Muse Spark 1.2

Meta released Muse Spark 1.2 on August 5, 2026, roughly four weeks after Muse Spark 1.1, shipping it alongside Muse Code — Meta's own terminal-based coding agent in beta, co-trained with the model it runs on. On Artificial Analysis' Intelligence Index, Muse Spark 1.2 scored 54, tying Grok 4.5. The OrcaRouter blog (the Five days after release, on August 10, 2026, Meta announced it would open the weights of Muse Spark 1.2, which would make it the strongest current U.S. open-weight competitor to Chinese open-weight models. The post also surfaces pricing context (via the provider's catalog showing Meta at $1.25 input / $4.25 output per

A primary-source dissection of Muse Code and Muse Spark 1.2 confirms that Meta announced both on August 5, 2026, with Muse Code positioned as a terminal-based coding agent and 1.2 as the model trained to run inside it. The analysis draws on the official Meta AI Research blog and the evaluation methodology report, organ On measurement, the article walks through Meta's actual benchmark figures for 1.2 rather than relying on chart images, and reports a case where GPU kernel optimization required roughly 24 hours of model time, illustrating the practical cost ceiling for long-horizon agentic runs on the standard tier. It also restates th

Eesel AI's analysis confirms Meta shipped Muse Spark 1.2 on August 5, 2026, describing it as "a moderate improvement over Muse Spark 1.1" optimized for multi-file refactors and long debugging sessions. Three elements carried over unchanged from 1.1: the 1,048,576-token context window, the standard rate card, and closed weights. The headline change is expanded public preview availability beyond the United States, which had been the main friction for 1.1 adopters abroad. On Artificial Analysis metrics, Muse Spark 1.2's agentic index jumped from 39.7 to 49.3 (+9.6) and GDPval-AA v2 Elo rose from 1374 to 1628 (+254), while Intelligence Index moved modestly from 53.2 to 56.8. Coding-specific benchmarks like Terminal-Bench, Omniscience, and SciCode stayed flat or declined, making the agentic gains the defining story. Artificial Analysis has already deprecated Muse Spark 1.1 in favor of 1.2, signaling Meta's intended lineage direction.

CNBC reported on August 10, 2026 that Meta CEO Mark Zuckerberg announced in an Instagram video that the company would open the weights of its latest model, Muse Spark 1.2, making the parameters downloadable and usable by the public. Zuckerberg framed the move as a competitive response to both U.S. closed-model labs lik The same announcement introduced Muse Glimmer, a new family of open-source models designed to run on laptops, positioned alongside the open-weight Muse Spark 1.2 release. The article places the decision within investor scrutiny of Meta's Superintelligence Labs capital expenditure plans, which Meta has forecast at up to

Meta released Muse Spark 1.2 on August 5, 2026 alongside the beta of Muse Code, a terminal-based coding agent that Meta co-trained with the model rather than leaving harness development to the ecosystem. The published Meta Model API list rate is $1.25 per million input tokens and $4.25 per million output tokens on the Beyond pricing and packaging, the launch guide notes that Meta published benchmark charts for Muse Spark 1.2 only as images rather than as raw numbers in body text, which limits independent verification of headline scores and is a meaningful caveat for any team evaluating the model. The contributor-tier discount is exp

Meta AI Research announced Muse Spark 1.2 on August 5, 2026, alongside Muse Code (beta), a terminal coding agent powered by the new model. Muse Code handles complex software engineering tasks across large repositories: planning changes, writing code, and validating results, while coordinating multiple persistent subagents to avoid redundant work. The runtime uses a local event log for replay-exact, restart-safe execution, and ships with bundled skills like /plan, /grill, and /goal for structured task progression. The research post positions Muse Spark 1.2 as "our next step toward the frontier, with larger and much more capable models on the way." Muse Code installs via a single curl command on macOS or Linux and maintains async background agents that persist across sessions rather than spawning per task. This combination of a coding-focused model and an integrated terminal agent represents Meta's push into agentic software engineering workflows that span entire codebases.

Artificial Analysis published a benchmark-driven breakdown of Muse Spark 1.2 on August 5, 2026, giving the model an Intelligence Index score of 54 (xhigh) — up three points from Muse Spark 1.1 and eleven points from the April Muse Spark 1.0 release. The piece credits Meta Superintelligence Labs for early access and pla Agentic knowledge-work gains drive the update: GDPval-AA v2 Elo rises 260 points to 1631 (fifth among all models benchmarked and ahead of Claude Opus 4.8), Terminal-Bench v2.1 improves from 78% to 80%, and τ³-Banking edges from 25% to 27%. The AA-Omniscience abstention rate climbs from 18 to 22 as hallucination falls 1

Meta debuted Muse Spark 1.2 on August 5, 2026 as a coding-centric update to its Muse Spark model line, released through Meta Superintelligence Labs (MSL). According to Yahoo Finance, Spark 1.2 ships alongside Muse Code, the first coding agent from MSL, which serves as the harness running the model and positions Meta to The Muse Spark launch forms part of Meta's broader AI rebuild after its Llama models underwhelmed developers last year, an effort driven by chief AI officer Alexandr Wang following Scale AI's $14.3 billion acquihire. The release sequence spans Muse Spark, Spark 1.1, Spark 1.2, the Muse Image image generation model, and

AI Release Tracker's spec page documents Muse Spark 1.2 as released by Meta on August 5, 2026, with a 1 million token context window, priced at $1.25 input and $4.25 output per million tokens (rates fetched from OpenRouter on September 2, 2026). It arrived 27 days after Muse Spark 1.1 and shipped alongside Muse Code, M The tracker records benchmark results as: DeepSWE 1.1 at 59.3% (12th of 24), Terminal-Bench 2.1 at 82.9% (13th of 29, behind Gemini 3.8 Flash at 89.4%), and BullshitBench v2 at 50% (24th of 74, behind Claude Opus 4.8 at 95%). The coding-focused update positioning and third-party-sourced pricing give developers an indep

Vals AI rates Muse Spark 1.2 at 49.29% accuracy (±1.10) on its Vals Index, ranking 31 of 45 models, with its strongest showing being 1st of 76 on Harvey's Legal Agent Benchmark. The model supports a 1M-token context window and 131k max output tokens, accepting text, image, video, and file inputs. Token costs are listed at $1.25 per million input tokens and $4.25 per million output tokens, with a per-test cost of $1.942 on the Vals Index. Across Vals benchmark categories, Muse Spark 1.2 scored notably on MedScribe (90.06%, rank 5/108), Public Benefits Bench (68.47%, rank 9/48), CyberBench v1.1 (69.52%, rank 21/45), and Finance Agent v2 (60.60%, rank 3/76). Weights remain private, and the model was developed for ranking purposes. Average run latency reached 22 minutes 56 seconds per evaluation cycle.