meta's open-weight 816B mixture-of-experts model (42B active) with 1M context, built for coding and agents you can self-host.

Key specifications

Specification score
46.7
Context window
1.05M
License
Muse Model License

The specification score summarizes recorded features and limits. It is not a benchmark result.

Pricing conditions: Open weights, no first-party API price. Hosted gateways indicatively around $1.25 in / $4.25 out per 1M tokens; varies by provider. Confirm with your host.

Inputs and outputs

Recorded modality support shows what this model can accept and produce. An undocumented modality is shown as unknown rather than assumed to be unsupported.

Accepts

  • Text input
  • Image input
  • Video input
  • Audio input

Produces

  • Text output

Capabilities

VisionAudioOpen weights

What is Muse Spark Open?

Can a frontier-class model run under your desk? Meta's answer, for this release, is yes, if your desk holds two Apple M5 Ultra Mac Studios with 512 GB of unified memory each. That claim sets the tone: this is the open-weight edition of Muse Spark, meant to be downloaded, fine-tuned and served by anyone who accepts the licence.

Architecture in brief

  • 816B total parameters, 42B active per token (mixture of experts: 256 experts, 8 routed per token, 62 layers).
  • A 1.8B-parameter perception encoder; text, image, video and audio in, text out.
  • 1M-token context.
  • Weights in BF16 and NVFP4. Local serving needs roughly 700 GB of GPU memory for NVFP4 or about 2 TB for BF16.

Released on 15 October 2026, it is aimed at coding and agentic work: long-horizon tasks, complex software engineering, multi-step reasoning. Meta also says it received extra safety post-training because open weights ship without system-level guardrails, and that it met the release thresholds of Meta's Superintelligence Scaling Framework.

Where to get it

Downloads are offered through Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Hosted endpoints come from Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI. Supported runtimes include SGLang, vLLM, TensorRT-LLM, Transformers, MLX, llama.cpp and Unsloth.

Meta publishes no first-party API price. Gateway listings cluster around $1.25 per million input tokens and $4.25 per million output, but that is a market rate that varies by host; treat it as indicative and check the provider you actually use.

The licence is the real decision

The Muse Model License allows use, modification, fine-tuning, distillation and redistribution. How restrictive it is depends on which summary you read. Meta's public description stops at "limited commercial terms" that let many developers use it free. The licence text itself, as summarised at release, adds two concrete conditions: products must display "Built with Muse Spark Open", and any business offering AI models as a service with more than $20 million in gross revenue over twelve months must notify Meta and sign a separate agreement. Read the licence file shipped with the weights before you build on it; that document, not a summary, binds you.

Version lineage is unclear

Meta calls it "the open-weight version of Muse Spark" without saying which one. It has been presented both as Muse Spark 1.2 going open and as sharing a base with Muse Spark 1.3. If you are comparing against the hosted Muse Spark 1.2, benchmark both on your own prompts rather than assuming they are identical. Maximum output is similarly unconfirmed: a figure of 131,072 tokens circulates on some listings, but Meta does not state one.

Publisher-reported numbers

Meta reports 39 on the Artificial Analysis Intelligence Index, 89.2 on MCP-Atlas, 61.8 on JobBench, 89.0 on GPQA Diamond and 76.4 on Terminal-Bench 2.1. On that last test, open rivals are listed higher, so this is not a clean sweep.

Sizing a pilot

Start on a hosted endpoint, not your own cluster. Run your agent's real tasks for a few hundred episodes and log tokens per episode; with 42B active parameters, throughput depends heavily on the host. Only once quality holds should you price self-hosting against gateway rates. For a smaller open Meta model, look at Muse Glimmer 30B, and see the full family on the Meta provider page.

How to evaluate Muse Spark Open

Compare the model on the work you need to complete.

Tasks and inputs

Muse Spark Open is listed for Multimodal, Text → Text, Coding & Code Generation. It accepts text, image, video, audio and produces text. Use the publisher’s input format and preprocessing for your chosen deployment.

Capacity and cost

Documented limits: 1.05M context. Measure a completed task using the prompts, documents and tools your application will use. Include reasoning and retries in the result.

Choosing a deployment

The weights are downloadable under Muse Model License. Check runtime support and size the complete model and working context for your expected concurrency.

Frequently asked questions

816B total parameters in a mixture-of-experts design, with 42B active per token and a 1M-token context window.

These records share a directory category. Check the source card, output type and deployment requirements before treating a model as a replacement.

baidu

ERNIE 5.5 Flash Preview

ERNIE 5.5 Flash Preview is Baidu's first post-trained 280B-A14B text model, with a 1M window, aimed at game dev, 3D scripting and knowledge work.

1M context

Meta

Muse Ember

Muse Ember is Meta's smallest, cheapest Muse model: a 1M-context multimodal worker for tool calls, coding subtasks, routing and extraction at volume.

1.05M context

iflytek

Spark X2.6

Spark X2.6 is iFLYTEK's 293B mixture-of-experts language model family with integrated device and cloud variants, unveiled in October 2026.

— context

ByteDance Seed

Seed 2.1 Lite

Seed 2.1 Lite is a lightweight Doubao Seed 2.1 model listed on Volcengine Ark with a 1M context, 256K output, audio and image input and tool calls.

1.02M context