Reflection AI Open-Sources Its 501B Model While Nvidia Weighs Buying the Company

On October 16, Reflection AI pushed the weights for Beam, a 501-billion-parameter mixture-of-experts model, onto Hugging Face under an Apache 2.0 license. The model card lists 23 billion active parameters per token, a million-token context window, and checkpoints in BF16, FP8, and NVFP4. Reflection, a startup founded in 2024 by two former Google DeepMind researchers, is pitching Beam as proof that a Western lab can ship an open-weight model at the scale and permissiveness of China's Qwen, Kimi, and GLM families, rather than ceding that category entirely.

The numbers Reflection volunteers are competitive on paper: 90.6% on GPQA Diamond, 81.6% on SWE-bench Verified, 97.7% on AIME 2026. The company's own framing is more modest than the benchmark table suggests. It describes Beam as "competitive" with Zhipu's GLM-5.2 and "approaching" Alibaba's Qwen3.8-Max on coding and agentic work, while conceding that Moonshot's Kimi K3 still leads on raw capability. Its actual sales pitch is efficiency, not supremacy: Reflection says Beam needs three to four times less inference compute than GLM-5.2 to reach comparable reasoning scores, the result of a reinforcement-learning phase that ran 100 million-plus rollouts across 10,500 Nvidia GB300 GPUs over four weeks. None of those figures carry independent verification; no outside benchmarking group had published its own numbers on Beam as of this week.

The timing nobody put in a press release

What makes Beam's release more interesting than another open-weight entrant is what was happening in the two weeks before the weights went live. On October 11, the Financial Times reported that Nvidia — Reflection's lead investor — was in early talks to either buy the company outright, take a larger stake, or acquihire its research team and license its technology without taking control. The FT's sources described an acquihire as the structure most likely to sidestep a lengthy antitrust review, and cautioned that no terms were agreed and the talks could still collapse. Neither company has commented publicly. Reflection shipped Beam's weights five days after that report broke, not before it.

"A more bifurcated world" is how Omdia analyst Lian Jye Su described the dynamic Beam represents — Western labs feeling they need to offer an open alternative specifically because Chinese labs got there first, with Reflection's Nvidia backing mattering as much as its benchmark position.

The entanglement between Reflection and its would-be acquirer goes well beyond a term sheet. Nvidia put $800 million of equity into a $2 billion investment in October 2025, at a valuation press reports later put as high as $25 billion pre-money. Beam itself was trained on Nvidia GB300 chips that Reflection doesn't own outright — it rents them through a June 2026 agreement with SpaceX worth up to $6.3 billion, including roughly $150 million in monthly payments for capacity at SpaceX's Colossus 2 facility in Memphis through 2029, on top of a separate $1 billion compute deal with Nebius. One analysis of Beam's own benchmark chart also notes that Nvidia's Nemotron models appear on it as a comparison point — a startup benchmarking itself partly against models made by the company that might soon own it.

There's a reason the sequencing matters beyond gossip. Apache 2.0 weights, once mirrored across Hugging Face and the dozen inference platforms that added day-zero support for Beam, can't be un-published — nobody, including a new owner, can retract code that's already been copied onto thousands of machines. That's precisely why the order of operations here is notable: Reflection got Beam out under a genuinely permissive license before any acquisition closed, which locks in openness for this specific model regardless of what Nvidia decides next. It says nothing about whether Beam-2 ships the same way, or whether anyone keeps paying the SpaceX and Nebius compute bills to support it, if Reflection's independence ends.

The credibility question sits underneath all of this, and it has history. A different company, the New York startup HyperWrite, released a model called Reflection 70B in September 2024 that it called "the world's top open-source model" — a claim that collapsed within days once outside testers found its public weights didn't match the announced scores, and accusations of outright fabrication followed. Reflection AI has no connection to HyperWrite beyond sharing a word in its name, but the coincidence is a useful reminder: open-weight launches in this category have a specific track record of self-reported numbers not surviving contact with independent evaluation, and Beam's benchmark table, like Mistral's for the still-unreleased Large 4 and Meta's for Muse Spark Open, is entirely vendor-supplied for now.

What happens next matters more than the launch itself. If Nvidia does buy or absorb Reflection, the open weights already in circulation stay open, but the roadmap, support, and the next model's license become decisions made inside the company that also sells the GPUs everyone is trying to use less of. If the talks fall apart, Reflection keeps burning roughly $150 million a month on compute to support a model it's giving away for free, betting that enterprises wary of Chinese-origin weights will pay for its hosted API instead. Either way, the test that actually settles whether Beam earned its billing is the one nobody involved can self-report: independent evals from groups like Epoch AI or Artificial Analysis, and whether real deployments bear out the efficiency claim once the marketing is stripped away.