Yesterday, the rumor became a product. Nvidia-backed Reflection AI unveiled Beam, its first open-weight model — a 501-billion-parameter mixture-of-experts system aimed squarely at China’s DeepSeek, Qwen, and Z.ai’s GLM. It is the highest-profile American answer yet to the open-model wave, and for readers of this site, it raises the question we get asked most: does “open weights” mean “uncensored”?
Short answer: not exactly. Longer answer below — plus what we’ll test the moment the weights actually drop.
TL;DR — What We Know
| Confirmed | Not yet |
|---|---|
| Model name: Beam, unveiled Oct 5, 2026 | Weights not yet released (coming later in October) |
| 501B total params, 23B active per task (MoE), text-only, 1M-token context | Apache 2.0 license planned, not yet attached |
| Coding + agentic focus; pre-trained on 23.8T tokens | Benchmarks are company-reported, not independently verified |
| Founded 2024 by ex-DeepMind researchers; Nvidia-backed | Early access via waitlist only (final red-teaming in progress) |
| $7B+ in compute deals (SpaceX, Nebius) for GB300s through 2029 | No hands-on refusal testing possible yet — that’s our job when weights land |
What Beam Claims
Reflection’s pitch isn’t “we beat everyone.” It’s efficiency: Beam reportedly matches Z.ai’s GLM-5.2 on advanced reasoning while using roughly one-quarter to one-third the inference compute — 3–4x more efficient than rival Western open models on coding and agentic tasks, achieved through high-compute reinforcement learning during training.
On the company’s own scoreboard: 80.9 on SWE-bench Verified (vs 70.7 for Nvidia’s Nemotron 3 Ultra) and 80.1 on Terminal Bench v2.1 — close to GLM-5.2’s 81.0, though DeepSeek V4.1 Flash (90.6) and Kimi K3 (88.3) still lead there. Treat every one of these numbers as provisional: no public checkpoint exists yet, so nobody outside Reflection can reproduce them.
The business model matters more than the scoreboard. Reflection isn’t selling a chatbot — it’s selling “AI factories”: enterprises and governments take the weights, add their own data and compute, and run customized systems locally. The target customer is the bank or agency that won’t touch a Chinese model for security reasons but wants Chinese-model economics.
Open Weights vs. Uncensored: The Honest Breakdown
This is the part most coverage skips, and it’s the whole reason we’re writing about Beam on this site.
Where “open” genuinely helps the unfiltered case:
- Downloadable weights mean self-hosting. No vendor API, no vendor content filter sitting between you and the model. That is one of the few real paths to unfiltered use — and it’s why open-weight releases matter to this audience at all.
- 23B active parameters is meaningful. A 501B dense model is datacenter-only; a 23B-active MoE, quantized, is potentially runnable on high-end consumer hardware. The barrier is lower than the headline parameter count suggests.
Where “open” does not mean “uncensored”:
- Beam went through safety training (deliberative alignment, per the company’s materials). Open weights ≠ unaligned weights. The refusals are baked into the checkpoint, not just the API wrapper.
- Beam is a coding/agentic model, not a conversational companion. “Can run it” is not the same as “want to chat with it.”
- The license draws the real boundary. Apache 2.0 is planned — permissive if it holds — but until the weights ship with actual license text, the commercial and modification terms are unknown.
Our standing position, unchanged: this site reviews and routes; we don’t host generation. Open weights shift the responsibility — and the freedom — to whoever runs the model. That’s worth understanding clearly before celebrating.
What We’ll Test When the Weights Land
Reflection says weights drop later this month with a technical report and model card. The day they do, Beam goes into our hands-on protocol, same as Venice, ChatGPT, Grok, and Nastia:
- Refusal behavior — our standard prompt set, run locally, scored against Venice / DeepSeek / Qwen on the same tasks.
- License teardown — what Apache 2.0 (or whatever ships) actually permits: commercial use, fine-tuning, redistribution.
- Local inference reality check — what hardware it actually takes, quantized, with honest tokens/sec numbers — not marketing claims.
No independent test exists anywhere right now. Ours will be the first in this format.
FAQ
Is Beam available now?
Announced, not downloadable. Early versions are waitlist-only while final red-teaming finishes; public weights are slated for later in October 2026.
Is Beam uncensored?
No — and Reflection doesn’t claim it is. It received safety/alignment training. What open weights give you is the ability to self-host, which removes the vendor’s runtime filter but not the model’s training.
How does Beam compare to DeepSeek and Qwen?
By the company’s telling: comparable reasoning to GLM-5.2 at a fraction of the inference cost, closing in on Qwen on coding/agentic. Independent verification has to wait for the weights.
Should I wait for Beam or use something existing?
If you want an unfiltered chat experience today, that’s a different article — see our hands-on tested roundup. Beam is infrastructure news, not a chatbot launch. Its importance is what it enables, not what it chats like.
Sources: Reuters, Oct 5 · AIStockWire summary of TechCrunch reporting · MarkTechPost spec breakdown · Startup Fortune analysis