Meta's open-weight frontier MoE — 816B parameters (42B active), 1M-token context, and agentic coding as the headline skill.
curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "muse-spark-open",
"messages": [{ "role": "user", "content": "Explain quantum tunneling simply." }]
}'Overview
Muse Spark Open is Meta's open-weight frontier model: an 816B-parameter sparse mixture-of-experts model with roughly 42 billion parameters active per token and a 1M-token context window, released October 15, 2026. It is trained for agentic coding and long-horizon tasks, accepts text and images, and can call tools, search the web, and reason across multi-step work.
On Venice, Muse Spark Open runs under the private tier with zero retention — your prompts and code are not stored, profiled, or used for training. Note that this model is not end-to-end encrypted or TEE-isolated on Venice, so it is private by non-retention rather than by hardware attestation. It is also not one of Venice's uncensored models: it ships with Meta's standard safety alignment, so if unfiltered output is the priority, Venice Uncensored 1.2 remains the option for that.
Agent quickstart
The API is OpenAI-compatible: change the base URL and the model id and existing client code works unchanged.
curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "muse-spark-open",
"stream": true,
"messages": [{ "role": "user", "content": "Draft the release note." }]
}'curl https://api.venice.ai/api/v1/chat/completions \
-H "Authorization: Bearer $VENICE_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "muse-spark-open",
"messages": [{ "role": "user", "content": "Find the rate limits." }],
"tools": [{
"type": "function",
"function": {
"name": "search_docs",
"description": "Search the API documentation.",
"parameters": {
"type": "object",
"properties": { "query": { "type": "string" } },
"required": ["query"]
}
}
}],
"tool_choice": "auto"
}'import os
from openai import OpenAI
client = OpenAI(
api_key=os.environ["VENICE_API_KEY"],
base_url="https://api.venice.ai/api/v1",
)
resp = client.chat.completions.create(
model="muse-spark-open",
messages=[{"role": "user", "content": "Build without permission."}],
)
print(resp.choices[0].message.content)Specifications
Assessment
Use cases
Prompting
Pin your coding standards and project context as a cached system prompt — cached input bills at $0.19 per 1M tokens instead of $1.56.
Ask it to plan before acting ("list the files you'll touch and why, then wait for confirmation") — it is trained to plan, act, observe, and self-correct.
Attach screenshots of the bug or the rendered page; it reasons over images, not just code text.
Request JSON schema output explicitly when you need machine-readable results instead of prose.
Let it use web search for current library docs rather than relying on training-data versions.
Paste whole files or long logs rather than summarizing — the 1M-token window is there to be used, and truncation costs more accuracy than tokens cost.
Alternatives
| Model | Best for | Context | Open weights | Price (Venice) |
|---|---|---|---|---|
| Muse Spark Open | Agentic coding & long-horizon tasks | 1M tokens | Yes | $1.56 in · $5.31 out / 1M |
| Kimi K3 | Open-weight long-context reasoning | 1M tokens | Yes | $3.75 in · $18.75 out / 1M |
| Gemini 3.8 Flash | Fast everyday tasks | 1M tokens | No | $0.94 in · $4.69 out / 1M |
| Qwen 3.7 Plus | Budget 1M-context work | 1M tokens | No | $0.50 in · $2 out / 1M |
| Venice Uncensored 1.2 | Unfiltered responses | 128K tokens | Yes | $0.20 in · $0.90 out / 1M |
Pick Muse Spark Open when you want the strongest open-weight option for agentic coding and long-horizon engineering work, with vision in the loop and no prompt retention. If budget dominates, Qwen 3.7 Plus covers the same context length for a third of the price.
Pricing
Billed per token on Venice: $1.56 per 1M input tokens and $5.31 per 1M output tokens.
New Venice accounts include a free daily allowance and 500 welcome credits — no credit card required.
Getting a key
Create a key in API settings. Nothing else is required to start.
Export it as VENICE_API_KEY so the snippets above run unedited.
Point an existing OpenAI client at https://api.venice.ai/api/v1. The scheme is part of the value: an OpenAI client given a bare host does not resolve it.
Pass muse-spark-open as the model and send the request.
FAQ
Muse Spark Open is Meta's open-weight frontier model, released October 15, 2026. It is an 816B-parameter sparse mixture-of-experts model with about 42 billion parameters active per token, a 1M-token context window, and training focused on agentic coding and long-horizon tasks. It accepts text and image input and produces text output.
Venice bills Muse Spark Open per token: $1.56 per 1M input tokens and $5.31 per 1M output tokens, with cached input at $0.19 per 1M tokens. There is no per-message surcharge — cost scales with how much you send and generate.
The weights are openly downloadable under the Muse Model License, which permits free use, fine-tuning, distillation, and redistribution. It is not OSI-style open source, though: products must display "Built with Muse Spark Open," and providers offering it as a service above $20M in annual revenue must reach a separate agreement with Meta. On Venice you pay per token rather than for a license.
Yes to both. On Venice it exposes function calling, vision with multiple image inputs per request, web search, structured JSON-schema output, and extended reasoning. Audio and video input, which the underlying model supports per Meta's documentation, are not exposed through the Venice API.
Both are open-weight models with 1M-token context windows. Muse Spark Open is purpose-built for agentic coding workflows and costs less per output token ($5.31 vs $18.75 per 1M), while Kimi K3 is positioned for general long-context reasoning at a higher price. For software-engineering agents, Muse Spark Open is the natural pick; for broad reasoning work, compare both on your own tasks.
Legally yes — the Muse Model License allows running the weights yourself. Practically it is an 816B-parameter model, so Meta's reference desktop setup uses two Apple M5 Ultra Mac Studios, and most developers will prefer hosted inference like Venice's API instead of owning the hardware.
Yes — Venice runs it under the private tier with zero retention: prompts, code, and outputs are not stored, profiled, or used for training. It is not TEE-isolated or end-to-end encrypted on Venice, so privacy comes from the no-logging policy rather than hardware guarantees.
No. Unlike Venice Uncensored 1.2, Muse Spark Open ships with Meta's standard safety alignment intact, so Venice does not present it as an uncensored model. If unfiltered responses are the requirement, Venice Uncensored 1.2 is the model for that, at the cost of a much smaller 128K context.
1,024K tokens of context with up to 128K tokens of output per response on Venice — enough to hold a large codebase, an extended agent transcript, or hundreds of pages of documents in a single session.
One key, free to start, no credit card.