America just fired its biggest open-weight shot yet in the AI race — and this time, nobody has to ask permission to use it. On October 16, Reflection AI, a Brooklyn startup backed by Nvidia and founded by two former Google DeepMind researchers, published the full weights of its first model, Beam, on Hugging Face under the permissive Apache 2.0 license. Beam is a 501-billion-parameter model built for coding and AI agents, and until this week it was waitlist-only. Now anyone can download it, fine-tune it, and run it on their own servers — no API contract, no per-token bill, and no data leaving the building.
The timing is the story. Chinese open models from DeepSeek, Kimi and Qwen have spent 2026 embarrassing expensive US models on coding scoreboards at a fraction of the price. Beam is the most serious American answer so far: a frontier-scale model, released completely open, in the same week that saw three major open-weight coding models land within 48 hours of each other.
Key Takeaways
- Reflection AI released Beam 1.0's open weights on Hugging Face on October 16, 2026, under Apache 2.0 — free for commercial use, with post-training recipes included.
- Beam has 501 billion total parameters but activates only about 23 billion per token, with a 1-million-token context window. It is text-only and built for coding, terminal work and long-running AI agents.
- Reflection's published scores are strong — 81.6% on SWE-Bench Verified and 97.7% on AIME 2026 — but on DeepSWE v1.1 it trails China's GLM-5.3 and Kimi K3. All scores are vendor-reported, not independent tests.
- You cannot run this on a laptop: the full download is about 1 TB, and even the smallest checkpoint needs several high-end GPUs. Most people will try it through cloud providers instead — Ollama Cloud is serving it free for a limited time.
- Reflection is Nvidia-backed, was founded in 2024 by ex-DeepMind researchers Misha Laskin and Ioannis Antonoglou, and trains partly on SpaceX's Colossus 2 data center.
What Actually Happened on October 16
Reflection first introduced Beam on October 5, while the model was still in final red-teaming, with access limited to an early-access waitlist and a promise that the weights would follow later in the month. On October 16, Hugging Face published its "Welcome Beam" launch post and the promise became real: Beam 1.0 (officially Beam-1.0-501B-A23B) went live with checkpoints in three formats — BF16 at roughly 934 GB, FP8 at about 469 GB, and NVFP4 at about 273 GB — plus a small DFlash2 drafter model for faster generation.
The release came with day-zero support in the tools developers actually use: Transformers, vLLM and SGLang could serve it immediately. Reflection says more than 40 partners and integrations launched alongside it, including Amazon Web Services, Microsoft Azure, Google Cloud, Baseten, Fireworks AI, Together AI, CoreWeave, Databricks and Ollama. Reflection also opened a public beta of its own API, free to start, and Hugging Face is offering $10,000 in Hub credits for strong community projects built on Beam. This was not a quiet research drop — it was a full ecosystem launch.
Inside Beam: A Giant Model That Only Wakes Up a Slice of Itself
Beam is a sparse mixture-of-experts model: 52 layers with 256 experts in each block, of which only 8 are routed to for any given token. That is how a 501-billion-parameter model activates just 23 billion parameters at a time — big total knowledge, much smaller running cost. Its context window stretches to 1 million tokens, it handles text only (no images or audio), and its knowledge cutoff is June 30, 2026.
The training run behind it was enormous. Reflection says Beam was pretrained on 23.8 trillion tokens in under four weeks on 6,144 Nvidia GB300 GPUs, then refined with reinforcement learning on 10,500 GB300 GPUs for four weeks, generating more than 100 million practice attempts across nearly a million software-engineering, terminal, coding and tool-use environments. One unusual design choice: Beam was trained with a length penalty that rewards correct answers reached with fewer tokens, which is why it offers reasoning-effort settings from low to max — quick answers when you want speed, deep thinking when the problem earns it.
The Benchmarks: Genuinely Strong, With One Honest Gap
Reflection's published numbers put Beam in serious company: 81.6% on SWE-Bench Verified (real GitHub issue fixing), 80.1 on Terminal Bench v2.1, 97.7% on the AIME 2026 math competition, 90.6% on GPQA Diamond science questions and 88.3% on LiveCodeBench. For an open-weight model anyone can download, those are frontier-adjacent figures.
Now the part the hype posts skip. On DeepSWE v1.1 — the same test where, as we reported yesterday, DeepSeek's new Flash checkpoint beat Claude Opus 5 — Beam scores 44.4, well behind GLM-5.3 at 61.0 and Kimi K3 at 68.0, and roughly level with the older GLM-5.2. Reflection's own pitch is efficiency, not the top score: competitive results while using three to four times less inference compute than rivals. Every figure above comes from Reflection and Hugging Face's republication of it — independent leaderboards have not caught up yet, so treat the table as a claim to be tested, not a verdict.
Why This Release Matters in the United States
For two years, the open-weight race has had an uncomfortable shape for Washington and Silicon Valley: the downloadable frontier models kept coming from China, while US labs kept their best systems behind paid APIs. That mattered beyond bragging rights. A company that downloads open weights can fine-tune the model on its own data, run it inside its own firewall, and never send a single prompt to anyone's cloud — which is exactly what banks, hospitals, defense contractors and law firms need before they will touch AI at all.
Beam changes the American side of that board in the same week Meta pushed the open model it built with Apple silicon in mind — the release we covered in our report on Meta's Muse Spark Open running on Mac Studios. The contrast between the two is instructive: Muse Spark Open is bigger and multimodal but carries Meta's custom license with conditions, while Beam is text-only with the cleanest license in the business. Apache 2.0 means a US startup can build a commercial product on Beam without a revenue clause or attribution rule hanging over it. Among developers this week, the summary making the rounds was simple: Nvidia's new Bolt is the one you can run at home, Muse has the broadest inputs — and Beam has the license you would choose for a business.
The Hardware Reality: Why Almost Nobody Will Download 1 TB
Open does not mean small. Hugging Face is blunt that the BF16 checkpoint is about a terabyte, and even the compressed NVFP4 version needs roughly 280 GB of GPU memory — a realistic serving setup is four Nvidia B200 cards, or eight H200-class cards for the FP8 build. This is a data-center model, full stop.
That is why the provider list matters more than the download button for most US readers. Beam is already served by Hugging Face Inference Providers, Baseten, Fireworks and Together; it sits in AWS SageMaker JumpStart in the US East regions; OpenRouter lists it at $1 per million input tokens and $3 per million output tokens; and Ollama Cloud made it generally available on October 16, free for a limited window with zero data retention. In other words, you can test Beam this weekend from a browser — just do not expect your gaming PC to join in.
⚠️ Scam Warning: Fake "Beam" Downloads Are a Real Risk
Every major open-model release triggers the same trap within hours: lookalike repositories, "cracked" or "pre-quantized" copies on random file hosts, and Telegram and Discord links promising a smaller, faster Beam. A model file is executable-adjacent baggage — a poisoned checkpoint or a fake installer is an easy way to hand someone your machine. Only download from Reflection's official Hugging Face organization, or use a named provider such as Ollama, Baseten, Fireworks or Together. Reflection will never ask you to pay a "release fee," connect a crypto wallet, or enter account credentials to get the weights — anyone who does is running a scam. And if a "Beam-powered" app asks for access to your photos, files or messages, close it: the real model is text-only.
Before You Try Beam: A Quick Checklist
- Start hosted, not local. Use Ollama Cloud's free window, the free Reflection API beta, or a Hugging Face Inference Provider before pricing any hardware.
- Check the source. Weights should come only from Reflection's official Hugging Face page, under Apache 2.0, with the model card attached.
- Test on your own tasks. Run your real coding and agent workloads against Beam and your current model — vendor benchmarks, including the ones above, are a starting map, not a buying decision.
- Count the real cost. Self-hosting means multiple data-center GPUs around the clock; for most teams the $1/$3 OpenRouter pricing or a dedicated provider will be cheaper.
- Remember it is text-only. If your workflow needs image or document understanding built in, you will need separate tooling around Beam.
Frequently Asked Questions
Is Reflection AI's Beam really free?
Yes. The weights were released on October 16, 2026 under the Apache 2.0 license, which allows commercial use without royalties. The model itself is free — the GPUs or cloud time needed to run it are not, which is why most people will access it through a provider.
Can I run Beam on my home PC or phone?
No. Even the smallest official checkpoint is about 273 GB and needs several high-end data-center GPUs. Community quantized builds may shrink that over time, but today Beam is a server model. Hosted options like Ollama Cloud and OpenRouter are the realistic way to try it.
How does Beam compare with DeepSeek and Kimi?
On Reflection's own tables, Beam is competitive with GLM-5.2 on coding and agent tasks while using less inference compute, and its SWE-Bench Verified score (81.6%) is strong. But on DeepSWE v1.1 it trails GLM-5.3 (61.0) and Kimi K3 (68.0) by a wide margin at 44.4. Independent testing has not yet confirmed any of these figures.
Who is Reflection AI?
A Brooklyn, New York startup founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou. It is backed by Nvidia, focuses on automating software development, and earlier this year signed a deal for additional computing capacity at SpaceX's Colossus 2 data center.
What can Beam actually do?
It is built for software engineering, terminal and DevOps work, tool calling through standards like MCP, and long multi-step agent tasks, with a 1-million-token context window and adjustable reasoning effort. It is text-in, text-out only — no native image, audio or video understanding.
The Bottom Line
October 16 did not end the US–China open-model race, and Beam's own numbers admit it: on at least one headline coding test, the Chinese models are still ahead. What changed is that the United States now has a downloadable, commercially clean, frontier-scale entry of its own — trained by ex-DeepMind founders, backed by Nvidia, and open enough that any American company can build on it without asking anyone's permission. In a week when three open coding models landed in 48 hours, that license page may matter as much as any benchmark.
One habit is worth keeping no matter which model wins: your private files should never become training material or chat attachments in the first place. Pro Calculator Secure Vault keeps photos, videos, audio, documents and private notes behind sign-in protection on your own device, with an intruder selfie, an incognito private browser, and 8 smart calculators (standard, floating, BMI, age, currency, tip, GST & VAT, loan EMI) — local on-device storage only, with no data collected or shared with third parties. More from the maker on the ZAS developer page on Google Play.
No comments:
Post a Comment