Five open-weight AI models in three days. StepFun beats Meta, and none beat the closed frontier.
Meta's Muse Spark Open, Reflection's Beam, StepFun's Step 5 Preview, NVIDIA's Nemotron 3.5 and a DeepSeek update all shipped downloadable weights this week. How they compare on size, license, price and independent scores, and when open weights are worth it.
October 16, 2026
In the space of three days this week, five notable AI models shipped with downloadable weights. Meta released Muse Spark Open, its biggest open release since the Llama era. Reflection AI put out the weights for Beam, StepFun released Step 5 Preview, NVIDIA shipped Nemotron 3.5 Super VL, and DeepSeek updated V4.1 Flash. Mistral's trillion-parameter Large 4 is next, with weights due October 31.
Open weights mean you can download the model and run it on your own hardware, fine-tune it, and keep your data in-house. Here is how this week's releases compare.
The five releases
| Model | Size (total / active) | License | Hosted price per 1M (in / out) | Artificial Analysis index |
|---|---|---|---|---|
| Step 5 Preview (StepFun) | 600B / 27B | StepFun Model License (commercial thresholds) | $1.00 / $2.70 | 44 |
| Muse Spark Open (Meta) | 816B / 42B | Muse Model License (limited commercial terms) | $1.25 / $4.25 | 39 |
| Beam (Reflection AI) | 501B / 23B | Apache 2.0 | $1.00 / $3.00 | 30 |
| Nemotron 3.5 Super VL (NVIDIA) | 120B / 12B | OpenMDW-1.1 | $0.10 / $0.40 | 26 |
| DeepSeek V4.1 Flash (Oct 17 update) | Not disclosed | MIT | $0.15 / $0.60 | - |
Hosted prices are from OpenRouter; scores are from Artificial Analysis, where available. All five are Mixture-of-Experts designs that activate only a fraction of their parameters per token, and all except DeepSeek's support a context window of about 1 million tokens.
What stands out
- StepFun scores highest. Step 5 Preview's 44 on the Artificial Analysis index is the best of the group, ahead of Meta's much larger model.
- Meta is back in open models, with conditions. Muse Spark Open is natively multimodal (text, image, audio and video in) and Artificial Analysis rates it the strongest US open-weight model at release. Its license carries commercial limits, so check them before building a product on it.
- Beam has the cleanest license. Reflection's model ships under Apache 2.0, the most permissive terms here. Its first independent score (30) trails the company's own launch claims, so test it on your own tasks.
- NVIDIA is the budget pick. Nemotron 3.5 Super VL is small enough to run on a single GPU, costs $0.10/$0.40 hosted, and reads images, documents and video.
- DeepSeek keeps iterating. The V4.1 Flash update lifted its Terminal-Bench 4.0 score from 31.2 to 42.3, according to third-party reporting.
Open does not automatically mean cheaper
If you simply call these models through a hosted API, most are not cheaper than the budget closed models. Claude Haiku 5.5 and GPT-6 Luna both cost $0.10/$0.50 per million tokens, less than every model in the table except NVIDIA's. And the closed frontier is still well ahead: Claude Opus 5.5 scores 58 and GPT-6 Astra 53 on the same index.
The reasons to choose open weights are control rather than price: running inference on your own infrastructure for data privacy or residency, fine-tuning on proprietary data, avoiding vendor lock-in, and predictable costs at high volume on hardware you own. If none of those apply, a cheap closed model is usually the simpler choice.
Which to try
For the best quality you can self-host, start with Step 5 Preview and check its license thresholds. For multimodal work under a big-company brand, try Muse Spark Open and read Meta's terms. For the fewest legal questions, Beam's Apache 2.0 license is the easiest to clear. For a small, cheap model on modest hardware, use Nemotron 3.5 Super VL. Our model index has full specs for each, and the AI Pricing Index tracks their prices over time.
Some links in this article are affiliate links. Learn more.