AI News Feed
Market watch
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
LG UltraFine Evo 32U990A-S Hits Record-Low $1,099 on Amazon, TechRadar ReportsNewegg Labor Day Deal: 32GB Kingston Fury Beast DDR4-3200 Kit Nearly Half Price at $179.99 with CodeGalaxy Z Fold 8 inner screen corners reported with excessive giveHuawei unveils Mate XT 2 tri-fold with privacy-screen edition as Apple prepares folding iPhoneWall Street Analysts Recommend Nvidia, Uber and Marvell for Long-Term GrowthLenovo, AMD and Huawei Put Local Large-Model AI at Center of New Devices
Companies

Meta Releases Muse Spark 1.3, an Agentic Coding Model Using 20% Fewer Tool Calls and 25% Fewer Tokens

Meta launched Muse Spark 1.3, an agentic coding model with 20% fewer tool calls and 25% fewer tokens.

AI News Feed · 2026-09-04 14:20

According to the report, Muse Spark 1.3 is available today in Muse Code and the Meta Model API for production use. It cannot be self-hosted because its weights are closed, and a max reasoning mode remains gated behind further safety testing. The model has a 1M-token context window.

The report said Meta trained Muse Spark 1.3 across multiple agent harnesses so its behavior generalizes beyond a single environment. Given an open-ended objective, the model gathers its own context from messy and conflicting sources and patches gaps in its plan. It asks clarifying questions on ambiguous prompts, pulls the user in when it stalls, and confirms before consequential actions. On long runs it adapts to preferences for frequent status updates or silent execution. Meta also reported better calibration on the model's own limits, so it flags hurdles instead of hallucinating an outcome. The report said the model maps an incoming prompt to the correct task inside a cluttered single thread, whether the user is steering or interrupting.

For coding, the report said Meta trained Muse Spark 1.3 on more long-horizon coding tasks. Meta engineers compared it with Muse Spark 1.2 and found approximately 20 percent fewer tool calls and 25 percent fewer tokens, as well as fewer unnecessary turns, less verbosity, and cleaner code style.

On Meta's own benchmark scores, Muse Spark 1.3 reached 75.4 on DeepSWE v1.1, ahead of Claude Opus 5 at 74.0 and GPT-5.6 Sol at 72.7. It scored 59.4 on SWE-Atlas Codebase QnA and tied GPT-5.6 Sol at 88.8 on Terminal-Bench 2.1, with Claude Opus 5 at 86.7. Long-context retrieval showed the widest gap: MRCR v2 scores of 98.5 at 256K–512K and 98.1 at 512K–1M, against 91.5 and 73.8 for GPT-5.6 Sol. The report noted that Meta's scorecard uses the max mode, which is not the mode developers can call today. Meta reported OSWorld 2.0 at 66.9 for max versus 57.2 for xhigh, GDPval-AA v2 Elo at 1,754 versus 1,709, and JobBench at 64.9 versus 61.2; DeepSearchQA tied at 89.4 for both. Since Muse Spark 1.2 was evaluated at xhigh, part of the generational jump is a reasoning-tier change, the report said.

Artificial Analysis scores the shipping xhigh variant at 61 on its Intelligence Index and the preview max variant at 62, according to the report. That places xhigh level with GPT-5.6 Sol max and Grok 4.6 high, behind Claude Opus 5 max at 63 and Claude Fable 5.1 max at 66. A separate Tau3-Bench Banking measurement came in at 47 percent for xhigh and 52 percent for max, the top score the platform has recorded on that evaluation. The report said pricing is unchanged at $1.25 per million input tokens and $4.25 per million output tokens, with a contributor tier at $0.10 and $0.20. Meta lists an open-weights Muse Spark release on its roadmap, but current weights remain closed.