guide · ai

Muse Spark 1.2

Muse Spark 1.2 (Meta / MSL, August 5, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.

August 12, 2026 · By Alastair Fraser

A retro robot representing Muse Spark 1.2

--- title: “Muse Spark 1.2” description: “Muse Spark 1.2 (Meta / MSL, August 5, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.” type: guide category: ai pubDate: 2026-08-12 image: /images/abs-model-muse-spark-1.2.png imageAlt: “A retro robot representing Muse Spark 1.2” imagePrompt: “Bold graphic editorial illustration, 1990s comic-book influence, heavy ink outlines, halftone texture, crimson and electric blue on cream. A single retro-futurist robot representing an AI model / a brain-in-a-server, no text, no logos, 16:9.” affiliate: false sources: - name: “Meta AI Research — Introducing Muse Code and Muse Spark 1.2” url: “https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2” - name: “Meta Developer model page — Muse Spark” url: “https://developer.meta.com/ai/models/muse-spark/” - name: “Reuters — Meta launches new AI coding tool powered by Muse Spark 1.2 (Aug 5, 2026)” url: “https://www.reuters.com/technology/meta-launches-new-ai-coding-tool-powered-by-muse-spark-12-2026-08-05/” - name: “CNBC — Meta Muse Glimmer open-weight AI model (Aug 10, 2026)” url: “https://www.cnbc.com/2026/08/10/meta-muse-glimmer-open-weight-ai.html” - name: “Artificial Analysis — Muse Spark 1.2 (Aug 5, 2026)” url: “https://artificialanalysis.ai/articles/muse-spark-1-2” - name: “Simon Willison — Muse Code and Muse Spark 1.2 (Aug 5, 2026)” url: “https://simonwillison.net/2026/Aug/5/muse-code-and-muse-spark-12/” - name: “mindshub.ai — Navigating the LLM comparison (updated Aug 12, 2026)” url: “https://mindshub.ai/blog/navigating-the-llm-landscape-a-comparative-analysis-of-leading-large-language-models” facts: - label: “Vendor” value: “Meta / MSL” - label: “Released” value: “August 5, 2026” - label: “License” value: “Closed (weights ‘soon’)” - label: “Type” value: “AI model release” related: - abs-model-claude-opus-5 - abs-model-gpt-5.6-cyber - abs-model-grok-4.6 - abs-model-kimi-k3 - abs-model-lfm2.5-vl-3b - abs-model-ling-3.0-flash - abs-model-muse-glimmer-30b - abs-model-nemotron-3.5-lightning - abs-model-qwen-image-3.0 - abs-model-qwen3.7-flash - abs-model-qwen3.8-2.4t-a95b - abs-model-qwen3.8-max-weights-update - abs-model-solar-pro-4 tags: - muse-spark-1.2 - ai-models - model-release draft: false --- --- title: “Muse Spark 1.2” description: “Muse Spark 1.2 (Meta / MSL, August 5, 2026): what the release is, why it matters for operators, specs, benchmarks, and the call.” type: guide category: ai pubDate: 2026-08-12 image: /images/abs-model-muse-spark-1.2.png imageAlt: “A retro robot representing Muse Spark 1.2” imagePrompt: “Bold graphic editorial illustration, 1990s comic-book influence, heavy ink outlines, halftone texture, crimson and electric blue on cream. A single retro-futurist robot representing an AI model / a brain-in-a-server, no text, no logos, 16:9.” affiliate: false sources: - name: “Meta AI Research — Introducing Muse Code and Muse Spark 1.2” url: “https://research.meta.ai/blog/introducing-muse-code-and-muse-spark-1-2” - name: “Meta Developer model page — Muse Spark” url: “https://developer.meta.com/ai/models/muse-spark/” - name: “Reuters — Meta launches new AI coding tool powered by Muse Spark 1.2 (Aug 5, 2026)” url: “https://www.reuters.com/technology/meta-launches-new-ai-coding-tool-powered-by-muse-spark-12-2026-08-05/” - name: “CNBC — Meta Muse Glimmer open-weight AI model (Aug 10, 2026)” url: “https://www.cnbc.com/2026/08/10/meta-muse-glimmer-open-weight-ai.html” - name: “Artificial Analysis — Muse Spark 1.2 (Aug 5, 2026)” url: “https://artificialanalysis.ai/articles/muse-spark-1-2” - name: “Simon Willison — Muse Code and Muse Spark 1.2 (Aug 5, 2026)” url: “https://simonwillison.net/2026/Aug/5/muse-code-and-muse-spark-12/” - name: “mindshub.ai — Navigating the LLM comparison (updated Aug 12, 2026)” url: “https://mindshub.ai/blog/navigating-the-llm-landscape-a-comparative-analysis-of-leading-large-language-models” facts: - label: “Vendor” value: “Meta / MSL” - label: “Released” value: “August 5, 2026” - label: “License” value: “Closed (weights ‘soon’)” - label: “Type” value: “AI model release” related: - abs-model-claude-opus-5 - abs-model-gpt-5.6-cyber - abs-model-grok-4.6 - abs-model-kimi-k3 - abs-model-lfm2.5-vl-3b - abs-model-ling-3.0-flash - abs-model-muse-glimmer-30b - abs-model-nemotron-3.5-lightning - abs-model-qwen-image-3.0 - abs-model-qwen3.7-flash - abs-model-qwen3.8-2.4t-a95b - abs-model-qwen3.8-max-weights-update - abs-model-solar-pro-4 tags: - muse-spark-1.2 - ai-models - model-release draft: false --- --- type: guide category: ai --- # Muse Spark 1.2 — Meta’s frontier-at-a-discount coding model, closed today and open-weight soon ## The release Meta Superintelligence Labs (MSL) shipped Muse Spark 1.2 on August 5, 2026 — a coding-focused update to the proprietary Muse Spark line that MSL launched in April to end the Llama era. The launch was paired with Muse Code (beta), Meta’s first-party terminal coding agent, and the two were co-trained together: > “Aug 5 (Reuters) - Meta Platforms on Wednesday launched Muse Code, a coding tool powered by its latest AI model, Muse Spark 1.2, designed to help developers write and debug software.” — Reuters, Aug 5 2026, 9:29 PM UTC Four days later, on August 10, Mark Zuckerberg announced on Instagram and in a 6,500-word essay that Meta would open the Spark 1.2 weights “soon” — following the August 9 Glimmer release. CNBC’s headline that day: “Meta launched Muse Code and plans to release Muse Spark 1.2 weights.” As of Aug 12, 2026, the model is still closed in production. Treat open-weights as a future option. ## What it is and where it sits in the Muse family Muse Spark 1.2 is a transformer-based, multimodal-reasoning LLM (large language model) with a 1M-token context window — the same architecture line as Spark 1.1 (released July 9, 2026), with training compute concentrated on coding tasks and explicit co-training with the Muse Code harness. Per Reuters, the launch came nearly a month after Meta introduced 1.1, and 1.1 generated difficult coding challenges that improved 1.2’s ability to follow complex instructions. 1.2 retains the rest of 1.1’s agentic surface; the work is concentrated on coding, not replacing 1.1 wholesale. Family timeline, all 2026: | Model | Release | Role | |---|---|---| | Muse Spark (original) | April 2026 | First MSL frontier model; proprietary, ended the Llama era | | Muse Spark 1.1 | July 9, 2026 | Multimodal reasoning upgrade; Meta Model API public preview | | Muse Spark 1.2 | August 5, 2026 | Coding-focused update to Spark 1.1 (closed today; weights “soon” per Zuck) | | Muse Code | August 5, 2026 | Terminal coding agent harness, co-trained with Spark 1.2 | | Muse Glimmer-30B | August 9, 2026 | First open-weight MSL model; distilled from Muse Spark | The reading: Spark = closed frontier (cloud) · Code = closed coding agent (terminal) · Glimmer = open-weights local agent. 1.2 sits between 1.1 and a future model Meta teased. ## How it works — the co-trained model + harness system The model and harness were co-trained. Meta’s blog post is direct: > “We co-trained Muse Spark 1.2 with Muse Code to ensure the model exhibits its best performance and coding usability when paired together. The training included rejection sampled harness trajectories and recipe optimizations for goals, compaction, and subagents, alongside the integration of the Muse Code toolset to maximize harness compatibility.” — Meta AI Research blog, Aug 5 2026 Three mechanics flow from that. The harness leaves traces the model was trained on — Muse Code keeps an event log of every tool run, and the model was trained on rejection-sampled trajectories from that log. Bundled skills ship with the training distribution: /plan decomposes a task, /grill stress-tests it, /goal enforces compaction. The model runs persistent async sub-agents and resumes from a crash mid-task — the long-horizon bet for 1,000+ tool calls over 24 hours. ### Specs that matter | Spec | Value | |---|---| | Context window | 1,000,000 tokens | | Modalities | Text + image input, text output | | Pricing (standard) | $1.25 input / $4.25 output per MTok | | Pricing (contributor tier) | $0.10 / $0.20 per MTok — opt in to let Meta train on your traffic | | License | Closed today; weights promised “soon” | The contributor tier is the wild card. Simon Willison flagged it on launch day: > “muse-spark-1.2 is priced at $1.25/million input and $4.25/million output — close to Gemini 3.6 Flash ($1.50/$7.50) — but if you agree to let Meta use your data ‘to improve our products’ you can use muse-spark-1.2-contributor which is $0.10/$0.20 — a huge discount, closer to GPT-5.6 Luna ($0.20/$1.20) and Gemini 3.1 Flash-Lite ($0.25/$1.50).” — Simon Willison, Aug 5 2026 On the contributor tier, output drops to roughly 250× cheaper than Claude Fable 5 per WorldofAI — a real cost lever, but only on traffic you accept Meta training on. ## Where it works — benchmarks and the day-one operator case The headline number is the AA Intelligence Index: 54 (xhigh) — up from 1.1’s 51. > “Meta’s Muse Spark 1.2 scores 54 on the Artificial Analysis Intelligence Index. Its Meta’s third release in four months, significantly improving agentic knowledge work capabilities over prior releases and putting Meta next to SpaceXAI in a tie for third place amongst US labs.” — Artificial Analysis, Aug 5 2026 Day-one positioning vs peers (per AA): Spark 1.2 (54) ties GPT-5.5 (55) and Grok 4.5 (54), behind Claude Opus 5 (61), Claude Fable 5 (60), GPT-5.6 Sol (59), and Kimi K3 (57). Deltas vs 1.1: Intelligence Index +3, GDPval-AA v2 Elo +260 (1631, #5), Terminal-Bench +2 (80%), hallucination rate −10 points — but attempt rate down 15. On cost, AA’s read is the operator-relevant one: > “Among the most cost-efficient models at its intelligence level: Muse Spark 1.2 costs $0.40 per Intelligence Index task at Meta’s unchanged $1.25/$4.25 per 1M token pricing, with only Grok 4.5 (high, $0.37) and GPT-5.6 Sol (medium, $0.39) cheaper in its intelligence cluster.” — Artificial Analysis, Aug 5 2026 mindshub.ai’s updated Aug 12 comparison puts Spark 1.2 at 11–14% off the public intelligence lead — ahead of GLM-5.2’s ~16%, where 1.1 sat tied a month earlier. Coding gains, cost efficiency, and 1M-token context with compaction make 1.2 the cleanest “strong but cheap” coding-agent candidate in the closed-API tier. ## Where it breaks — regressions and “still closed today” Six places Spark 1.2 underperforms: