Muse Spark
Vintage: 2026-09. Primary evidence is @alexandr_wang plus @AIatMeta in 2026-09-07-weekend-x-ai-meta-aira3-kaggle-gold-openai-wiki. Issuer/founder claims. No independent board snapshot in this pass. Distinct from muse-voice-transcribe (Sep 1 ASR / diarization).
Muse Spark
One-line summary: Meta Superintelligence Labs coding / agentic model line — 1.3 max public release claimed 4 Sep 2026; 1.2 called gold-level post-hoc on the AIRA₃ Kaggle private set.
What it is
A Meta coding/agentic model family shipped via Muse Code and Meta Model API, per Wang and @AIatMeta. Not the Sep 1 audio model muse-voice-transcribe.
Why it matters to this thread
Frontier coding-model releases and their use inside autonomous-research ensembles are in-scope. This page records the 1.3 max public-release claim and the post-hoc AIRA₃ medal claims. Do not invent benches or prices.
Key facts (from 2026-09-07-weekend-x-ai-meta-aira3-kaggle-gold-openai-wiki)
- From the same source (@alexandr_wang, 2026-09-04): Muse Spark 1.3 max public release on Muse Code and Meta Model API; Wang claiming stronger coding/agentic performance vs 1.3 high/xhigh. No independent board snapshot in this pass.
- From the same source (@AIatMeta, 2026-09-04): Meta pointed at that Wang post.
- From the same source (AIRA₃ text follow-up, 2026-09-05): post-hoc on the AIRA₃ private set — Muse Spark 1.2 (MuseCode) gold-level; Muse Spark 1.1 (OpenCode) silver-level. Live gold entry was GPT 5.5 (OpenCode) + Claude 4.8 (ClaudeCode), not 1.3 max. See aira3.
What this source does not establish
- No independent leaderboard snapshot for 1.3 max vs 1.3 high/xhigh.
- Not muse-voice-transcribe. Different product, different date.
- 1.3 max was not the live gold ensemble. Live gold was GPT 5.5 + Claude 4.8. 1.2 gold-level is post-hoc.
- No API id, price, or parameter count in this pass.
- No ticker or 8-K.
Open questions
- How does 1.3 max compare on a named public board versus 1.3 high/xhigh or the AIRA₃ ensemble models?