Muse Spark Open: run Meta’s open-weights model with Hermes Agent

Published October 15, 2026 · Weights, prices and license checked on the publication date · 中文版

What Meta released

On October 15 Meta published Muse Spark Open, the first Muse Spark model whose weights you can download. Earlier Muse Spark releases (1.1 to 1.3) are API-only. The release post aims it at long-running agent work, large software tasks and multi-step reasoning, which is the kind of work a Hermes Agent does.

Only 42B parameters are active on each token, so generation runs much faster than the 816B total suggests. That is what makes Apple-silicon machines a realistic option. You still have to hold the full model in memory.

Muse Spark Open is not on Meta’s own Model API. Meta’s model docs say to download the weights or use a hosted inference provider. That means the meta-ai provider in Hermes, which talks to api.meta.ai, cannot reach this model. Use OpenRouter or your own server instead.

Pick a route

For most Hermes users, OpenRouter is the sensible place to start. Self-hosting pays off when the data must not leave your network, or when you already own the hardware and run the bot all day.

Use it on a managed Hermes instance

1. Open the dashboard and find your Hermes instance.
2. Click its current text model, choose the OpenRouter provider and search for Muse Spark Open.
3. If it is not in the shortlist, search the full OpenRouter catalogue for meta/muse-spark-open. To type it in by hand, choose Custom platform… and enter platform meta and model ID muse-spark-open. This routes meta/muse-spark-open through OpenRouter. meta is only the catalogue namespace here, not a direct connection to Meta.
4. Confirm the switch, wait for the instance to come back, and start a new conversation. Check that the instance card shows the new model.

Muse Spark Open is a paid model. It is not part of the free trial. OpenRouter requests use your own OpenRouter key if you have saved and enabled one, and your eligible managed credits otherwise. To pay OpenRouter directly, add a key on the API keys page. Managed instances cannot load the self-hosted weights; that route is for Hermes running on your own machine.

Self-hosted Hermes through OpenRouter

Run hermes model, choose OpenRouter, enter your OpenRouter key when asked, and pick or type meta/muse-spark-open. The matching block in ~/.hermes/config.yaml is:

model:
  provider: openrouter
  default: meta/muse-spark-open

Keep OPENROUTER_API_KEY in Hermes’s credential setup or ~/.hermes/.env, not in this block. Add these fields to your existing configuration rather than replacing the file. Reference: the Hermes configuration docs.

Start a new CLI session after the change. If a messaging gateway is already running, run hermes gateway restart and begin a new conversation.

Self-hosted Hermes against your own server

vLLM, SGLang, llama.cpp and MLX all expose an OpenAI-compatible /v1 endpoint. Meta’s developer post gives this vLLM launch for the NVFP4 build on four GPUs. The muse parsers matter most for Hermes; --served-model-name is our addition:

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 \
  --served-model-name muse-spark-open

At launch the PyPI release of vLLM cannot serve this model yet. On NVIDIA, run the command inside the vllm/vllm-openai:muse-spark-open Docker image. AMD GPUs use the vllm/vllm-openai-rocm:muse-spark-open image with a different checkpoint (AMD’s Quark MXFP4 or BF16) and their own settings, listed in vLLM’s recipe. The vLLM team says nightly wheels follow once its support PR is merged. Meta’s version of the command also turns on speculative decoding with its DSpark draft model. SGLang takes the same --reasoning-parser muse --tool-call-parser muse pair. Hermes then connects as a custom provider (vllm and llamacpp are aliases for custom):

model:
  provider: custom
  base_url: http://192.168.1.50:8000/v1
  default: muse-spark-open
  context_length: 262144

• default must match the name your server reports at /v1/models. The --served-model-name flag above sets it; without it, the name is the full Hugging Face repo ID.
• Set context_length to what your server really allows. Hermes normally detects it. Self-hosted setups often cap the window below 1M to save memory, and Hermes needs the real limit to know when to compress history. The 262,144 above is an example, not a recommendation.
• Keep the muse tool-call parser. A Hermes agent works by calling tools, and Muse Spark Open writes tool calls in its own format. Without --tool-call-parser muse and --enable-auto-tool-choice, calls come back as plain text and the agent stalls. Our vLLM guide covers the server side in more detail.
• Reasoning effort accepts only low, medium, high or xhigh. The model card says the default is xhigh, recommends high or xhigh for agent work, and says any other value returns an error. Hermes also has levels such as none, minimal and max, so if you set an effort for this model, use one of the four. The card recommends temperature 1.0 and top_p 1.0.
• Getting the weights. Anyone can read the model card. To download the official weights you must give your full legal name, date of birth and organization, and Meta reviews each request by hand. The community MLX and GGUF conversions are not gated.

What hardware it needs

These are file sizes, not total memory use. The KV cache for long conversations comes on top, and it grows with the context you allow.

Meta lists these minimums:

On a Mac, the machines are linked directly over Thunderbolt 5 with RDMA, and every layer is split across them. Each M5 Ultra has up to 512 GB of unified memory with 1.2 TB/s of bandwidth. Meta’s Mac cookbook loads the official 533 GB NVFP4 checkpoint and serves an OpenAI-compatible API on port 8080:

mlx.launch --hostfile hostfile.json --backend jaccl --env MLX_METAL_FAST_SYNCH=1 -- \
  mlx_lm.server --model meta-models/Muse-Spark-Open-816B-A42B-NVFP4 --port 8080

mlx_lm.server listens only on 127.0.0.1 by default. If Hermes runs on the same Mac, set base_url to http://127.0.0.1:8080/v1. To reach it from another machine, add --host with the Mac’s LAN address, and keep that port off the public internet, since the server has no login. Set default to the name that /v1/models returns. Meta’s cookbook does not say whether tool calling works on this path, so test a tool call before you rely on it. The ungated community MLX builds are smaller: 462 GB at 4-bit and 359 GB at 3-bit. The 3-bit build fits in a single 512 GB machine, but that is below Meta’s stated minimum and loses more quality.

On NVIDIA and AMD, Meta reports kernel-level speedups of 1.03× to 3.95× in SGLang. AMD publishes MXFP4 and FP8 builds, and Intel publishes an INT4 build. A single 24 GB or 48 GB GPU cannot run this model. For a model that fits one card, see Hermes with llama.cpp or Hermes on a Mac.

Pricing and a worked budget

Every provider OpenRouter routes to charges the same input and output rate at launch. Cache-read pricing is listed for three of them and not for Nebius. Check the OpenRouter model page before a large run.

For 100,000 uncached input tokens + 10,000 output tokens, the token cost is about $0.17: $0.125 in plus $0.0425 out. If the whole input is a cache hit, it drops to about $0.06. A cache hit is never guaranteed. Long tool loops resend history every turn, so a long Hermes task can still cost far more than one request.

This is the same list price as the closed Muse Spark 1.3 on OpenRouter. The reasons to choose Spark Open are that you can also run it yourself, and a copy of the weights you download stays exactly as it is. For the API-only model, see Hermes + Muse Spark.

What the license allows

The Muse Model License lets you run, modify, fine-tune, distill and redistribute the model, subject to Meta’s usage policy. Three conditions matter for anyone building on it:

• Attribution. A product or service that uses the model, or a model derived from it, must show “Built with Muse Spark Open” in its interface, documentation, website or settings menu.
• Model-as-a-Service above $20M. If you give third parties API access to AI models (free or paid) and your group’s gross revenue exceeds US$20 million in any 12-month period, you must notify Meta and sign a separate agreement before offering this model. This does not apply to internal use, research, or using it to power features of your own product.
• Usage policy. The policy lists prohibited uses and says the model is not intended for people under 18.

A personal or company Hermes Agent that runs the model internally is outside the revenue clause. You still need to follow the attribution condition and the usage policy.

Frequently asked questions

Related guides

• Hermes + Muse Spark (API model)
• Self-host models for Hermes with vLLM
• Run Hermes on a Mac