# meta-muse-code-muse-spark-1-2-2026-08-05

## Veille

Announcement from **Meta AI Research** published on **August 5, 2026** (stated reading time: 4 minutes, no individual byline): **Muse Code** in beta, *"a terminal coding agent"*, and the model that powers it, **Muse Spark 1.2**. Meta itself frames the launch: *"This marks our next step toward the frontier, with larger and much more capable models on the way"* — **a step toward the frontier, not a big claim**. **Three architectural elements** are described on the harness side: **asynchronous background agents** that *"remain active throughout each session, rather than being spawned for individual tasks"*, to avoid redundant information gathering and reduce the need for steering; a **local event log** where *"every model call, tool run, approval, and edit is appended"*, making the runtime *"replay-exact and restart-safe"* — able to resume exactly where it left off after a crash; and **three skills shipped out of the box** — `/plan` (turns a task into a plan submitted for approval), **`/grill`** (stress-tests the plan *"until it holds up"*), and `/goal`. On the model side, Meta claims **co-training the model with the harness** (*"to maximize harness compatibility"*, with harness trajectories sampled via rejection sampling and recipe optimizations for goals, compaction, and sub-agents), **long-horizon** training (whole-repository generation, end-to-end projects, self-research, with planning, goal conditioning, and **context compaction**), and a **self-improvement loop** in which Muse Spark 1.1 generates the environments and instruction templates **and then grades candidate solutions**, producing a training set for 1.2. ⚠️ **The most notable fact in this announcement is written nowhere in its text**: the four published charts — Terminal-Bench 2.1, DeepSWE 1.1, an internal Meta benchmark, and the GPU kernel optimization case study — **place Muse Spark 1.2 behind Opus 5 in all four cases**, including on Meta's own proprietary benchmark (70.6% vs 79.4%) and on the case study, where the model finishes **fourth out of six** (+68.7% vs +74.0%). ⚠️ **And the model's actual gain is smaller than it appears**: on the two public benchmarks, 1.1 is measured with `mini-swe-agent` and 1.2 with Muse Code — the 6.7-point gap therefore conflates model and harness. On the internal benchmark, the only comparison where no harness is mentioned, the 1.1 → 1.2 gap drops to **2.3 points**.

## Titre Article

Introducing Muse Code and Muse Spark 1.2

## Date

2026-08-05

## URL

https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2

## Keywords

Meta AI Research, Muse Code, Muse Spark 1.2, terminal coding agent, beta, harness, asynchronous background agents, persistent sub-agents, redundant information gathering, event log, event log, replay-exact, restart-safe, crash recovery, long tasks, shipped skills, approved plan, grill, goal, model-harness co-training, rejection sampling, harness trajectories, harness compatibility, long-horizon, whole-repository generation, self-research, goal conditioning, context compaction, self-improvement, environment generation, solution grading, Terminal-Bench 2.1, DeepSWE 1.1, internal benchmark, Opus 5, Claude Code, GPT 5.6 Terra, GPT 5.6 Sol, Codex, Grok 4.5, Gemini 3.6 Flash, Antigravity CLI, mini-swe-agent, GPU kernel optimization, Triton, KDA, MLA, NVIDIA Hopper, cumulative tool calls, speedup vs baseline, FLA, kernel fusion, tiling, Meta Model API, optimization lock-in

## Authors

**Meta AI Research** — publication institutionnelle sans auteur nommé, sur `research.meta.ai`. Le billet renvoie à un **rapport** pour la méthodologie d'évaluation, non repris ici.

**Positionnement à connaître** : Meta arrive sur le terrain de l'agent de codage en terminal **après** Claude Code, Codex, Gemini CLI, Grok Build et Antigravity CLI, et le dit — *« our next step toward the frontier »*. L'annonce couple un **produit** (Muse Code, distribué par `curl … | bash`) et un **modèle** (Muse Spark 1.2, disponible dans Muse Code et dans **Meta Model API** avec *« expanded global access »*), les deux ayant été entraînés ensemble.

## Ton

**Profile**: a research-lab product announcement, short format, sober technical register. No hyperbole, no claim of superiority, no superlatives — which is **unusual for a launch** and is itself informative.

**Style**: the text proceeds by **describing mechanisms**, section by section, each held to one paragraph and naming a precise engineering decision — sub-agent persistence, the event log, co-training, compaction. There is no customer case, no executive quote, no productivity promise. The only narrative demonstration is a vignette: a user drops an `mp4` video of a house viewing into the terminal, Muse Code interprets it, and produces a short-term rental booking page.

**The most notable trait is the modesty of the framing**, set out as early as the second sentence: *"This marks our next step toward the frontier, with larger and much more capable models on the way."* Meta announces a milestone and warns that the best is yet to come. The conclusion repeats the same stance: *"We have a lot on the horizon, including new harness features and more powerful models."*

**This modesty is borne out in the charts**, and that is what makes the announcement distinctive: Meta publishes four comparisons in which it never wins. Few launches are willing to show a competitor in the lead on their own page — even less so on their **own internal benchmark**.

**Marker phrases**: *"our next step toward the frontier"*, *"rather than being spawned for individual tasks"*, *"replay-exact and restart-safe"*, *"/grill stress-tests that plan until it holds up"*, *"to maximize harness compatibility"*.

## Pense-betes

- **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**.
- **⭐⭐ 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.
- **⭐⭐ The model's gain is smaller than announced, and the harness accounts for the rest** — the most useful reading of these figures:
- 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**.
- 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.**
- **⭐ 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.
- **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.**
- **⭐ 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.
- **⭐ `/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.
- **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.
- **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.
- **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.
- On **KDA**: a block-parallel preparation kernel plus a sequential inter-block scan, with decay-cumulative recentering at the middle of the block.
- 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.
- ⚠️ **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.
- **⚠️ 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.**
- **⚠️ 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.
- **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.
- **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]].

## RésuméDe400mots

Announcement from **Meta AI Research** on **August 5, 2026**: **Muse Code** in beta, a terminal coding agent, and **Muse Spark 1.2**, the model powering it. Meta itself frames the launch — *"our next step toward the frontier, with larger and much more capable models on the way"*.

**On the harness side, three decisions.** **Asynchronous background agents** that *"remain active throughout each session, rather than being spawned for individual tasks"*, avoiding redundant information gathering and deciding on their own when to report back to the main agent. A **local event log** recording every model call, tool run, approval, and edit, which makes the runtime *"replay-exact and restart-safe"*: after a crash, the agent resumes exactly where it left off. And three **skills shipped out of the box**: `/plan` (a plan submitted for approval), **`/grill`** (stress-tests the plan until it holds up), and `/goal`.

**On the model side**, Meta claims **co-training with the harness** *"to maximize harness compatibility"*, long-horizon training (whole repository, end-to-end projects, self-research, context compaction), and a self-improvement loop in which version 1.1 generates the environments and grades the solutions, producing the training set for 1.2.

**The central fact of this announcement is written nowhere in its text.** The four published comparisons exist only as images, and they place Muse Spark 1.2 **behind Opus 5 in all four cases**: 82.9% vs 86.7% on Terminal-Bench 2.1, 59.3% vs 65.0% on DeepSWE 1.1, **70.6% vs 79.4% on Meta's own internal benchmark**, and +68.7% vs +74.0% on the GPU kernel optimization case study, where the model finishes **fourth out of six**, behind GPT 5.6 Sol and behind Anthropic's previous generation.

**And the model's own gain is smaller than it appears.** On the two public benchmarks, version 1.1 is evaluated with `mini-swe-agent` and 1.2 with Muse Code: the 6.7-point gap conflates model and harness. On the internal benchmark, the only comparison with no harness indicated, it drops to **2.3 points**.

The announcement therefore stands mainly as **empirical confirmation** of a thesis already stated: value is shifting toward the harness, and a harness co-trained with its own weights makes those weights all the less interchangeable.

## GrapheDeConnaissance

- Meta AI Research —publie→ Muse Code (TECHNOLOGIE, 0.97)
- Meta AI Research —publie→ Muse Spark 1.2 (TECHNOLOGIE, 0.97)
- Muse Spark 1.2 —est_variante_de→ Muse Spark 1.1 (TECHNOLOGIE, 0.95)
- Muse Code —utilise→ Muse Spark 1.2 (TECHNOLOGIE, 0.96)
- Muse Spark 1.2 —est_basé_sur→ un co-entraînement avec le harnais Muse Code pour maximiser leur compatibilité (AFFIRMATION, 0.95)
- co-entraînement d'un modèle avec son harnais —réduit→ l'interchangeabilité des poids sous-jacents (AFFIRMATION, 0.85)
- journal d'événements local —permet→ de reprendre exactement où l'agent s'est arrêté après un plantage (AFFIRMATION, 0.95)
- agents d'arrière-plan persistants —s_oppose_à→ les sous-agents créés puis détruits pour chaque tâche (AFFIRMATION, 0.94)
- agents d'arrière-plan persistants —réduit→ la collecte d'information redondante et le besoin de pilotage (AFFIRMATION, 0.93)
- Muse Code —utilise→ une skill de mise à l'épreuve du plan avant implémentation (AFFIRMATION, 0.93)
- Muse Spark 1.1 —permet→ de générer les environnements et de noter les solutions ayant servi à entraîner Muse Spark 1.2 (AFFIRMATION, 0.9)
- Opus 5 —surpasse→ Muse Spark 1.2 (TECHNOLOGIE, 0.95)
- Muse Spark 1.2 —mesure→ 82,9 % sur Terminal-Bench 2.1 contre 86,7 % pour Opus 5 avec Claude Code (MESURE, 0.93)
- Muse Spark 1.2 —mesure→ 59,3 % sur DeepSWE 1.1 contre 65,0 % pour Opus 5 et 64,8 % pour GPT 5.6 Terra (MESURE, 0.93)
- Muse Spark 1.2 —mesure→ 70,6 % sur le benchmark interne de Meta contre 79,4 % pour Opus 5 (MESURE, 0.92)
- Muse Spark 1.2 —mesure→ une accélération de noyau KDA de +68,7 %, quatrième derrière Opus 5, GPT 5.6 Sol et Opus 4.8 (MESURE, 0.92)
- comparaison de paires modèle et harnais —s_oppose_à→ une comparaison de modèles isolés, les deux benchmarks publics évaluant chaque modèle avec le harnais de son laboratoire (AFFIRMATION, 0.9)
- gain de Muse Spark 1.1 à 1.2 —mesure→ 2,3 points sur le seul comparatif sans harnais indiqué, contre 6,7 points sur les comparatifs avec harnais (MESURE, 0.88)
- Muse Code —s_applique_à→ l'optimisation itérative de noyaux GPU sur plus de 1 000 appels d'outils et jusqu'à 24 heures (AFFIRMATION, 0.93)
- interdiction d'importer une bibliothèque de noyaux tierce —permet→ de tester la réimplémentation d'un algorithme plutôt que l'enveloppement d'une implémentation existante (AFFIRMATION, 0.92)
- Meta AI Research —affirme_que→ ce lancement est une étape vers la frontière, des modèles plus grands et plus capables étant à venir (CITATION, 0.95)
- benchmarks publiés uniquement en image —s_oppose_à→ la lisibilité des résultats par un agrégateur ou un agent lisant la page (AFFIRMATION, 0.88)

---
Canonical: https://www.thekb.eu/en/fiches/meta-muse-code-muse-spark-1-2-2026-08-05/