# Meta releases Muse Spark Open, an 816-billion-parameter open-weight model that still trails China's Kimi K3 on coding

The mixture-of-experts model activates 42 billion of its parameters per token, reads up to 1 million tokens of context and runs locally on two Apple M5 Ultra Mac Studios, though its license limits commercial use.

- Published: 2026-10-16T06:26:08.965Z
- Canonical: https://polylog.news/ai/2026-10-16/meta-releases-muse-spark-open-an-816-billion-parameter-open
- Publisher: Polylog (AI desk)
- Section: tech
- Sources: [Meta AI Research (via Hacker News)](https://research.meta.ai/blog/muse-spark-open), [Meta AI](https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/), [Polylog editors](https://polylog.news)

Meta Superintelligence Labs put the weights of [Muse Spark Open](https://research.meta.ai/blog/muse-spark-open) on Hugging Face on October 15. It is the first open release from the Muse family. Until now, Meta had kept Muse closed and offered it only through its [own API and apps](https://ai.meta.com/blog/introducing-muse-spark-meta-model-api/). The [model card](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) lists 816 billion total parameters, with 42 billion active per token. [Reporting on the release](https://pasqualepillitteri.it/en/news/24330/meta-muse-spark-open-816b-open-weights) describes a mixture-of-experts Transformer with 256 experts, 8 of them routed per token. The model has a 1-million-token context window, accepts text, image, video and audio, and outputs text. It shares a base model with the closed Muse Spark 1.3 and received extra safety post-training so it can run without Meta's system-level guardrails.

Meta ships two weight formats. One is BF16. The other is NVFP4, NVIDIA's 4-bit floating-point format, which Blackwell hardware supports natively. Serving NVFP4 takes at least about 700 GB of accelerator memory, and BF16 takes about 2 TB. The most notable deployment claim is a desktop setup: [two M5 Ultra Mac Studios](https://venturebeat.com/technology/meta-unveils-muse-spark-open-the-new-top-u-s-open-weights-model-and-it-runs-on-two-paired-mac-desktops), each with 512 GB of unified memory. VentureBeat notes that this memory configuration has not shipped publicly yet. [CNBC](https://www.cnbc.com/2026/10/15/apple-meta-partner-ai.html) described the work as a joint Apple and Meta effort to run the model on Macs. A Russian-language AI channel called it [one of the strongest open models in the United States](https://t.me/ai_machinelearning_big_data/11138) and noted that it runs without cloud servers.

The benchmark results are less favorable than the launch presentation suggests. A [comparison of Meta's published numbers](https://www.trendingtopics.eu/muse-spark-open-meta-open-weight-license/) puts Muse Spark Open at 76.4 on Terminal-Bench 2.1, against 85.0 for Moonshot AI's Kimi K3. On SWE-Atlas-QnA it scores 58 to Kimi K3's 66. It also trails Qwen, GLM and Kimi configurations on Terminal-Bench 2.1, and on DeepSWE 1.1 it beats Qwen but trails GLM and Kimi. It does beat the other US open models, NVIDIA's Nemotron 3 Ultra and Thinking Machines' Inkling, often by more than 20 points. Meta's model card claims a score of 39 on the Artificial Analysis Intelligence Index. VentureBeat [could not verify](https://venturebeat.com/technology/meta-unveils-muse-spark-open-the-new-top-u-s-open-weights-model-and-it-runs-on-two-paired-mac-desktops) any independent Artificial Analysis assessment at publication. If the score is confirmed, it would tie DeepSeek V4.1 Flash for fifth place.

The license is the main practical constraint. The Muse Model License [closely resembles Kimi K3's license](https://www.trendingtopics.eu/muse-spark-open-meta-open-weight-license/), with commercial limits and naming requirements. Alibaba's Qwen and Zhipu's GLM have mostly shipped under Apache 2.0 or MIT, which have no revenue thresholds.

## What this means

The model's real value is where it comes from, not its benchmark rank. US government agencies and regulated enterprises that will not deploy Chinese weights now have a US-origin open model with 1 million tokens of context and native multimodality. That cuts into the Chinese labs' advantage in those accounts and undercuts mid-tier closed APIs on price. Apple gains a reason to sell high-memory workstations, and the NVFP4 weights favor NVIDIA Blackwell for cloud serving. The framing that this is the new top US open model is correct but narrow: on Meta's own numbers, it is 8.6 points behind Kimi K3 on Terminal-Bench 2.1, and its license gives back the advantage that Qwen and GLM hold through their Apache and MIT terms. The closest precedent is Llama 2 in July 2023. Its custom license, which excluded services with more than 700 million monthly users, did not stop a large wave of fine-tunes, but it added legal review costs for enterprises, and Chinese labs then won developer share with more permissive licenses. The likelier outcome is that Artificial Analysis publishes an independent score within weeks that lands at or slightly below Meta's 39, and that adoption concentrates in US regulated sectors rather than the global developer base. That read depends on Meta continuing to ship open versions of later Muse bases. It would be proven wrong if the independent score comes in well below 39, or if Meta stops releasing open weights after this one, as it did after Llama 4.

## What to watch

- Whether Artificial Analysis publishes an independent Intelligence Index score near Meta's self-reported 39. A much lower score would show the launch claims were vendor framing.
- Whether Moonshot, Alibaba or Zhipu respond with new releases or price cuts in the coming weeks. That would show the US-China open-weight contest is now being fought release by release.
- Whether US federal agencies or defense contractors name Muse Spark Open in procurement. That would confirm that model origin, not raw benchmark rank, is driving adoption.