Meta, Reflection and DeepSeek each released a very large open-weight model between Thursday and Saturday. All three are pitched at coding and agents, all three come with 1-million-token context windows, and all three come with licenses that are worth reading before you get excited.
First, a definition. “Open-weight” means the company publishes the trained model’s parameters, the billions of numbers that make it work, so anyone can download and run it. That’s not the same as open source in the full sense: the training data and code usually stay private, and the license decides what you can do with it.
Independent scores have now arrived for all three, and they reorder the pecking order a bit. Here’s the rundown.
Meta: Muse Spark Open (Oct 15)
Meta released the weights for Muse Spark Open, which it calls “the most capable US open-weight model to date.” Note the qualifier: US. It’s a natively multimodal mixture-of-experts model with 816 billion total parameters, 42 billion active per token, and a 1M-token context window.
Mixture-of-experts, or MoE, means the model is split into many specialist sub-networks and only activates a few for each token. That’s why the “active” number is so much smaller than the total: it’s cheaper to run than its size suggests, but you still need memory for the whole thing.
The party trick is local inference. Meta worked with Apple’s MLX framework so the model runs on two M5 Ultra Mac Studios. Each machine’s integrated GPU can use up to 512GB of unified memory, and the chip’s 1.2TB/s of memory bandwidth is fast enough to read “the small slice each token needs,” so the pair “behave like one supercomputer with a terabyte of unified memory,” Meta says.
Meta’s demo was more than a chatbot. It says it used the model on-device to ship real iOS changes to Ice Cubes, an open-source Mastodon client, and to drive DaVinci Resolve video editing end to end.
That is a genuinely interesting milestone, and also a reminder that “runs at home” now means “runs on two top-spec Mac Studios.” Your laptop is not invited.
It’s available through hosts including Baseten, Fireworks AI, Groq, Modal, Nebius, OpenRouter and Together AI, with weights on Hugging Face and the big clouds.
The catch: it ships under the “Muse Model License with limited commercial terms,” which Meta says “allow many developers to use the model for free.” Many is not all. Meta also says the model got extra safety post-training because the weights are public, with a full evaluation report “available soon.”
Reflection: Beam (Oct 16)
A day later, Reflection released Beam, its first open-weight model: a sparse mixture-of-experts with 501 billion total and 23 billion active parameters, built for coding, reasoning and long-horizon agent work.
The license is the headline: Apache 2.0, which is about as permissive as it gets. Reflection shipped weights in BF16, FP8 and NVFP4 (successively smaller number formats that trade a little precision for lower memory use), plus a speculative-decoding drafter model that speeds up generation. It has day-zero support in vLLM, SGLang, Transformers and others, and an open-source coding harness called Mirror CLI.
Reflection is pitching more than a model. It calls the release “an open operating system for agents,” with over 40 partners at launch. On-premises versions come via Dell and HPE for customers who can’t send data to anyone’s cloud.
The scale of the training effort is eye-opening. Reflection says its largest reinforcement learning runs used roughly 1.3 billion sandboxes, and at peak the fleet hit 883,006 vCPUs across 23 clusters on Google Cloud and AWS. It also published a safety technical report alongside the model card, which more labs should do on day one.
DeepSeek: V4.1-Flash-1017 (Oct 17)
Then DeepSeek did its usual thing: called something a “minor update” and posted big numbers. DeepSeek-V4.1-Flash-1017 is a refresh of V4.1-Flash, and the weights are under the MIT License, the other gold standard for permissiveness.
DeepSeek’s own model card claims Humanity’s Last Exam (no tools) rising from 36.8 to 42.6, Terminal-Bench 4.0 from 31.2 to 42.3, and DeepSWE v1.1 from 74.2 to 77.0. On DeepSeek’s table, it beats GPT-5.6 Sol on several agentic tests but still trails Anthropic’s Opus 5 on most of them. The company also says the update cuts task time and output length, which matters more to your bill than any single benchmark.
What the independent tests say
Benchmarking firm Artificial Analysis has now run all three through its Intelligence Index, a composite of ten evaluations. The scores:
- DeepSeek V4.1 Flash 1017: 45, the highest of the three, per Artificial Analysis.
- Muse Spark Open: 39, the top US open-weight score.
- Beam: 30, second among US open-weight models, per its Beam writeup.
For scale, Artificial Analysis says proprietary GPT-6 Astra scores 53. Open models are closing in, but the frontier is still closed.
The details are more interesting than the totals. On Terminal-Bench 4.0, a test of agentic coding in a command line, DeepSeek scored 34% while Muse Spark Open managed 8% and Beam just 3%. Beam did well on AutomationBench-AA, a test of agentic business workflows, at 48.5%. But it hallucinated on 47% of the knowledge questions it tried to answer, against 26% for the UAE’s K2 Horizon.
Price is where DeepSeek runs away with it. Artificial Analysis measured $0.30 per million input tokens and $1.20 per million output for DeepSeek, versus $1.25 and $4.25 for Muse Spark Open. DeepSeek is also wordier, using about 89,000 output tokens per task to Muse’s 40,000, so the gap per task is smaller: $0.26 versus $0.84. Beam was the most frugal, using about 26,000 output tokens per task.
Our read: DeepSeek’s self-reported Terminal-Bench gain doesn’t fully replicate. It claimed 42.3, and Artificial Analysis measured 34%. That doesn’t prove anything sinister, since setups differ, but it’s exactly why self-reported charts are an opening bid.
Why “US” keeps showing up
Both Meta and Reflection lean on “US open-weight” in their positioning, and that’s not an accident. Chinese labs, DeepSeek among them, have dominated the top of the open-weight charts. For enterprises and governments wary of running Chinese models, a capable American alternative is the selling point, even if it scores lower.
Why it matters: the open-model race now runs on two tracks, raw capability and geopolitical comfort. Muse and Beam are competing on the second one as much as the first.
The takeaway
Our read, in three parts:
- Licenses diverge more than benchmarks. MIT and Apache 2.0 versus a custom license “with limited commercial terms” is the difference between “build a business on it” and “call legal first.”
- Self-reported numbers need checking. The first independent results ranked DeepSeek first, exposed weak terminal scores for the US models and didn’t fully confirm DeepSeek’s own claims.
- “Open” is getting heavy. Hundreds of billions of parameters and terabyte-class memory footprints mean open weights mostly benefit people with serious hardware or a hosting bill. That’s still valuable, just not the same thing as “AI for everyone.”
What to watch: Meta’s promised full evaluation report, independent safety testing of all three, Reflection’s self-distillation recipes, and whether any of these models shows up in real coding tools rather than benchmark tables.
Image: Jernej Furman from Slovenia / Wikimedia Commons, CC BY 2.0
.jpg){kind=link}