Intel / post

Meta's Muse Spark Open tops US open-weight benchmarks

Source: x.com
Date: Published Oct 16, 2026

Context

Meta's Muse Spark Open is a model with 816 billion parameters, of which 42 billion are active for each : a router sends each token to 8 of 256 expert sub-networks rather than through the whole model. On Artificial Analysis' Intelligence Index, a composite of ten evaluations, it scores 39, which the benchmarker says makes it the highest-scoring US open-weights model (Artificial Analysis). That is 22 points above Muse Glimmer, Meta's 30B open-weights model from August, and ahead of NVIDIA's Nemotron 3.5 Super VL at 26. It is also within a point of Muse Spark 1.2 (40), Meta's August flagship. Meta's current Muse Wonder models score higher still, at 52 and 56. Artificial Analysis tested the model with access Meta gave it before the public release. Meta's own announcement is separate Intel.

The more useful finding is where the score comes from. Among the nine highest-scoring open-weights models, Muse Spark Open ranks last on the index's four agentic evaluations but fourth on the other six combined. It leads all open-weights models on AA-Omniscience (25), helped by a 24% rate, the lowest of any open-weights model scoring 30 or more. Artificial Analysis says Z.ai's GLM-5.3, the closest Chinese model by size, owes all of its 6-point lead over Muse Spark Open to the agentic evaluations. In short, its overall standing rests on the index's non-agentic tests, and it is comparatively weak at the multi-step, tool-using tasks the agentic evaluations measure.

It is also economical with output. It uses 40k output tokens per index task, the fewest among those nine models and 43% fewer than GLM-5.3. At the launch price from Together AI, Fireworks and Baseten ($1.25 per million input tokens, $4.25 per million output tokens), that works out to $0.84 per task. That is less than half of GLM-5.3's $2.01 but about three times DeepSeek's V4.1 Flash at $0.27.

The weights can be run outside Meta's cloud. Artificial Analysis reports that the full BF16 weights are about 1.6TB, and that Meta's NVFP4 checkpoint compresses only the 256 experts to 4-bit, bringing the full set of weights to about 530GB. Meta says that version runs on four Blackwell Ultra GPUs, eight H200s, or two 512GB M5 Ultra Mac Studios. The custom Muse Model License requires a "Built with Muse Spark Open" attribution. Model-as-a-Service businesses with more than $20M in annual revenue need a separate agreement with Meta. All scores here come from one benchmarker's thread; Meta's own claims are reported secondhand, and the chart images with the full results were not read.

Terms in this piece · Glossary

mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
token — The chunk of text a model reads and writes in — roughly three-quarters of a word — and the unit AI usage is billed in.
hallucination — When a model states something false with full confidence — inventing facts, citations, or APIs that don't exist.
eval — A repeatable test for AI quality — a set of tasks plus scoring — used the way software teams use test suites, because model output is too variable to judge by eyeballing.

Read the source x.com