The US–China Open-Weight Race: Muse Spark Open, Beam, Kimi K3 and GLM-5.3

The US Is Back in the Open-Weight Race

Through the summer, the open-weight leaderboard was a Chinese story. Moonshot’s Kimi K3 arrived in July as the largest open-weight model ever released, Z.ai published GLM-5.3’s weights in August, and Xiaomi’s MiMo-V2.6-Pro, released in September, now holds the top open-weight score on Artificial Analysis’s index.

October brought two US replies. Reflection AI announced Beam 1.0 on October 5, a 501B-parameter model with 23B active under Apache 2.0, with weights reported on October 16. Meta followed on October 15 with Muse Spark Open, an 816B model with 42B active, a 1-million-token context and text, image, video and audio input. It runs on SGLang, vLLM and Apple’s MLX, and is available through Amazon SageMaker AI, Google Cloud Model Garden, Microsoft Foundry and seven API providers, including OpenRouter.

Neither release ends the Chinese lead outright. Reflection itself says Kimi K3 stays ahead on raw capability, and Meta’s own benchmark table shows Muse winning on tool use and losing on coding. What changed is that a buyer who has been told to avoid Chinese-built models now has a large, current open model from a US lab to consider.

Ten Open-Weight Models Side by Side

Memory is what the weights need at 4-bit quantization, before working memory for long prompts. A mixture-of-experts model must hold all its parameters in memory, even though it computes with only the active ones for each token.

As of October 19, 2026, from vendor announcements, model cards, Artificial Analysis and OpenRouter. “—” means no confirmed figure. AA Index is Artificial Analysis Intelligence Index v4.3.x, not comparable with v4.1; an unconfirmed report of 39 for Muse Spark Open is omitted, and MiniMax M3.1 Flash’s 42 is a first listing. M3.1 Flash’s parameter count is per MarkTechPost. Memory is roughly 0.55 bytes per parameter at 4-bit (Beam’s figure is our calculation). DeepSeek V4.1 Flash is the exception: its mostly FP8 weights and a separate Engram memory of about 196B parameters put its 4-bit builds at about 400–445 GB, against about 510 GB as released. Muse Spark Open’s model card gives a floor of about 700 GB of GPU memory for NVFP4 and 2 TB for BF16 on data-center GPUs. GLM-5.3’s price is aggregator-reported; DeepSeek’s off-peak rate is $0.15 / $0.60.

Where Each Model Leads

Meta’s launch benchmarks put Muse Spark Open, Kimi K3 and GLM-5.3 on the same tests. These are Meta’s runs, not independent ones. Moonshot’s own Terminal-Bench 2.1 score for Kimi K3 is higher than Meta measured, and Reflection has published its own numbers for Beam.

All figures are vendor-reported. Meta did not include Beam in its runs; Reflection’s numbers come from its own launch, which also reports 80.9 on SWE-bench Verified. MiniMax reports 66.0 on Terminal-Bench 2.1 for MiniMax M3. Scores from different harnesses aren’t strictly comparable.

Read together with the independent index, the field splits by job rather than by country:

1. Muse Spark Open — Tool use. 89.2 on MCP-Atlas in Meta’s runs, five points ahead of GLM-5.3, with text, image, video and audio input.
2. Kimi K3 — Coding. 85.0 on Terminal-Bench 2.1 in Meta’s run and 88.3 per Moonshot, plus first place on LMArena’s Frontend Code Arena (1,679) at launch.
3. GLM-5.3 — Balance. Second on both of Meta’s benchmarks and 45 on Artificial Analysis v4.3.
4. MiMo-V2.6-Pro — Independent score. The top open-weight model on Artificial Analysis v4.3.2 at 46, at about $0.13 per index task, under MIT.
5. DeepSeek V4.1 Flash — Value. Vals.ai ranks it the top open-weight model on its Vals Index (57.86%, 15th of 56 overall), and its API costs $0.30 / $1.20 at peak.
6. MiniMax M3.1 Flash — Price. $0.15 / $0.60 per million tokens and about $0.22 per index task at a first-listed score of 42; open weights are promised for October 23.
7. Beam 1.0 — License. Apache 2.0, the only US model here under a standard permissive license; Reflection reports 80.1 on Terminal-Bench 2.1 with 23B active parameters.

The honest summary: the US entries are competitive, not dominant. Muse Spark Open wins on tool use in Meta’s own runs, and Chinese labs hold the top three confirmed open-weight scores on the one index that tests every model the same way.

Muse Spark Open on Two Mac Studios

Meta calls a Mac Studio running Muse Spark Open the “best way to run a powerful model on your own desktop.” Apple’s MLX research measured decode speed, the rate at which the model writes its answer, on linked M5 Ultra Mac Studios at an 8,192-token context, one request at a time.

The 512 GB Mac Studio is due in late October and has no price yet; the $30,000–$40,000 figure for a pair is a press estimate, not an Apple price. The four-unit route uses the 256 GB model, which already costs over $10,000, so four of them start above $40,000.

At about 25 tokens per second, a 1,000-token answer takes roughly 40 seconds to write, and a 10,000-token output takes close to seven minutes. That is workable for one person. Apple’s figures don’t cover several simultaneous users, the time to read a long prompt, or contexts far beyond 8,192 tokens.

On token cost alone, the numbers are sobering. Running flat out around the clock at 24.99 tokens per second, the pair would write about 790 million tokens a year, which costs about $3,350 at OpenRouter’s $4.25 output rate, before input tokens. The case for two Mac Studios is control over where data goes, not cheaper tokens.

This is editorial analysis, not a Faraday offering. Faraday Machines installs open-weight models on Mac Studio hardware in Toronto, on M5 Max units with 128 GB each. At 4-bit, MiniMax M3 (~240 GB) fits in three of them (384 GB). DeepSeek V4.1 Flash fits in three as a 1- to 3-bit build of about 250–350 GB, which trades some quality, most noticeably on long or precise work; at 4-bit (about 400–445 GB) it needs four. Muse Spark Open’s M5 Ultra setup is a different machine.

What Origin Changes for a Canadian Buyer

Model origin and API host are separate questions. An open-weight model can be served by anyone who downloads the weights: Muse Spark Open, for example, is already on seven API providers and three cloud platforms. The jurisdiction that matters is the one over whoever processes your prompts.

For a Canadian business, a US-hosted API raises the same CLOUD Act question as any US cloud service, covered in our Canadian AI sovereignty guide. A Chinese lab’s own API puts your prompts under Chinese jurisdiction instead. Running the weights on your own hardware removes the provider from the picture, whether the model came from Meta, Moonshot or Xiaomi.

What self-hosting doesn’t remove is the license. Muse Spark Open’s usage policy bans, among other things, operating critical infrastructure and the unauthorized practice of regulated professions such as finance, law and medicine, and those terms travel with the weights. Your own customers or procurement rules may also care where a model was trained. That is a policy question, and the answer differs by buyer.

Who Is Using Chinese Open Models

The best public usage data comes from OpenRouter, an API marketplace, so it covers one aggregator’s traffic, not direct API use or self-hosting. The figures also use different definitions, so they can’t be strung into one trend line.

The 2025 study covered more than 100 trillion tokens of all OpenRouter traffic. Its 13% figure is a weekly share for Chinese open models alone, averaged over the study period; the August 2026 figure is one month’s share for all open-weight models, limited to US traffic. Both point the same way, but the jump from 13% to 60% compares different things. OpenRouter’s COO described the pattern to The New Stack:

Chinese open models are “disproportionately heavy in agentic flows run by U.S. firms.” — OpenRouter’s COO, quoted by The New Stack (2026)

That is the market Muse Spark Open and Beam are entering: US companies already use Chinese open models heavily for agent work, which is where Muse Spark Open’s tool-use lead would matter most, if it holds up in independent tests.

License Terms Differ More Than Origin

License terms don’t follow the flag. Among the US releases, Beam is Apache 2.0 while Muse Spark Open uses a custom license that requires products to display “Built with Muse Spark Open”. Among the Chinese ones, DeepSeek V4.1 Flash, MiMo-V2.6 and GLM-5.3-Flash are MIT, while Kimi K3, GLM-5.3 and MiniMax M3 use custom licenses. Most custom conditions target very large model-as-a-service providers and leave internal use by ordinary businesses free. Some reach smaller companies too: Muse Spark Open’s attribution rule and usage policy apply at any size, and MiniMax M3’s license requires prior written authorization from any organization above $20 million in yearly revenue. Our open-weight AI license tracker sets out the terms model by model. It isn’t legal advice; check the license text before you deploy.

References