---
type: analysis
slug: research-muse-spark-1-2-was-co-trained-with-muse-code-cline-shipped-the-model-not-a
title: "Muse Spark 1.2 was co-trained with Muse Code; Cline shipped the model, not a verified prompt transplant"
description: "Meta’s 5 August 2026 launch says Muse Spark 1.2 was co-trained with the Muse Code harness and toolset. The next day Cline made the contributor model selectable. Docker browser sign-in is a documented non-interactive gap with an official API-key path; claimed 2.7x token savings from ported Muse instructions were not verified from a Cline-owned primary page."
published_at: 2026-11-08T00:00:00.000Z
topics: ["cli","tools","agents","research"]
source_urls: ["https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2","https://research.meta.ai/static/muse-spark-1-2-methodology","https://artificialanalysis.ai/articles/muse-spark-1-2","https://dev.meta.ai/models/muse-spark-1-2","https://github.com/cline/cline/releases/tag/v4.1.6","https://github.com/cline/cline/releases/tag/cli-v3.0.51","https://github.com/cline/cline/releases/tag/sdk/sdk/v0.0.71","https://simonwillison.net/2026/Aug/5/muse-code-and-muse-spark-12/","https://dev.meta.ai/docs/muse-code/auth","https://dev.meta.ai/docs/muse-code","https://dev.meta.ai/docs/muse-code/extending","https://github.com/shelajev/muse-code-sbx-kit","https://github.com/meta-models/muse-code-sdk/issues/38"]
author: "Andrey Reshetnikov"
author_profile: https://newruntime.com/owner-profile.md
---
On 5 August 2026 Meta released Muse Code, a beta terminal coding agent, and Muse Spark 1.2, a coding-focused update to Spark 1.1. The launch post says the model was co-trained with Muse Code and is also available on the Meta Model API, so other agents can call the same checkpoint without running Meta’s CLI. Muse Code authenticates with a browser session or an API key, defaults to muse-spark-1.2, and can run headless with muse exec.

The next day Cline described trying Muse Code, failing to sign in from Docker, extracting instructions from Muse Code’s system prompt, and adding them to Cline while using Spark 1.2 on a real bugfix. That anecdote is a live test of a now-common builder question: whether co-training gains travel with the weights into a foreign harness, or only appear when the original prompt, tools, compaction, and runtime stay together. The answer changes whether you buy Muse Code, wrap Spark in your own agent, or copy someone else’s prompt fragment.

## What we checked

- **What exactly did Cline add or release as a result of the extracted Muse instructions?** partly confirmed. On 6 August 2026 Cline shipped meta/muse-spark-1.2-contributor in IDE v4.1.6, CLI v3.0.51, and SDK v0.0.71. Those release notes add the model to the Cline provider catalog and do not mention extracted Muse system-prompt text, a new default harness, or a published experiment artifact. [5][6][7]
- **Is there a Meta primary source confirming Muse Spark 1.2 was co-trained with the Muse agent harness?** confirmed. Meta’s launch post states it co-trained Muse Spark 1.2 with Muse Code so the pair would perform best together, using rejection-sampled harness trajectories, recipe work for goals, compaction, and subagents, and the Muse Code toolset. Independent write-ups quote that same co-training paragraph. [1][8]
- **Is the Docker sign-in bug independently reproducible or acknowledged by Meta?** partly confirmed. Meta documents first-run browser sign-in or an API key, and tells CI and other non-interactive environments to set META_API_KEY. It does not document Docker and did not acknowledge a Docker-specific login bug at launch. A later, separate 1.4.0 credential-save bug on Windows and headless Linux was acknowledged and fixed in 1.4.1. An independent Docker Sandboxes kit shows container use via a printed verification URL on the host browser or META_API_KEY. [9][10][13][12]
- **Does calling Muse Spark 1.2 from a foreign agent recover Muse Code’s co-training gains without the Muse Code runtime?** unresolved. Meta’s own evaluations pair Spark 1.2 with Muse Code and warn that other agents’ tools and system prompts may not be tuned. Artificial Analysis measured Spark 1.2 in its Stirrup harness and reported only a two-point Terminal-Bench v2.1 move from 78% to 80%. That is consistent with harness dependence, but it does not test Cline’s claimed prompt-only transplant. [2][3]
- **Did Cline publish a reproducible prompt-port experiment showing large token, time, and cost cuts on the same Muse Spark 1.2 task?** unresolved. Cline’s original X post returned 403 and the LinkedIn post timed out, so those numbers could not be read from a Cline-owned page in this session. Cline’s same-day GitHub releases do not include experiment logs, the extracted prompt, the bug that was fixed, run count, or token and cost figures. [5][6]

## What we found

- Meta’s primary launch post is explicit about co-training. Muse Spark 1.2 was trained with Muse Code using rejection-sampled harness trajectories, recipe optimizations for goals, compaction, and subagents, and integration of the Muse Code toolset “to maximize harness compatibility.” The same post also describes persistent background agents, a local append-only event log, and bundled skills including /plan, /grill, and /goal. That is a model-plus-product stack, not a generic API wrapper. [1]
- Meta’s evaluation methodology pairs Spark 1.2 with Muse Code on Terminal-Bench 2.1 and DeepSWE v1.1, while Spark 1.1 is run with mini-swe-agent and competitors with their own agent products. GDPVal-AA v2 and MCP Atlas use the benchmark providers’ harnesses instead. Meta warns that its tools and system prompts may not be tuned for third-party models. Cross-agent score gaps therefore mix model quality with harness fit. [2]
- Artificial Analysis, using its Stirrup harness rather than Muse Code, scored Spark 1.2 at 54 on its Intelligence Index, up 3 points from 1.1, with Terminal-Bench v2.1 moving from 78% to 80%. GDPval-AA v2 Elo rose 260 points to 1631. Standard API pricing on the model page remains $1.25 input, $0.15 cached input, and $4.25 output per million tokens, with a 1M context window. Independent agentic gains exist, but they are smaller on terminal coding than Meta’s harness-paired framing would suggest, and they are not a Cline prompt-port result. [3][4]
- Cline’s verified product change on 6 August 2026 was catalog support. IDE v4.1.6, CLI v3.0.51, and SDK v0.0.71 make meta/muse-spark-1.2-contributor selectable on the Cline provider. That matches Meta’s dual-SKU pattern, which Simon Willison recorded at launch as muse-spark-1.2 at $1.25/$4.25 versus a contributor ID at $0.10/$0.20 if Meta may use the data. It is not evidence that Cline merged Muse Code’s system prompt into its default agent. [5][6][7][8]
- Muse Code sign-in is specified as browser approval or paste-an-API-key. For CI, Meta says set META_API_KEY; an environment key outranks a stored browser session. Headless runs use muse exec. A third-party Docker Sandboxes kit installs the official muse binary and tells users to complete the printed URL in a host browser or to use META_API_KEY, persisting ~/.config/muse/auth.json. Browser OAuth without a usable browser is expected to fail; that is not the same as Meta admitting a Docker bug. [9][10][11][12]
- In late September 2026, issue 38 on meta-models/muse-code-sdk reported that Muse Code 1.4.0 completed device login then failed to save credentials on Windows and headless Linux with “keychain write failed (internal error -2147483648).” A Meta commenter said 1.4.0 took a macOS-only save path on those platforms and that 1.4.1 writes ~/.config/muse/auth.json again. That is a real, acknowledged login bug, weeks after Cline’s Docker report, and it is a credential-backend defect rather than a missing TTY. [13]

## What it means

Treat Muse Spark 1.2 and Muse Code as a co-trained pair. If you call the model from Cline, your own loop, or another CLI, Meta’s methodology already implies you should not expect Muse Code’s published coding scores. Measure the model in your tools, prompts, compaction, and subagent scheme. Contributor versus standard IDs are a data-use and price choice, not a harness transplant. For containers and CI, skip browser OAuth and inject META_API_KEY; if you need an interactive session, complete device login on a host with a browser and persist auth.json, as the independent sandbox kit does.

Do not operationalize recapped “Muse system prompt” bullets as Meta’s training recipe. Cline’s GitHub did not ship those instructions, Meta did not publish the Muse Code system prompt as a portable spec, and the original social experiment could not be opened here. If you want the co-training surface, run Muse Code or rebuild the behaviors Meta named: goal conditioning, compaction, subagents, and the actual toolset, then A/B them on your tasks.

## Still open

- Cline’s original experiment post, extracted prompt text, task identity, run count, and token or cost logs were not available from a Cline-owned page in this session.
- Whether restating Muse-style engineering rules in a foreign system prompt recovers a stable fraction of Muse Code’s co-training benefit, or only changes one unrepeated bugfix run.
- Whether Meta will document an official Docker or TTY-less login path beyond META_API_KEY.
- How much of Spark 1.2’s coding lift is the 1.2 checkpoint versus Muse Code’s runtime, skills, observers, and toolset, given that Meta’s 1.1 comparison used a different agent.

## Sources

1. [Introducing Muse Code and Muse Spark 1.2 | Meta AI Research](https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2) (primary)
2. [Muse Spark 1.2 & Muse Code Methodology](https://research.meta.ai/static/muse-spark-1-2-methodology) (primary)
3. [Muse Spark 1.2: Benchmarks and analysis | Artificial Analysis](https://artificialanalysis.ai/articles/muse-spark-1-2)
4. [Muse Spark 1.2 | Meta](https://dev.meta.ai/models/muse-spark-1-2) (primary)
5. [Cline v4.1.6](https://github.com/cline/cline/releases/tag/v4.1.6) (primary)
6. [Cline CLI v3.0.51](https://github.com/cline/cline/releases/tag/cli-v3.0.51) (primary)
7. [Cline SDK v0.0.71](https://github.com/cline/cline/releases/tag/sdk/sdk/v0.0.71) (primary)
8. [Introducing Muse Code and Muse Spark 1.2 — Simon Willison](https://simonwillison.net/2026/Aug/5/muse-code-and-muse-spark-12/)
9. [Authentication and billing - Meta Model API](https://dev.meta.ai/docs/muse-code/auth) (primary)
10. [Muse Code - Meta Model API](https://dev.meta.ai/docs/muse-code) (primary)
11. [Extending and automating - Meta Model API](https://dev.meta.ai/docs/muse-code/extending) (primary)
12. [shelajev/muse-code-sbx-kit](https://github.com/shelajev/muse-code-sbx-kit)
13. [muse login 1.4.0 can't save the credential on Windows and headless Linux](https://github.com/meta-models/muse-code-sdk/issues/38) (primary)

Author: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- topic hub: [Tools](https://newruntime.com/topics/tools.md): Explore the tools topic hub.
- topic hub: [Agents](https://newruntime.com/topics/agents.md): Explore the agents topic hub.
- collection: [Analysis](https://newruntime.com/analysis.json): Open the latest authorial analysis.
- collection: [Sources](https://newruntime.com/sources/): See the original sources this archive cites.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)