Meta Open-Sources Muse Spark Open on AI Gateway: A Million-Token Context Reasoning Model

Meta Open-Sources Muse Spark Open on AI Gateway: A Million-Token Context Reasoning Model

Meta's open-weight reasoning model Muse Spark Open has officially landed on AI Gateway. Built specifically for coding and agentic workflows, it covers scenarios like multi-file refactoring, code debugging, and understanding large codebases — giving developers an open-source option that combines long context with image understanding capabilities.

An Open-Source Reasoning Model Built for Coding

Muse Spark Open's positioning is clear: it's not a general-purpose chat model, but one designed to serve coding and agentic tasks. According to official information, it supports multi-file refactoring, debugging, and understanding large codebases — scenarios that traditional short-context models struggle to handle.

The "Open" in its name signals that this is an open-weight model, and developers can call it directly through AI Gateway without setting up a separate vendor account. For teams looking to reduce integration friction while keeping model behavior predictable, this is a notable shift.

Open Weights refers to the public release of a model's trained parameter files, which anyone can download, self-host, fine-tune, or run on their own infrastructure — in contrast to closed-source models accessible only via API. Open weights are not equivalent to fully open source, since training data, training code, and the complete training pipeline are often not disclosed together, and licenses may carry commercial restrictions. For developers, the core value of open weights lies in auditability, portability, and predictability: the ability to avoid depending on a single vendor, deploy in private environments to meet compliance requirements, or continue training for specific domains. Although Muse Spark Open is released as an open-weight model, the rollout described in this article uses a hosted version called through AI Gateway, letting developers gain inference capabilities without self-deploying — while retaining the option to self-host or fine-tune in the future.

1M Token Context + Mixed Text and Image Input

Muse Spark Open's most standout technical features are twofold.

First is the 1 million token context window. This scale means an agent can read a large number of source files in a single request, which is especially critical for cross-module dependency analysis, batch refactoring, and similar tasks.

Second is support for both text and image input. Developers can have the model process code and screenshots in the same request — for example, locating the corresponding code based on a UI screenshot, or troubleshooting by combining an error screenshot with source code. This combination of multimodality and long context aligns perfectly with the real needs of agentic coding tools.

Tokens and Context Windows: A token is the smallest unit a model uses to process text; one English word corresponds to roughly 1 to 1.5 tokens, and symbols, indentation, and line breaks in code are also split into tokens. The context window refers to the maximum total number of tokens a model can "see" at once in a single request, including both input and output. A 1 million token window is intuitively equivalent to thousands of pages of text or hundreds of thousands of lines of code, allowing the model to read through key parts of an entire mid-sized codebase at once without relying on vector retrieval or chunked concatenation. However, the longer the context, the higher the VRAM usage and compute cost during inference, so in practice you typically need to balance context length against latency and cost based on task complexity. Mixed text and image input means the model receives both text and image modalities within the same request; under the hood, a vision encoder typically converts images into embedding vectors, which are then fed together with text tokens into a unified decoding process.

Zero Data Retention, No Additional Account Needed

On the data compliance front, Muse Spark Open supports ZDR (Zero Data Retention) when provided through AI Gateway, and by default does not use request data for training.

One practical detail: developers don't need to create a separate vendor account for this model. This is consistent with AI Gateway's overall design philosophy — serving as a unified entry point that consolidates access, billing, and routing for multiple models under a single API.

ZDR (Zero Data Retention) is a commitment aimed at enterprise compliance: after processing a request, the service provider does not persistently store the request content or model output, discarding it after completing inference in memory only. This is a different guarantee from "not used for training" — the latter concerns whether data enters model iteration, while the former concerns whether data leaves a copy on the server. For handling private codebases or screenshots containing sensitive information, ZDR is often a key condition for passing internal security review. BYOK (Bring Your Own Key) refers to developers using a key they obtained from the model vendor, forwarding requests through an aggregation platform, with costs settled directly to the original vendor without the platform adding a markup. Both are often used together with unified gateways, budget controls, and routing rules, forming the compliance and cost governance foundation for enterprise-grade LLM access.

Compatible with Mainstream APIs and Coding Tools

For invocation, the model name is uniformly meta/muse-spark-open, and it can be used across all API formats and coding agents. Currently supported interfaces include:

• AI SDK: Suitable for direct integration in JavaScript/TypeScript projects
• OpenAI-compatible Chat Completions and Responses API: Makes migration easier for existing OpenAI ecosystem code
• Anthropic-compatible Messages API: Connects with Claude-style invocation methods

On the tooling side, coding agents like Claude Code, Codex, and Cursor can all connect. The process is to first upgrade Vercel CLI, then run the coding agent's configuration command, and afterward select meta/muse-spark-open in the agent.

The Value of AI Gateway: Unified Entry Point and Cost Transparency

The significance of this launch goes beyond simply adding another model. AI Gateway itself provides a unified API for calling models, tracking usage and costs, and configuring retries, failover, and performance optimization to improve availability relative to single vendor endpoints.

It also includes features like custom reports, zero data retention support, API key budgets, and routing rules. On billing, AI Gateway charges at vendor pricing without extra markup, and does not charge a platform fee for inference, including BYOK (Bring Your Own Key) requests.

For development teams, this means they can experiment with a new open-source reasoning model in their coding agent pipeline at low management cost, while retaining room for switching and cost control.

Usage Recommendations

If your team is building code comprehension, automated refactoring, or screenshot-based debugging tools, Muse Spark Open's million-token context and mixed text-image input are worth testing. We recommend starting with the model playground for small-scale validation, then gradually integrating into production with AI Gateway's budget and routing rules.

One-line Summary

Meta's open-weight reasoning model Muse Spark Open lands on AI Gateway, supporting 1M context and text-image input for coding and agentic workflows.

Paragraph Summary

Meta's open-weight reasoning model Muse Spark Open is now available through AI Gateway, positioned as a specialized model serving coding and agentic workflows. It supports a 1 million token context window capable of processing large codebases in one pass, and accepts mixed text and image input for locating code from screenshots or troubleshooting with error screenshots. On compliance, the model offers ZDR zero data retention and is not used for training by default, and developers don't need to create a separate vendor account. The model name is uniformly meta/muse-spark-open, compatible with AI SDK, OpenAI Chat Completions/Responses, and Anthropic Messages API, and can connect to coding agents like Claude Code, Codex, and Cursor. AI Gateway charges at vendor pricing without extra markup and provides usage tracking, budgets, routing, and failover capabilities.