Meta's MuseSpark1.3 ties Fable5 on ArtificialAnalysis while cutting token use 25%
量子位 · wechat · 2026-09-03
Meta released MuseSpark1.3 on September 2, the fourth version of the line since April. On the ArtificialAnalysis leaderboard, MuseSpark1.3 Max scored 62, tying Fable5 and 5.1 xhigh, and now beats the higher-ranked Opus5 max on Coding and Agent sub-scores.
- Efficiency: 20% fewer tool calls and 25% fewer tokens than 1.2, with better retention of initial instructions on long-horizon tasks.
- Community tests: same-prompt 3D Viking figurine HTML generation cost $0.10 with visibly better detail; in another test, the Ultra Contributor tier ran 20 self-improvement rounds and 60+ agent runs under a "don't stop below 9.5/10" judge rule, all under $1.
- Unchanged pricing: $1.25/M input, $4.25/M output, $0.15 cached input. ArtificialAnalysis estimates $0.55 per Intelligence Index task for 1.3 xhigh, well under GPT-5.6 Sol Max ($0.95) and Grok4.6 High ($0.94).
- Team pays off: core researchers Shengjia Zhao and Shuchao Bi detailed a training-stage investment in reward-scoring correction targeting laziness, reward hacking, and evasive answers.
Open weights are coming, and a larger Muse model is still teased.
More from Models
- Loop Transformer's Real Winner Is SRAM-Based Ultrafast Inference, Argues Bing Xu — bingxu_ · 2026-09-03
- Meta's Muse Spark 1.3 is now free on OpenCode — ramagetime · 2026-09-03
- DeepLoop paper makes looped transformers scalable; rumor claims frontier models are 48 layers looped twice — StartupYou · 2026-09-03
- Leaked: OpenAI's Secret 'Bel' Model Slated for Year-End After Astra's Agent-1 Stage — haider1 · 2026-09-03
- Reddit user on Gemini 3.8 Flash: more effort, similar result — Correct_Tomato1871 · 2026-09-03
- Qwen3.8-Flash-Next on 2x3090 + DDR4: expert cache PR lifts decode from 17 to 25-29 t/s — Extension-Bid-639 · 2026-09-03