Summary: Meta’s Muse Spark Open is now available on Together AI, which served as a Day 0 launch partner and ran an early access program before the October 15, 2026 launch. The model targets long-running agentic coding with a 1M-token context window, and it’s callable serverless at $1.25 per 1M input tokens. If you’re building agents that hold whole codebases, the pricing and tool-calling behavior are the two things to check first.
What the model is built for
Muse Spark Open was trained on real developer work — multi-file refactors, debugging sessions, and long-running tasks, according to Together AI’s launch post. The claimed results are higher first-attempt accuracy, reliable tool calling, and codebase understanding across repositories rather than isolated files.
The spec sheet reads like a checklist for agent builders:
- A 1M-token context window that holds entire projects, including legacy code and dependencies, without chunking
- Long-horizon autonomy: planning, acting, observing, and self-correcting, with subagent coordination for parallelizable work
- Computer use across desktop, mobile, and browser, plus multimodal browsing
- Text, image, audio, and audio-video input
- Intent commentary — the model states what it’s doing and why before acting, which makes long runs easier to audit
That last point is easy to overlook but matters in production. If your eval pipeline parses traces, an explicit pre-action rationale gives you a free signal for post-hoc review. It pairs well with the discipline we described in our piece on starting agent evals with ten traces.
What early access showed
Two customers tested the model before launch through Together AI’s early access program: Factory and Ollama, both offering it from day 0.
Zimu Li, a member of technical staff at Factory, reported that in their agentic engineering workflows the model “consistently selected the right tool and kept its responses compact” — the decisiveness they look for in production agents.
Dong Chen, senior AI engineer at Ollama, highlighted tool calling and long context specifically, including multi-step and parallel calls with clean structured outputs. For a platform whose whole job is serving open models to millions of developers, structured-output cleanliness is the difference between a demo and a shippable integration.
Note these are quoted customer impressions from the launch post, not independent benchmarks. If tool-calling reliability is your bottleneck, run your own harness before committing.
Running it yourself
Together AI serves it serverless under the model ID meta-models/Muse-Spark-Open-816B-A42B. The client call is standard chat completions:
from together import Together
client = Together()
response = client.chat.completions.create(
model="meta-models/Muse-Spark-Open-816B-A42B",
messages=[{"role": "user", "content": "Refactor this module to remove the circular import."}],
)
print(response.choices[0].message.content)Pricing is $1.25 per 1M input tokens, $4.25 per 1M output tokens, and $0.15 per 1M cached input tokens. That cached rate is the number to model if your agents replay large codebase context across many turns — the savings compound fast in long-horizon workflows.
How it fits the Muse Spark line
This continues a pattern we’ve tracked across the family: Meta keeps pushing open-weight models that assume an agent harness, not just a chat box. Muse Spark 1.3 shipped agents that ask questions and know their limits, and Muse Spark 1.1 brought agentic tooling to the Meta Model API. Spark Open extends that into a 1M-context, multimodal open-weight package hosted by a third party — meaning you can evaluate it against your own traces the same day you read the announcement.
The practical next step: run one of your hardest long-horizon tasks through it, compare first-attempt success and tool-call error rates against your current model, and let the cached-input economics decide whether it earns a slot in your routing table.
Sources
View the original post on X