Open-weight models, the ones you can download and run yourself, have been dominated by Chinese labs for most of 2026. October changed that. After Meta's Muse Spark Open, two more US releases landed in the same week: Nvidia's Nemotron 3.5 family on 15 October, built for AI agents from data centres down to laptops, and the downloadable weights of Reflection AI's Beam on 16 October, a 501-billion-parameter model under the permissive Apache 2.0 licence.
Nvidia Nemotron 3.5: agents from server to laptop
Nvidia's open models are meant to be customised: the company publishes the weights together with training data and recipes, so businesses can post-train them on their own data and run them on their own infrastructure. The 3.5 release has three members.
Super VL
120B- Inputs
- text and images, including PDFs, tables, charts and web pages
- Runs on
- data centre GPUs, cloud, DGX Station and RTX Pro workstations
- Nvidia says accuracy is close to its largest reasoning model at about a fifth of the size
- Used or integrated by Salesforce, CrowdStrike, Perplexity, Atlassian, Ramp and EY
- Speed claims (up to 5x rivals' throughput) are Nvidia's own
For agents that work through business documents.
Bolt 74B-A9B
24GB- Runs on
- a single RTX GPU or DGX Spark
- Tools
- llama.cpp, Ollama, LM Studio, Unsloth, vLLM
- Native 2-bit quantisation, so it leaves room for other apps
- Top open model in its class on PinchBench, per Nvidia
For a private research or coding agent on your own machine.
Bolt E4B
4GB- Runs on
- GeForce RTX, Jetson and Spark devices
- Status
- coming later, no date yet
For local voice assistants and embedded interfaces.
The pitch to companies is control. A bank or an insurer can take Super VL, train it on its own claims or statements, and keep every document inside its network. Nvidia names use cases such as document intelligence, financial accounting and insurance claims, and says that, once post-trained for a specific task, Super VL finishes it faster than the big closed models. Those comparisons come from Nvidia and its partners, so treat them as a starting point for your own tests.
Reflection Beam: big, open and efficient
Reflection AI announced Beam on 5 October and published the weights on Hugging Face on 16 October, in BF16, FP8 and NVFP4 versions plus a small "drafter" model that speeds up generation. It is a mixture-of-experts model: 501 billion parameters in total, but only 23 billion work on each token, which keeps it relatively cheap to serve.
- Size
- 501B total, 23B active per token
- Context
- trained up to 256K, extended to 1M tokens
- Inputs
- text only
- Licence
- Apache 2.0, with no limits on fine-tuning, distillation or merging
- Training
- 23.8 trillion tokens on 6,144 Nvidia GB300 GPUs in under four weeks
- Hardware to self-host
- one node with four Nvidia B300 GPUs, using the 273 GiB NVFP4 build
- Serving
- Transformers, vLLM, SGLang, Unsloth, Axolotl and Nvidia Dynamo, plus Reflection's own API
Higher is better. Reflection-reported, 8 October 2026.
Reflection is candid that Beam is not the strongest open model: Kimi K3 stays ahead on raw capability, and DeepSeek V4.1 Flash beats it on some agent-style coding tests. Its selling points are efficiency (Reflection says it matches GLM-5.2 on reasoning with three to four times less compute) and the clean Apache 2.0 licence. The independent Artificial Analysis Intelligence Index puts it at 30, 19th of 120 models, while using fewer output tokens than most.
Higher is better. Independent scores, mid-October 2026.
Which open model for your agent?
- 1
Start from where it must run
A laptop or single GPU points to Nemotron Bolt; your own data centre opens up Super VL or Beam; anything bigger is easier through a hosted API.
- 2
Check the licence
Beam's Apache 2.0 is the simplest for products. Muse Spark Open and Nvidia's models have their own terms, so read them before you ship.
- 3
Test on your tasks
Leaderboards rarely reflect your documents or your SQL. Run 30 to 50 real examples through two or three candidates.
- 4
Plan for fine-tuning
Nvidia publishes data and recipes; Beam allows any fine-tuning. Both make post-training on your data realistic.
Check your understanding
1/3Which Nemotron 3.5 model is designed to run on a single 24GB GPU?
Native 2-bit quantisation fits its 74B parameters into 24GB.
2/3Under which licence did Reflection release Beam's weights?
Apache 2.0 places no limits on fine-tuning, distillation or merging.
3/3What hardware does Reflection suggest for self-hosting Beam?
Using the 273 GiB NVFP4 version of the weights.
You got 0 out of 3.
More open models: Meta Muse Spark Open, Mistral Large 4 and DeepSeek's DeepInfer and V4.1 Flash. For the full league table, read the best AI models in October 2026.
Sources: Nvidia's Nemotron 3.5 announcement as reported by IT Brief and TechCrunch (October 2026); Reflection AI's "Introducing Beam" post (benchmarks updated 8 October 2026) and the Beam 1.0 model release (16 October 2026); Artificial Analysis Intelligence Index v4.3.2. Vendor benchmarks are the companies' own.
Alessandro Ecclesie AgazziWritten by Alessandro Ecclesie Agazzi, analytics engineer in London. Updated 21 October 2026. About me