Meta Model Releases, Tracked

Meta Muse Spark 1.3 built DUNGEONFALL, a 3D roguelike dungeon crawler in a single 736-line HTML file on its first attempt. Featuring procedural non-overlapping multi-room carving, distinct Boss and Treasure vaults, clean Three.js geometry lifecycle cleanup for true permadeath resets, and Web Audio sound effects.

Meta's Muse Spark 1.3 delivered a flawless one-shot pass on our destructible voxel benchmark: 658 lines of Three.js featuring 59,300+ blocks, face culling, 16-chunk partitioned meshing, procedural pixel textures, and locked 60 FPS editing performance.

Meta's Muse Spark 1.3 under xhigh effort generated a 736-line 3D top-down arena shooter with true A* grid pathfinding, raycast line-of-sight cover logic, and 4 differentiated bot personalities. After a round-one firing trigger bug, one follow-up prompt dialed in multi-tier difficulty and lethal combat AI.

Meta's flagship Muse Spark 1.3 under xhigh effort generated a complete 715-line 3D circuit racer with 2 AI opponents, 12 anti-cheat checkpoints, and procedural Web Audio engine sounds. After a missing road surface bug on pass one, a single follow-up prompt fixed the track geometry.

Meta's flagship agentic model Muse Spark 1.3 generated a complete 793-line 3D open-world wilderness survival game in a single HTML file on its first attempt — featuring custom value noise terrain, 5 distinct biomes, instanced foliage at 60 FPS, real-time day/night lighting, and interactive ghost preview crafting.

Meta Superintelligence Labs' Muse Spark 1.3 claims top coding marks with 75.4% on DeepSWE 1.1 and 182 tok/s throughput, backed by an aggressive $0.10 contributor data-sharing pricing tier.

Meta Superintelligence Labs has launched Muse Spark 1.3, an agentic foundation model featuring a 1M-token context window, 75.4% on DeepSWE 1.1, and an aggressive $0.10 contributor pricing tier.

Meta released Muse Spark 1.2 and Muse Code on August 5, 2026 — a 1M-context reasoning model paired with a persistent terminal coding agent that manages parallel sub-agents in isolated git worktrees. A unique Contributor pricing tier at $0.10/$0.20 per million tokens offers a 90% discount in exchange for opting in to training data use.

Meta Superintelligence Labs has released Muse Glimmer, a 30-billion parameter open-weight model distilled from its proprietary Muse Spark flagship. Published under an Apache 2.0 license, Glimmer is purpose-built for offline, on-device agentic workloads like coding, debugging, and file management on consumer hardware.