Muse Code & Muse Spark 1.2: harness co-training, behind Opus 5

Key takeaways

1. What is being launched. Muse Code (beta), a terminal coding agent installed via curl -fsSL https://dev.meta.ai/install.sh | bash on macOS and Linux, and Muse Spark 1.2, a model available in Muse Code and in Meta Model API.
2. ⭐⭐ The overall result, absent from the text and legible only in the images. — Meta loses all four comparisons: | Comparison | Winner | Muse Spark 1.2 | Rank | |---|---|---|---| | Terminal-Bench 2.1 | Opus 5 + Claude Code — 86.7% | 82.9% | 2nd / 6 | | DeepSWE 1.1 | Opus 5 + Claude Code — 65.0% | 59.3% | 3rd / 6 | | Meta Internal Coding Bench | Opus 5 — 79.4% | 70.6% | 2nd / 5 | | KDA case study (speedup) | Opus 5 — +74.0% | +68.7% | 4th / 6 | → On its own proprietary benchmark, Meta sits 8.8 points behind Opus 5. And on the case study it chose as a showcase, it also falls behind GPT 5.6 Sol (+71.2%) and behind Anthropic's previous generation, Opus 4.8 (+69.6%). This is the central fact of the announcement, and one has to go read the PNGs to find it.
3. ⭐⭐ The model's gain is smaller than announced, and the harness accounts for the rest. — the most useful reading of these figures:
4. On Terminal-Bench and DeepSWE, 1.1 is evaluated with mini-swe-agent and 1.2 with Muse Code. The displayed gap (76.2 → 82.9, i.e. +6.7 points) therefore conflates model progress and harness progress.
5. On the internal benchmark, the only table where no harness is indicated, the 1.1 → 1.2 gap drops to 68.3 → 70.6, i.e. +2.3 points. → The model gains about two points; the rest comes from the tooling around it. Meta half-admits as much by claiming co-training "to maximize harness compatibility". None of these comparisons is model versus model: they are model + harness pairs.
6. ⭐ And this empirically validates Mozilla's thesis, three weeks apart. [[mozilla-state-of-open-source-ai-2026-07]] wrote "the model is eating the harness", observed that "on every model where both appear, the lab's own harness now wins", and stated the mechanism — "a harness tuned tightly to one lab's weights becomes a fitted component of that lab's product… the tighter the tuning, the less swappable the weights underneath. Lock-in arrives as a side effect of optimization." Muse Code + Muse Spark 1.2 is exactly that object: Meta explicitly announces having trained the model with the harness to maximize their compatibility. The Terminal-Bench table shows it implicitly — each model there is measured with its own lab's harness.
7. The event log — the best piece of engineering in the post. "a local event log in which every model call, tool run, approval, and edit is appended", a single source of truth that makes the runtime "replay-exact and restart-safe". After a crash, the agent resumes exactly where it left off, which enables long tasks. → Direct convergence with the run manifest and the run resume command of [[skill-gibbs-hyperresearch-2026-08-03]], where the manifest is described as "your durable memory". Two independent teams arrive at the same mechanism: context doesn't survive, an on-disk log does.
8. ⭐ Persistent background agents — a break from the dominant model. instead of being created for one task and then destroyed, they "remain active throughout each session", which avoids "redundant information gathering", lets them decide on their own when to report back to the main agent, and reduces "latency and the need for steering". → To be compared with Hugo Lassiège's opposite observation on disposable sub-agents — "I use them less and less, recent agents delegate fairly targeted work on their own" ([[lassiege-usine-logicielle-heure-ia-2026-07-28]]). Two opposite answers to the same problem: Lassiège delegates less, Meta delegates to agents that no longer die.
9. ⭐ /grill ships out of the box, and the name is no accident. "/grill stress-tests that plan until it holds up". This is exactly the function of Matt Pocock's grill-with-docs skill, filed in this corpus since June ([[skill-pocock-grill-with-docs-2026-06]]) — an adversarial interview that puts a plan to the test before implementation. A pattern born in the skills community arrives preinstalled at a frontier lab. A sign of maturation: harness best practices are becoming product primitives.
10. The delivered triptych. — /plan (a plan submitted for approval), /grill (stress-testing), /goal (pursuit of the objective) — describes a plan → challenge → execution loop with a human gate at the first step. This is the canonical form documented by [[osmani-agent-harness-engineering-2026-04-19]], now packaged.
11. Self-improvement, read precisely. "We also used Muse Spark 1.1 to generate challenging coding environments and instruction-following templates. The model then graded candidate solutions… producing a scalable training dataset for Muse Spark 1.2." → It is 1.1 that generates the environments and grades, with 1.2 being the product of this dataset. ⚠️ The subject of "the model" is ambiguous, and some retellings have read it as 1.2; the natural reading is 1.1, the only one named. A distillation loop of one generation on itself, whose known limitation is that it can only teach what the previous generation already knows how to evaluate.
12. GPU kernel case study — the protocol is more interesting than the result. iterative optimization over more than 1,000 tool calls, up to 24 hours, on the KDA and MLA kernels for NVIDIA Hopper GPUs. The model writes, compiles, profiles, and progressively improves. ⭐ Notable constraint: "Models were prohibited from importing third-party kernel libraries such as FLA directly" — the algorithm had to be reimplemented in Triton, not wrapped around an existing implementation. This is what makes the test honest, and it is the kind of guardrail missing from many agent demonstrations.
13. On KDA: a block-parallel preparation kernel plus a sequential inter-block scan, with decay-cumulative recentering at the middle of the block.
14. On MLA: a two-kernel Triton pipeline, reusing the shared KV latent as both K and V, measured at batch size 1, 64 heads, sequence length 8,192, latent dimension 512.
15. ⚠️ KDA is Kimi's linear attention (cf. [[sfeir-kimi-k3-moonshot-frontier-open-weights-2026-07-16]]): Meta thus has its agent optimize the kernel of a competing open-weights architecture. A telling and flavorful detail of how architectures circulate.
16. ⚠️ Three methodological caveats on the charts. 1. No figure appears in the text. The four comparisons are raster images with no textual data or numeric accessibility labels. A hurried reader, an aggregator, or an agent reading the page sees no results. The values in this fiche were read visually off the PNGs. 2. The KDA chart's y-axis is non-linear, which Meta honestly flags ("Axis Spaced Linearly In X", gridlines at 0-27-43-53-60-65-69-72-75%). The scale stretches the upper part of the range, which visually amplifies the gaps between the leading models — to Meta's disadvantage, in this case. 3. The compared configurations are not homogeneous: max for Opus 5 and GPT 5.6 Terra, high for Grok 4.5 and Gemini 3.6 Flash, nothing stated for Muse Spark. And Grok is absent from the internal benchmark, GPT 5.6 Sol appears only in the case study. The panel's composition changes from one chart to the next.
17. ⚠️ The internal benchmark is unverifiable. "Meta Internal Coding Bench" is not public, its composition is not described in the post, and Meta uses it to measure its competitors. The fact that it loses on it makes the figure more credible than a flattering score — but a proprietary benchmark remains unreproducible.
18. What the announcement does not say. no pricing, no usage limits, no model license, no weight release — a notable break from Meta's Llama tradition, on which the post is silent. Nothing either on supported languages, context size, or confidentiality guarantees for submitted code.
19. Meta / related. empirical confirmation of the optimization-lock-in mechanism described in [[mozilla-state-of-open-source-ai-2026-07]]; event log and resume to be compared with [[skill-gibbs-hyperresearch-2026-08-03]]; preinstalled /grill against [[skill-pocock-grill-with-docs-2026-06]]; harness doctrine in [[osmani-agent-harness-engineering-2026-04-19]] and [[lassiege-usine-logicielle-heure-ia-2026-07-28]]; competitors cited in [[sfeir-gpt56-sol-terra-luna-coding-agentique-pricing-2026-07-13]] and [[cherny-wu-reflecting-year-claude-code-2026-07-17]]; KDA kernel and Kimi architecture in [[sfeir-kimi-k3-moonshot-frontier-open-weights-2026-07-16]].