Muse Spark 1.3: Meta Is Back at the Top, and the Best Open-Weight Model Could Follow

Mark Zuckerberg pushed Muse Spark 1.3 out on X, and this time the brag comes with numbers that hold up. He calls it Meta’s biggest jump yet on coding and agentic work, and the model is already in Muse Code and Meta’s API. The post also teased what comes next: another model marked only by a watermelon emoji, plus open-weight releases for Muse Spark.

The clearest sign of progress is the Artificial Analysis Intelligence Index. Muse Spark 1.3 (xhigh) lands at 61, up from 57 on version 1.2 and 53 on version 1.1. That puts it beside GPT-5.6 Sol (max), Grok 4.6 (high) and Claude Opus 5 (high). The stronger max variant, which is still in limited preview for Meta partners, reaches 62. Only two models sit above it overall, both from Anthropic: Claude Fable 5.1 (max) at 66 and Claude Opus 5 (max) at 63.

The gains are concentrated in the stuff Meta cares about most. On Tau3-Bench Banking, the xhigh model rises from 35 to 47 percent and the max version to 52 percent, which is the best score of any model tested. Terminal-Bench 2.1 moves from 80 to 85 percent and 86 percent. GDPval-AA v2 climbs from 1,615 to 1,709 and 1,754 in Elo terms. Scientific reasoning improves too, with CritPt up from 18 to 26 percent and GPQA Diamond from 90 to 94 percent. Humanity’s Last Exam and SciCode also tick up by two to three points.

That stronger agentic performance costs more compute. Meta says the max variant reasons 62 percent longer on GDPval-AA v2 and 28 percent longer on Tau3-Bench Banking than xhigh. There are two slips: AA-LCR drops from 83 to 79 percent, and AA-Omniscience falls by three points. Artificial Analysis says that last drop is tied to the model refusing to answer more often when it is unsure, which also lowers hallucinations.

Price is another part of the pitch. Muse Spark 1.3 (xhigh) costs $0.55 per Intelligence Index task, cheaper than every model scoring 59 or above. Meta’s API pricing stays at $1.25 per million input tokens and $4.25 per million output tokens, with cache hits at $0.15. The model takes text, image and video, and it has a one million token context window.

The open-weight piece is where this gets messy in the useful way. Chinese labs currently lead that ranking, with Kimi K3 (max) and GLM-5.3 (max) both at 60, and Alibaba’s Qwen3.8 at 58. A Muse Spark 1.3 open release would jump straight to the front. But Meta has only committed to releasing the weights of Muse Spark 1.2 so far, and whether version 1.3 follows is still undecided.