Muse Spark: Meta Unleashes NEW AI Model (Are they back?)

Meta's Muse Spark lands on the leaderboard, beats Opus 4.6 on cherry-picked benchmarks, and gets shoved into WhatsApp, Instagram, and Ray-Ban glasses — but admits it still loses on coding.

What Meta Claims

Muse Spark scored 42.8 on Humanity's Last Exam, putting it above Opus 4.6 / "Opus Max" and just below Gemini 3.1 Pro and GPT 5.4. In their own blog post, Meta openly admits the model has "performance gaps, such as agentic systems, encoding workflows" — a rare concession compared to Chinese labs that claim number-one status on release. The benchmarks shown are "very cherry-picked," with comparisons placed right next to Opus 4.6 to look competitive.

Coding Reality Check

For agentic and coding workflows, Opus 4.6 "is still the best at that. That is without a doubt." The creators point out that benchmark questions leaking into training data lets models become elite test-takers without real intelligence gains. The Artificial Analysis leaderboard now shows Meta "on the map" for general intelligence, but the coding column is not where Meta wants to be compared.

Distribution Is The Actual Play

Meta's strategy is shove-not-convince. Expect Muse Spark inside WhatsApp, Instagram, and Ray-Ban AI glasses, mirroring how Elon shoved Grok into X. Convenience wins: users already inside Instagram are unlikely to open a separate Opus or ChatGPT tab. The endgame is the same flywheel that made Claude Code smart — millions of users feeding interaction data back into training the next model.

Competitive Position

The creator's call: Muse Spark is a catch-up play, not an outcompete play. To actually close the gap, Meta needs to attract developers the way Anthropic did with Opus, because stolen developer "patterns and thinking patterns" are what made Mefaust possible. Until then, Meta is "not going to have the balls enough" to ship security-grade tooling like Project Glasswing.

Near-Term Watch

Anthropic's Mefaust is "miles ahead" and is rumored for release soon. If self-improving models consolidate into a single vendor, the race to AGI becomes a monopoly problem — a point the creator flags as a real concern, not hype.