Muse Spark 1.3 on OpenCode Free: Setup, Limits and the Contributor Catch

Meta's Muse Spark 1.3 is available at zero token cost inside OpenCode as opencode/muse-spark-1-3-contributor-free. You connect OpenCode to OpenCode Zen, pick the free model from the picker, and code. The trade is clear: Meta may train on your prompts and completions, the free quota is dynamic and unpublished, and the model only works over the Responses endpoint.

If you want the short version, it is this. Install OpenCode, run /connect to add your Zen key, run /models and choose muse-spark-1-3-contributor-free, or launch with opencode -m opencode/muse-spark-1-3-contributor-free. Keep a paid model configured as fallback, because the free pool rotates without notice and rate limits arrive without warning.

What is Muse Spark 1.3?

Muse Spark 1.3 is Meta's agentic coding model, released on 2 September 2026 for Muse Code and the Meta Model API. Meta describes it as trained for agentic workflows, with strong competitive coding performance, long-context tracking across prior steps and results, and native multimodal input across images, documents and video. The technical context window is 1,048,576 tokens, marketed as 1M, with a large output allowance in the 128k class. It supports tool use, reasoning and structured output, which is exactly what an agent harness like OpenCode needs. Two Meta-direct SKUs exist: Standard, which Meta states is not used to improve its products, and Contributor, which is heavily discounted in exchange for permission to train future models on prompts and completions. OpenCode Zen adds a third route: Contributor Free, served at $0 input and $0 output for a limited feedback period.

The names look almost identical but the contracts differ. Muse-spark-1.3 is Standard. Muse-spark-1.3-contributor is the discounted Meta-direct tier at $0.10 per million input tokens and $0.20 output. Muse-spark-1-3-contributor-free, addressed as opencode/muse-spark-1-3-contributor-free, is the $0 Zen promotion. Same family, three different deals on price, data rights and limits.

Is muse-spark-1-3-contributor-free really free?

Yes, at the OpenCode layer. OpenCode's Zen pricing table lists Muse Spark 1.3 Contributor Free as Free, Free, Free per million tokens for input, output and cached reads, alongside six other $0 models including Muse Spark 1.2 Contributor Free, Big Pickle, MiMo-V2.5, Ling 3.0 Flash Fin and the Nemotron pair. An independent model mirror confirms zero input and output cost for the same ID. Do not confuse this with Meta's direct Contributor tier, which is discounted rather than free.

Free means no token charge, not no limits. Every free Zen model carries the same wording: available for a limited time while the team collects feedback. No end date, no notice promise. The pool rotates constantly, and Muse 1.3 Contributor Free joined at launch week on 2 September 2026. Treat the picker as the source of truth, never a bookmarked list.

How do I set it up on OpenCode?

Start with a standard OpenCode install, then authenticate to Zen, then select the model. First, install with curl -fsSL https://opencode.ai/install piped to bash, or with npm install -g opencode-ai on Node setups, brew on macOS, or choco or scoop on Windows. Then change into your project directory and start OpenCode. Second, sign in at opencode.ai/auth, copy your Zen API key, and inside the OpenCode TUI run /connect, choose OpenCode Zen and paste the key. The command-line alternative is opencode auth login, which stores credentials in your auth file, and you can verify with opencode auth list. Third, run /models in the TUI and filter for free, or run opencode models piped to a free filter on the CLI, with a refresh flag if the new entry is missing. Select opencode/muse-spark-1-3-contributor-free.

For repeatable setups, pin the model in config. The global file lives at opencode.json under your OpenCode config directory, and the per-project file is opencode.json in the project root, with project settings overriding global ones. Set the model field to opencode/muse-spark-1-3-contributor-free and optionally set small_model to another free ID for background subtasks. For one-off runs, launch non-interactively with the -m flag followed by the full provider and model ID, or use the run subcommand with the same flag and a quoted prompt. Switching mid-session is the same picker again via /models, or a config edit, or a relaunch with a different -m value.

What is the Contributor catch on data and privacy?

The Contributor contract trades price for training rights. Zen's footnote and Meta's pricing and rate-limits page state the position plainly: heavily discounted or free pricing in exchange for permission to use prompts and completions to train future Meta models. Meta's own table labels Contributor as used to improve products versus Standard as not used. In an agentic coding session this includes the code and files the agent reads, not just what you type.

The practical rule is simple. Use Contributor Free for learning, side projects, open-source work and anything you are comfortable being trained on. Move client work, proprietary codebases, credentials-adjacent material and regulated data to Standard or another non-training route, and only use Contributor SKUs for client work with explicit written consent that covers training use. An NDA alone is not that consent. Zen workspace admins can disable data-collecting models across a team, which is worth doing as a default if you mix personal and client projects on one machine.

What are the usage limits and why do credits not help?

There is no published free quota. Users encounter FreeUsageLimitError over HTTP 429, phrased as free usage exceeded with a suggestion to add credits or subscribe, after very different amounts of use. Some people code for hours before hitting it, others hit it within minutes at peak. A maintainer explanation puts it down to dynamic capacity shifting with demand, and community analysis has estimated the daily free budget at roughly thirty cents of equivalent value. Expect multi-hour retry windows rather than instant resets.

Adding Zen balance or subscribing does not lift the free-model cap. This is repeatedly confirmed across Zen issues: free models cannot consume paid balance, so the add-credits wording in the error is misleading and a steady source of complaints. If you hit the wall, the fix is to switch to a paid model for the rest of the session, try another free model, or go local with Ollama until the window resets. Keep that paid fallback configured before you need it, because discovering the error text mid-task is the worst moment to set up billing.

Why does the endpoint matter for Muse Contributor Free?

Muse Contributor Free models on Zen are Responses-API only. The correct base is the Zen Responses URL with the OpenAI SDK adapter, using the opencode provider prefix and the contributor-free model ID. Calls sent to the chat-completions path return a generic HTTP 500, which looks like an outage but is really a wrong-endpoint error. This trips people up because most other free Zen models accept the chat-completions route, so copied configs fail only for Muse. Prefer Zen unless you specifically need Standard's no-train contract.

Is Muse Spark 1.3 open-source or open-weight?

No. Muse Spark 1.3 is closed and API-only today, reached through Muse Code, the Meta Model API and routers such as Zen. This is the biggest provenance trap in current coverage. Several videos and posts present going open weight soon as fact. The reported position is narrower: Meta has signalled open weights on the roadmap and intends to publish weights for 1.2, while publication for 1.3 remains undecided. Separately, Muse Glimmer from August 2026 is open-weight under Apache 2.0, and some outlets conflate Glimmer with Spark. Open-source and open-weight are also different claims. For planning purposes, treat 1.3 as closed until Meta publishes weights or code itself.

Benchmark numbers need the same care. Widely repeated figures include 88.8 on Terminal-Bench 2.1 and 75.4 on DeepSWE, framed as beating named rivals. At least one tracker noted the DeepSWE figure did not appear on the official board on the same date, and launch charts often depend on undisclosed reasoning-effort and variant settings. Treat these as Meta-reported until independent, like-for-like reproductions land. They suggest a strong agent, not a settled leaderboard.

What should I do when it disappears or rate-limits?

Refresh and verify first. Reopen the /models picker or refresh the list, confirm resolved config, and re-run /connect on auth failures. If the model is gone, it has likely rotated out, so switch to another free model or your paid fallback. Never hard-code the free list into team docs. Keep a paid model one keystroke away.

The verdict is simple. It costs nothing in tokens, it costs your prompts as training signal, it can vanish or throttle without notice, and it demands the right endpoint. Set it up in minutes through Zen, enjoy the 1M-context agent for suitable work, and keep a paid, non-training model beside it for everything that matters. That two-model posture gives you the upside of the promo without letting its limits decide your deadline.