Reflection AI's 501-Billion-Parameter "Beam" Goes Open: The Biggest Open-Weight AI Release From a US Lab Yet

Reflection AI, a Brooklyn startup founded by former Google DeepMind researchers and backed by Nvidia, has published the weights of "Beam" — a 501-billion-parameter mixture-of-experts model for coding, reasoning and agentic work — on Hugging Face under the permissive Apache 2.0 license. Announced October 5 and released as open weights on October 16, 2026, Beam is described as the largest frontier-class open-weight release from a US lab, arriving as the global race to give away — or lock up — the most powerful AI models intensified.

Illustration: a glowing neural network over a laptop with an open padlock, symbolizing Reflection AI's open-weight release of Beam. This is an AI-generated illustration, not a real scene.

What Happened

On October 16, 2026, Reflection AI published Beam's full model weights, a technical report, a model card and developer artifacts on Hugging Face under Apache 2.0 — one of the most permissive open-source licenses, which lets anyone download, modify and deploy the model, including commercially. The release was accompanied by Mirror, an open-source terminal agent harness built specifically for multi-step coding and tool-use tasks, and by a companion "drafter" model for speculative decoding, according to coverage compiled by TechSambad.

The launch mobilized an unusually broad ecosystem: more than 40 partners supported the release from day zero, with day-one availability in Transformers, vLLM, SGLang and Unsloth, hosting on AWS, Google Cloud, CoreWeave, Baseten, Fireworks and Ollama, and on-premise options from Dell and HPE, TechSambad reported.

The numbers behind Beam are striking. According to Reflection AI, the model carries 501 billion total parameters but activates only about 23 billion per token — roughly 4.6% of the full count — the defining trick of mixture-of-experts design. Pretraining took under four weeks on 6,144 Nvidia GB300 GPUs over 23.8 trillion tokens, the company says, and reinforcement learning used 10,500 GB300 GPUs for four weeks, generating more than 100 million rollouts.

Benchmark figures published by Reflection put Beam at 80.9 on SWE-bench Verified and 80.1 on Terminal Bench v2.1, with a 1-million-token context window. But analysts stress these are vendor-reported: Layer3Labs noted that TechCrunch pointed out outside labs had not verified them at the announcement, and ExplainX's comparison showed Beam trailing GLM-5.3 and Kimi K3 on most coding benchmarks — roughly level with GLM-5.2.

Background

Reflection's pitch is not that Beam is the smartest model — it is the most efficient intelligence you can actually own. The company's framing is "the Western open-weight frontier": where Chinese labs — DeepSeek, Qwen, Kimi, Z.ai's GLM — have dominated open-weight releases, Beam is the American counterweight competing on inference efficiency. Complete AI Training summarized the claim: Beam matches Z.ai's GLM-5.2 reasoning at three to four times less inference compute.

The lab's pedigree matters to that pitch. Founded by former Google DeepMind researchers and backed by Nvidia, Reflection is validated on the Dell AI Factory with Nvidia, is a certified consortium member of the US Department of Energy's Genesis Mission serving the 17 US National Laboratories, and is building a 250-megawatt sovereign AI cloud in South Korea with the Shinsegae group, backed by the US and Korean governments, according to Unite.AI. Press reports cited by ExplainX put the company's compute commitments above $7 billion through 2029.

Beam did not arrive alone. TechSambad's roundup noted that DeepSeek also opened new frontier weights in the same days, and that Meta released an 816-billion-parameter "Muse Spark Open" model — making mid-October 2026 one of the densest open-model release windows on record. On the closed side, the contrast is sharp: the most capable models are increasingly withheld or gated. Beam's Apache 2.0 license is the philosophical opposite — full weights, reproducible safety documentation, and the internal safety evaluations Reflection says it developed, all downloadable.

Reflection's safety approach is itself part of the story: the lab trained a second model from the same pretrained checkpoint with a separate fine-tuning and RL pipeline targeting behavioral principles — rules the model must not break, qualities it should satisfy — then merged it with the large-scale RL teacher, according to Unite.AI. The safety dataset was built adversarially, with each round's successful attacks folded into the next, and the company says it will publish its safety evaluations and open-source them.

Illustration: glowing code streams flowing into a terminal window, symbolizing the coding and agentic workloads Beam is built for. This is an AI-generated illustration, not a real scene.

Why It Matters

Beam matters for three reasons, and only one of them is the model itself.

First, the license. Apache 2.0 means Beam can be downloaded, fine-tuned, embedded in products and run on a company's own servers — including in regulated industries and government labs that cannot send data to someone else's API. ExplainX's analysis makes the practical point bluntly: do not plan production around weights you cannot download, and as of mid-October you can download Beam. For enterprises told AI is infrastructure, this is the first 501B-parameter piece of infrastructure from a US lab they can actually own.

Second, the efficiency argument reframes the open-weight race. The Chinese labs still lead on raw benchmark scores, and Beam's vendor-reported numbers trail GLM-5.3 and Kimi K3 by clear margins. Reflection's counter is "intelligence per token": if you can get GLM-5.2-class reasoning at a third of the inference compute, the economics of running agents at scale change. Whether the efficiency claim survives independent replication is the question the open weights now let anyone test.

Third, the timing. Beam's release landed in the same week Beijing drafted sweeping export controls on AI models, solar tech and biotech — and as Washington debates how much of the AI stack should stay under American control. An American startup giving away a frontier-class model under Apache 2.0 complicates both narratives at once: it demonstrates US AI capability while diffusing that capability to anyone with a GPU cluster. The geopolitics of open weights is now the central unresolved question of AI policy, and Beam is its latest test case.

What Happens Next

The near-term test is adoption. With day-zero support across the major inference frameworks and hosting partners, Beam should show up quickly in independent benchmarks, in community-quantized builds that can run on smaller hardware — ExplainX estimates roughly 500 GB of memory at 8-bit precision, a multi-GPU server rather than a workstation — and in enterprise pilots weighing it against the Chinese open models that currently lead the category.

The competitive response is already forming: Mistral's trillion-parameter Large 4 has open weights promised for late October, and the Chinese labs show no sign of slowing. The open-weight leaderboard that Beam just joined is being rewritten monthly; Beam's position will depend on whether independent benchmarks confirm Reflection's efficiency claims once the community gets its hands on the weights.

Illustration: rows of servers in a futuristic data center, symbolizing the massive GPU infrastructure behind frontier AI training. This is an AI-generated illustration, not a real scene.

Longer term, Beam is a bet that the future of AI is not a handful of giant APIs but an open operating layer for agents — downloadable, inspectable, runnable anywhere. If Reflection is right, the most important number in its launch is not 501 billion parameters, but two words in the license: Apache 2.0.

Sponsored: explore today's featured sponsored offers

FAQ

What is Reflection AI's Beam?

Beam is Reflection AI's first model: a 501-billion-parameter sparse mixture-of-experts model activating about 23 billion per token, built for coding, reasoning and long-horizon agentic work, with a 1-million-token context window. The Brooklyn startup — ex-Google DeepMind researchers, backed by Nvidia — announced it on October 5, 2026, and published the weights on October 16.

What does "open-weight" mean here?

Reflection published Beam's full weights on Hugging Face under the Apache 2.0 license. Anyone can download, modify, fine-tune and deploy the model — commercially and on their own servers — along with a technical report, model card, and the open-source Mirror terminal agent harness.

How good is Beam compared to rival models?

On Reflection's own numbers, Beam scores 80.9 on SWE-bench Verified and 80.1 on Terminal Bench v2.1 — but these are vendor-reported and outside labs had not verified them at announcement. It trails Chinese open models like GLM-5.3 and Kimi K3 on coding benchmarks, landing roughly level with GLM-5.2. Reflection's pitch is efficiency: GLM-5.2-class reasoning at three to four times less inference compute.

Why does an open release from a US lab matter?

Open-weight releases have recently been dominated by Chinese labs, while the most capable US models are increasingly closed or gated. Beam is described as the largest frontier-class open-weight release from a US lab, and its permissive license lets enterprises and government labs — including the 17 US National Laboratories Reflection serves — run a frontier-adjacent model on their own infrastructure.

What should I watch next?

Independent benchmarks now that the weights are downloadable, community-quantized builds for smaller hardware, and enterprise adoption against Chinese rivals. Watch too for Mistral's trillion-parameter Large 4, whose open weights are promised for late October — and for the geopolitics of free model weights, with Beijing drafting new AI export controls the same week Beam shipped.

Sources

• TechSambad (October 17, 2026) — October 16 weight release, Apache 2.0, Mirror harness, 40+ partners, 1.3B RL sandboxes, Meta Muse Spark Open context.
• Shattered.io (October 2026) — 501B/23B specs, DFlash2 drafter, Mirror CLI, Hugging Face distribution.
• DataCamp (October 2026) — October 5 announcement, "Western open-weight frontier" framing, Chinese-lab comparisons.
• ExplainX (October 2026) — benchmark comparisons, training details, $7B+ compute commitments (press reports), memory estimates.
• Layer3Labs (October 2026) — vendor-reported benchmarks, TechCrunch note on unverified numbers at announcement.
• Complete AI Training (October 2026) — Nvidia backing, ex-DeepMind founders, GLM-5.2 parity claim, early coverage framing.
• Unite.AI (October 2026) — safety methodology, DOE Genesis Mission, 250MW South Korea cloud.