Two American Open-Weight Models Landed in 24 Hours, and They Are Fighting Over Tokens and Licences
Reflection AI released Beam under Apache 2.0 while Meta shipped Muse Spark Open under a tighter licence, giving developers two Western rivals to China's open models with different trade-offs.
Outspoken Digest Technology Desk
Sunday, October 18, 2026/2 min read
For a year the best open-weight AI models have mostly come from China. This week, two American labs answered within a day of each other, and the more interesting part is how differently they did it.
Reflection AI released the weights of Beam on 16 October under the permissive Apache 2.0 licence. Meta shipped Muse Spark Open the same week, with a licence that comes with strings. Developers who want to run a frontier-grade model on their own hardware now have a real choice.
What is Reflection's Beam?
Beam is a mixture-of-experts model with 501 billion parameters, of which 23 billion are active for each token, VentureBeat reported. Reflection says it was trained from new weights, not on other models' outputs. Eight-bit and four-bit versions fit on a single eight-GPU server using H100, H200 or GB300 chips.
The pitch is efficiency. Reflection reports about 80 percent on the coding benchmark Terminal-Bench 2.1 while using roughly 25,000 tokens per attempt, against about 80,000 for the Chinese model GLM-5.2, which scores slightly higher. The company says that works out to three to four times less inference compute at comparable scores, an estimate based on active parameters. Artificial Analysis said early indicators suggest Beam is highly token-efficient, but those are first impressions, and the benchmark figures are Reflection's own.
VentureBeat notes Reflection has not set a price and its hosted service is free during public beta. The article also says the Financial Times reported on 10 October that Nvidia, a close partner and investor, was in early talks to buy Reflection or invest more. Co-founder Ioannis Antonoglou called that "baseless rumors" without saying whether talks exist.
What is Meta's Muse Spark Open?
Meta's model is larger: 816 billion parameters, 42 billion active, a one-million-token context window and native multimodality, according to OfficeChai. Alexandr Wang called it the most capable US open-weight model to date. By Meta's own figures it leads open rivals including GLM 5.3 and Kimi K3 on several benchmarks but trails both on Terminal-Bench 2.1.
The weights are on Hugging Face, Amazon SageMaker AI, Google Cloud Model Garden and Microsoft Foundry Models. Meta's demo ran the model on two linked Apple M5 Ultra Mac Studios, giving about a terabyte of memory.
Which one can you actually use?
The licences matter as much as the benchmarks. VentureBeat says Meta's Muse Model License requires the line "Built with Muse Spark Open" to be displayed, and companies with more than $20 million in revenue that sell access as a service need a separate agreement. Beam's Apache 2.0 carries no such conditions.
Neither has been independently tested at length. The real contest will be fought in cost per finished task, which is what VentureBeat says it is now measuring.
Published in The Outspoken Digest
Editorial desk
Outspoken Digest Technology DeskSoftware, hardware, artificial intelligence and what they change for everyone else.
Newsletter
The Digest, in your inbox
One edition, sent when it is ready. No noise, and your address is never passed on.
Read Next
More Technology →
Webb Caught a Probable Supernova From the Early Universe Only Days After It Exploded
Oct 18, 2026/2 min read
Webb Maps an Exoplanet's Atmosphere as Egg-Shaped, With a Puffed-Up Evening Side Hotter Than Its Morning Edge
Oct 18, 2026/2 min read
A Hole in a 75,000-Year-Old Shoulder Blade Points to a Wooden Spear, and Perhaps to Neanderthal Care
Oct 18, 2026/2 min read
Anthropic's Own Staff Data: Junior Technical Hires Fell From 27 to 9 Percent as Its AI Agents Outworked Humans
Oct 18, 2026/2 min read