← All guides

Muse Spark 1.3 in OpenCode: use Meta's free coding model (4 steps)

What Muse Spark 1.3 actually is, why OpenCode Zen lists it at $0, the exact setup, and the one catch (your data may train Meta's models).

Muse Spark 1.3 is Meta's agentic coding model with a 1M-token context window and native vision. On Meta's Model API the standard tier costs $1.25 input / $4.25 output per million tokens; the Contributor tier ($0.10 / $0.20) is cheap because your prompts and outputs may be used to train Meta's models, and OpenCode Zen lists that tier free at $0. Setup takes four steps: register at opencode.ai, create an API key, run /connect in OpenCode, then pick muse-spark-1.3-contributor-free in /models.

Meta shipped Muse Spark 1.3 in September 2026 as its coding model for agents, and OpenCode users can run it for free: Zen's model list includes a $0 "Muse Spark 1.3 Contributor" entry. Free here has a real price tag attached, though, one that is printed in Meta's own docs rather than a pricing page: on the Contributor tier your prompts and outputs may be used to train Meta's future models. This post covers what the model is, whether the benchmarks justify the hype, the exact four-step OpenCode setup, and when you should pay for the standard tier instead.

What Muse Spark 1.3 actually is

Muse Spark 1.3 is a model trained for long-horizon, agentic coding work: it tracks context and prior results across a long task, works through messy or conflicting inputs, and asks for clarification instead of guessing. It also perceives video, images, and documents natively, so you can feed it a screenshot or a screen recording and ask it to build what it sees. Every version carries a 1M-token context window.

Why developers are looking at it

Frontier-class coding benchmarks

Meta's official tables put it at 75.4 on DeepSWE v1.1 long-horizon agentic coding, ahead of Claude Opus 5 (max) at 74.0 and GPT-5.6 Sol (max) at 73.0.

Extremely long context that holds up

98.5 on MRCR 256K–512K and 98.1 on MRCR 512K–1M — the long-context scores are where it separates most clearly from the field.

Built for tool-heavy agent loops

Trained for reliable tool calling with fewer unnecessary turns, which is exactly what burns tokens (and patience) in day-to-day agent use.

Native multimodal perception

Video, images, and documents go in directly; its visual reasoning runs through a real execution environment rather than scripted steps.

Benchmarks: Muse Spark 1.3 vs the models you're weighing it against

BenchmarkMuse Spark 1.3GPT-5.6 SolClaude Opus 5
DeepSWE v1.1 (agentic coding)75.473.074.0
MRCR 256K–512K (long context)98.591.5—
MRCR 512K–1M (long context)98.173.8—
Terminal-Bench 2.188.888.886.7
OSWorld 2.0 (computer use)90.393.190.4

Meta's published benchmark tables (max reasoning settings). It wins the agentic-coding and long-context rows; GPT-5.6 Sol still leads OSWorld. Treat any single number as a starting point, not a verdict.

Why OpenCode can list it at $0

Meta's Model API sells the same model twice, at wildly different prices. The difference is not a coupon, it is consent: the Contributor tier is explicitly "used to improve our products," meaning your prompts and outputs feed Meta's training data. The standard tier is "not used to improve our products" and costs about 12x more per input token. OpenCode's special arrangement with Meta puts the Contributor tier on Zen free, and OpenCode's $10/month Go plan includes it as well.

The two tiers (Meta Model API list prices, per 1M tokens)

TierInputOutputYour data trains Meta?
muse-spark-1.3-contributor$0.10$0.20Yes — that is the trade
muse-spark-1.3 (standard)$1.25$4.25No
OpenCode Zen listing$0$0Yes (Contributor tier)

Cached input is $0.002 / $0.15 respectively, and both tiers have a 1M-token context window. Zen's free entry is the Contributor tier — the $0 is paid for with data, not with nothing.

Do not paste secrets, personal data, production customer data, or anything your company forbids from leaving into the free Contributor tier. If your policy forbids training use, the standard tier ($1.25 / $4.25) is the compliant choice — "free" is only free if your data is worth less to you than the token savings.

Set it up in OpenCode: 4 steps

From zero to Muse Spark 1.3 in about two minutes

  1. 1. Register an OpenCode accountSign up at opencode.ai. OpenCode Zen is pay-as-you-use with $0 free-tier models; the Go plan is a flat $10/month that also includes this model (available in limited regions).
  2. 2. Create an API keyIssue a key in the Zen console. Go subscribers reuse the same key. The endpoint is OpenAI-compatible, so the key works in any OpenAI-compatible client, not just OpenCode.
  3. 3. Connect inside OpenCodeIn the OpenCode TUI run /connect, choose OpenCode Zen, paste the key.
  4. 4. Pick the model and goRun /models and select muse-spark-1.3-contributor-free, then work normally. It is strongest on large-repo refactors, cross-file changes, and long agentic tasks where its context tracking pays off.

Where else you can run Muse Spark 1.3

OpenCode is the free route, not the only one. Meta's Model API is a self-serve, OpenAI-compatible endpoint, so any tool that accepts a custom OpenAI base URL can point at it with your own key. OpenRouter carries the Muse model family if you'd rather swap it into an existing workflow. Meta also ships Muse Code, a multi-agent terminal coder built for these models. One honest caveat: Claude Code speaks the Anthropic API format, so running Muse Spark there requires an OpenAI-compatible tool (OpenCode, or an agent like Wuwei) or a translation proxy in between — it is not a drop-in Claude replacement for that specific client.

The model-agnostic way to try it

The reason guides like this keep appearing is that no single model wins everywhere: Muse Spark 1.3 leads long-context and agentic coding, GPT-5.6 Sol leads computer use, Claude leads some reasoning tasks. Wuwei is a free, open-source (MIT) coding agent that runs locally on Windows, macOS, and Linux and is model-agnostic by design. You can point it at any OpenAI-compatible endpoint — Meta's Model API included — with your own key, or start on its free hosted models with no login at all. Swapping Muse Spark in and out as the benchmarks shift takes one setting change, not a new tool. Download at /en#download.

FAQ

Is Muse Spark 1.3 really free?

On OpenCode Zen, yes: its model list includes muse-spark-1.3-contributor-free at $0, and the $10/month OpenCode Go plan includes the Contributor tier too. The catch is built into the tier itself — Contributor means your prompts and outputs may be used to train Meta's future models. The standard tier that excludes your data costs $1.25 input / $4.25 output per 1M tokens on Meta's Model API.

What is the difference between the Contributor and standard tiers?

Identical model, different data terms. The Contributor tier ($0.10 / $0.20 per 1M tokens, free on OpenCode Zen) allows Meta to use your prompts and outputs to improve its products. The standard tier ($1.25 / $4.25) explicitly does not use your data for training. Pick Contributor for personal projects and learning; pick standard for anything confidential.

Is Muse Spark 1.3 better than Claude Opus 5?

On Meta's published benchmarks it trades blows: 75.4 vs 74.0 on DeepSWE v1.1 agentic coding in Muse Spark's favor, a large lead on MRCR long-context tasks, while GPT-5.6 Sol still leads OSWorld computer use. Benchmarks from a vendor's own table are a starting point, not a verdict — the real question is whether it is good enough for your workload at your price, and at $0 on OpenCode Zen the trial costs nothing.

Can I use Muse Spark 1.3 with Claude Code?

Not directly. Claude Code talks the Anthropic API format, while Muse Spark 1.3 is served through OpenAI-compatible endpoints (Meta Model API, OpenCode Zen, OpenRouter). You would need an OpenAI-compatible client such as OpenCode or Wuwei, or a translation proxy between the two formats.

Does OpenCode Go include Muse Spark 1.3?

Yes — the $10/month Go plan lists the Contributor tier (in limited regions) alongside its other models on a generous rolling request quota. The free Zen route also works if you're fine with pay-as-you-use pricing elsewhere in the catalog.

Muse Spark 1.3 is a frontier coding model you can try for $0 on OpenCode Zen, paid for with training data rather than money, or for about $0.10/$0.20 per million tokens direct from Meta if you keep your data. The four steps above get you running; the tier you pick decides who sees your prompts.

Windows · macOS · Linux · free, no login