Muse Spark 1.2

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

Meta released Muse Spark 1.2 on August 5, 2026, its third model in four months, alongside Muse Code (beta), a terminal-based coding agent targeting full software engineering workflows of planning, writing, and validating code rather than just autocomplete. Zuckerberg announced that Muse Spark 1.2 weights will be releas The article compiles multiple benchmark comparisons: DeepSWE 53% (trailing GPT-5.6 Sol at 73% and Claude Opus 5 at 74%), WeirdML 60.3% (comparable to GPT-5 or GPT-5.4 Mini), GDPval-AA v2 Elo 1631, and Text Arena rating 1498 versus Claude Fable 5's 1507 at roughly 91% lower cost. The model abstains more aggressively tha

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

Yahoo Finance's August 5, 2026 report frames Muse Spark 1.2 as a coding-centric update to the Muse Spark model Meta rolled out the prior month, and identifies Muse Code as the first coding agent from Meta Superintelligence Labs (MSL), the organization rebuilt around the $14.3 billion Scale AI acquihire of Alexandr Wang The launch sits inside Meta's broader model cadence, which has progressed through Muse Spark, Spark 1.1, Spark 1.2, alongside the Muse Image image generation model and a previewed Muse Video model, with capital expenditure raised to a $134–145 billion range (up from $124–145B) to fund data-center buildout. Meta's stock

The Hacker News discussion of Meta's Muse Spark 1.2 announcement surfaced concrete API pricing details not emphasized in the press coverage: developers who opt in to letting Meta train on their data receive a roughly 10x discount on input ($0.10 per million tokens vs. $1.25/Mtok) and a 20x discount on output ($0.20/Mto Beyond pricing, the 333-point, 265-comment thread (4 days old at time of crawl) confirms strong developer interest in Muse Code's terminal-agent workflow and surfaces practical concerns about the data-for-discount tradeoff, with token-budget caveats (e.g., 1M–10M daily limits on comparable OpenAI free tiers, implying o

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 launched Muse Code in beta on August 5, 2026 alongside Muse Spark 1.2, a coding-focused update to the Muse Spark frontier-model family, marking the company's most direct entry into the agentic-coding category dominated by Anthropic's Claude Code and OpenAI's Codex. CEO Mark Zuckerberg announced the terminal-based The harness's headline architectural choice is what Meta calls "persistent async background agents" — a set of specialized agents that remain alive across an entire session rather than being spawned fresh per task as in most rival coding harnesses, alongside parallel worktrees for concurrent work. In an X reply, Zucker

Meta AI released Muse Code in beta on August 5, 2026, a terminal coding agent powered by the new Muse Spark 1.2 model from Meta Superintelligence Labs. Muse Code targets complex software engineering across large repositories by planning changes, writing code, and validating results, with async background agents that pe The agent uses a local append-only event log recording every model call, tool run, and approval, and is positioned by Meta as its next step toward the frontier with larger models on the way. This launch-day coverage provides developer-relevant architectural detail for evaluating Muse Code as a competitor to existing te

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