AivexaNewsSearch
AI news for builders and product teamsChecked every hour

Meta Muse Spark Open is live on Together AI

Collected Oct 15, 2026

Together AI announced that Meta's Muse Spark Open, described as an open-weight model for agentic coding, is available on its platform starting today, with Together AI acting as a Day 0 launch partner. The company calls it the most capable US open-weight model to date, built for long-running agentic coding workflows that keep entire codebases in context.

The model is offered with a 1M-token context window on both serverless and dedicated infrastructure. According to Together AI, Muse Spark Open is trained on real developer work such as multi-file refactors, debugging sessions and long-running tasks, which the company says results in higher first-attempt accuracy, reliable tool calling and deep codebase understanding across the repositories developers actually work in.

Several capabilities stand out. The 1M-token context window is meant to hold entire projects, including complex dependencies, legacy code and thousands of files, without splitting work into pieces. The model also supports long-horizon autonomy, meaning it plans, acts, observes and self-corrects over long horizons and coordinates subagents for work that can be split up. It uses computer-use capabilities across desktop, mobile and browser, and browses the web with multimodal browsing. It accepts text, image, audio and audio-video input. Another feature is intent commentary: the model states what it is doing and why before acting, which Together AI says makes long runs easier to audit.

Pricing on Together AI's serverless offering is $1.25 per 1M input tokens, $4.25 per 1M output tokens and $0.15 per 1M cached input tokens. Developers can call the model through the Together API using the model identifier meta-models/Muse-Spark-Open-816B-A42B, or try it in the playground.

Two early access customers, Factory and Ollama, used Muse Spark Open before launch and both offer it to their users from day 0. Zimu Li, a member of technical staff at Factory, said enterprises choose Factory because they want their Droids running on the best model for the task, and that in their agentic engineering workflows Muse Spark Open consistently selected the right tool and kept responses compact, which Li described as the decisiveness production agents need. Li added that early access through Together AI let Factory stress-test the model ahead of launch.

Dong Chen, a senior AI engineer at Ollama, said millions of developers rely on Ollama as the destination to get up and running with open source models, and that Ollama is delighted to offer Muse Spark Open in partnership with Together AI. Chen said the model performs particularly well on tool calling and long context, including multi-step and parallel calls with clean structured outputs, and that partnering with Together AI for early access allowed Ollama to evaluate and optimize the model so developers experienced a fully supported, production-ready model from launch.

For developers, the practical appeal of a 1M-token context window is that an agent can keep an entire repository in view rather than retrieving and stitching together fragments. This is my own reading, but it suggests the model is aimed at workflows where the bottleneck is not a single prompt but a long sequence of steps that must stay coherent: refactors that touch many files, debugging that requires tracing dependencies, or agents that call tools and correct themselves over time. The intent commentary feature is a notable design choice because auditability is often a pain point in long-running agent runs; having the model explain its next action before taking it can make failures easier to diagnose. The support for text, image, audio and audio-video input also points to uses beyond pure code, such as document understanding and web automation, which Together AI explicitly mentions.

A likely trade-off is cost and latency at scale. The listed pricing of $1.25 per 1M input tokens and $4.25 per 1M output tokens is not trivial for high-volume agentic loops, and while cached input tokens at $0.15 per 1M can reduce repeat-context costs, caching behavior depends on how prompts are structured. Running entirely within a 1M-token window can also be slower and more expensive than retrieval-based approaches that send only relevant snippets. Teams will need to weigh the simplicity of whole-codebase context against those costs, and the dedicated infrastructure option gives an alternative to serverless for predictable workloads.

One detail worth noting is the model identifier, meta-models/Muse-Spark-Open-816B-A42B, which suggests a large mixture-of-experts style configuration, though the naming alone does not confirm architecture specifics. Together AI did not publish benchmark numbers in this announcement, so claims about first-attempt accuracy and tool-calling reliability rest on the company's description and customer comments rather than comparative evaluations. Developers evaluating the model should test it against their own repositories and tool-calling harnesses before standardizing on it.

Why it matters: Teams building coding agents now have a US open-weight option with a 1M-token context window available through Together AI on day one, including serverless access with published per-token pricing. That lowers the integration barrier for product teams that want to keep whole codebases in context without managing their own inference infrastructure, and the day-0 support from Factory and Ollama indicates some tooling ecosystem buy-in. The main work for adopters will be validating cost, latency and tool-calling behavior against their own workloads, since the announcement provides no independent benchmarks.

Read at Together AI

Based on reporting from the original publisher. Visit the source for full context and later updates.

Publisher excerpt

Meta Muse Spark Open, Meta's open-weight model for agentic coding, is live on Together AI with a 1M-token context window on serverless and dedicated infrastructure.