Reflection AI Launches Beam, a 501-Billion-Parameter Rival to China’s Open Models

Last updated: 6 October 2026. Figures below are sourced to primary publications and contemporaneous reporting, linked inline; figures self-reported by a company are flagged as such.

The 60-second version

• On 5 October 2026, Brooklyn-based startup Reflection AI released Beam, a 501-billion-parameter open-weight “mixture of experts” language model, with full weights and technical details promised later this month.
• Reflection says Beam matches Zhipu/Z.ai’s GLM-5.2 on advanced reasoning benchmarks while using 3–4x less inference compute — a claim the company has made but that has not been independently verified.
• Beam is explicitly positioned as a Western answer to China’s open-weight models (DeepSeek, Qwen, GLM-5.2), at a moment when Bloomberg Intelligence puts the US-China top-model performance gap at roughly 3%, down from around 15% earlier in 2026.
• Reflection has raised roughly $4.7 billion since its 2024 founding, with its latest round valuing the company at $25 billion pre-money; backers include Nvidia, Sequoia Capital and Lightspeed Venture Partners.
• No model license has been published yet, no independent benchmark has confirmed Reflection’s numbers, and the weights themselves are not yet downloadable — all genuinely open questions as of this writing.

Key numbers

What Reflection AI actually announced

Reflection AI, founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou, used an announcement on 5 October 2026 to introduce Beam: a text-only mixture-of-experts (MoE) model with 501 billion total parameters, of which only 23 billion are active for any given token. That ratio is the entire point of an MoE design — most of the network sits idle for a given query, which is how Reflection can claim strong performance without the inference cost of a dense model that size.

Beam was trained on 23.8 trillion tokens and ships with a one-million-token context window, putting it in the same long-context bracket as the current generation of frontier systems. Reflection’s own benchmark claim is that Beam performs on par with Zhipu’s GLM-5.2 — a Chinese open-weight model with roughly 744 billion total parameters and 40 billion active — on advanced reasoning tasks, while using three to four times less inference compute to get there. The company also claims Beam outscores Inkling, the open-weight model from Mira Murati’s Thinking Machines Lab released in July 2026, on four coding benchmarks where both have published results, though Inkling is multimodal and Beam is text-only, so the comparison is not perfectly apples-to-apples.

Crucially, none of this is confirmed by anyone outside Reflection. TechCrunch’s report on the launch states plainly that “Reflection’s performance claims haven’t been independently verified” — and at the time of writing, the weights themselves are not yet public, so no outside lab, researcher or developer has been able to run the benchmarks themselves.

Why this matters: the open-weight race has a Western entrant with real money behind it

For the past year, open-weight large language models have been dominated by Chinese labs — DeepSeek, Alibaba’s Qwen family, and Zhipu’s GLM series — releasing frontier-class systems under permissive terms while the leading Western labs kept their best models closed. That dynamic has shaped procurement decisions across the industry: developers who want to self-host, fine-tune or avoid per-token API billing have increasingly had only Chinese options at the frontier. Alibaba’s own infrastructure buildout underscores how seriously Beijing is backing this strategy — the company is betting $53 billion on 20 gigawatts of new AI data centres to keep that open-model pipeline fed with compute.

Beam is Reflection’s attempt to give the West a comparable option. It’s not the first Western open-weight release — Meta’s Llama family and Mistral’s models have been in the market for years, and Thinking Machines Lab’s Inkling arrived in July — but Reflection is pitching Beam specifically as compute-efficient enough to compete with China’s best on cost as well as capability, which is a different and arguably more commercially relevant claim than simply matching benchmark scores.

The backdrop makes the timing notable. A Bloomberg Intelligence analysis circulating this week put the performance gap between DeepSeek’s V4.1 Flash and the leading closed US model (Anthropic’s) at roughly 3% on LiveBench, down from around 15% earlier in 2026 and about 9% in May. Whether or not that specific figure holds up, the direction of travel — a narrowing gap between US and Chinese models, open and closed alike — is the context every one of these announcements now sits inside. It’s also the backdrop against which labs are raising eye-watering sums: Anthropic’s own recent IPO filing, which combined a $42 billion loss with a $2 trillion valuation ask and an explicit existential-risk warning, is the clearest sign yet of how much capital is being staked on staying ahead.

The compute-efficiency claim is the real story, if it holds up

Raw benchmark parity is a commodity claim at this point — plenty of models claim to match GLM-5.2 or GPT-class systems on one leaderboard or another. What would actually matter, if verified, is the 3–4x inference-compute reduction. Inference cost, not training cost, is what dominates the economics of running a model in production at scale, and a genuine multiple-fold efficiency gain would make Beam commercially relevant even to developers who don’t particularly care about the US-China framing. This is exactly the kind of claim that was similarly open-weight-adjacent and compute-focused when OpenAI’s own open-weight release timeline slipped behind a paused frontier training run earlier this year — compute constraints are shaping release strategy across the industry, not just at Reflection.

The money behind Beam

Reflection’s war chest is substantial for a two-year-old company. It emerged from stealth in March 2025 with $130 million (a $25 million seed plus $105 million Series A), valuing it at roughly $545 million. By October 2025 it had raised $2 billion at an $8 billion valuation, with backers including Nvidia, Eric Schmidt, Citigroup, 1789 Capital, Lightspeed Venture Partners and Sequoia Capital. TechCrunch’s reporting on the Beam launch puts cumulative funding at roughly $4.7 billion, with the most recent round valuing the company at $25 billion pre-money — a roughly 45x increase in valuation in under two years.

Nvidia’s involvement is not incidental. The chipmaker has backed Reflection financially and, per TechCrunch, Reflection separately struck compute agreements worth a combined $7 billion-plus with SpaceX and Nebius this past summer, securing access to Nvidia’s GB300 chips through 2029 — reportedly including a SpaceX arrangement worth around $150 million a month. That pattern — a chip supplier becoming a financial backer and compute landlord to the labs that consume its hardware — echoes what’s happened elsewhere in the industry; Nvidia’s $35 billion Lambda arrangement with Anthropic was described in almost identical terms. It’s a structural feature of how frontier and near-frontier AI companies are now financed, and it raises the same conflict-of-interest questions each time: Nvidia has a direct financial stake in every lab it backs hitting its performance targets, which is worth remembering when reading any benchmark a backed company publishes about itself.

Practical takeaways for builders and publishers

• Don’t switch infrastructure on a press release. Beam’s weights aren’t public yet. Until independent benchmarks exist, any production decision should treat Reflection’s numbers as a hypothesis, not a spec sheet.
• Watch the license, not just the weights. “Open-weight” covers everything from fully permissive (Apache 2.0-style) to restrictive commercial terms. Reflection hasn’t published Beam’s license yet — that detail will determine whether this is usable in a commercial product at all.
• If compute cost is your bottleneck, this is worth tracking closely. A genuine 3–4x inference-compute reduction at frontier-adjacent quality would materially change self-hosting economics for anyone running high-volume inference, from chatbots to coding agents.
• Treat all comparative benchmarks — Reflection’s and everyone else’s — as self-reported until a third party reproduces them. This applies equally to GLM-5.2’s own claims, Inkling’s, and any other model cited in this piece; none of the cross-model comparisons here have independent confirmation.
• Factor in geopolitics, not just specs. Procurement teams avoiding Chinese-origin open models for compliance or supply-chain reasons now have a higher-profile Western alternative to evaluate once weights ship — but “once weights ship” is the operative caveat.

What we still don’t know

• Beam’s actual license terms — whether it’s genuinely permissive or carries commercial restrictions — have not been published.
• No independent party has reproduced Reflection’s benchmark scores or its 3–4x compute-efficiency claim; all of it is self-reported.
• The exact release date for the full weights within October 2026 hasn’t been specified beyond “this month.”
• How Beam performs against Alibaba’s Qwen3.8-Max or against Anthropic’s and OpenAI’s closed frontier models on standardised, third-party benchmarks is not yet known — Reflection’s own comparisons focus on GLM-5.2 and Inkling.
• Whether Reflection’s hyperscaler and neocloud distribution partners (named but not detailed) will host Beam at prices that actually deliver the claimed compute savings to end users, or whether that saving is theoretical at the model-weights level only.
• No Reflection executive is quoted by name in reporting on the launch; the company reportedly did not respond to TechCrunch’s requests for comment in time for publication.

FAQ

What is Beam, in plain terms?

It’s a large AI language model that Reflection AI built and plans to release publicly, including its underlying weights, so other companies and developers can run and adapt it themselves rather than only access it through Reflection’s own paid service.

Is Beam actually better than China’s open models?

Reflection says Beam matches GLM-5.2 on reasoning while using significantly less compute, but that claim is self-reported and hasn’t been independently tested. Until outside researchers can run the weights themselves, it’s a claim, not a confirmed result.

When can I download or use Beam?

Reflection says full weights and technical documentation will ship “this month” (October 2026), distributed through hyperscalers, neoclouds and open-source library integrations. No exact date has been given.

Who is paying for all this?

Reflection has raised roughly $4.7 billion since 2024, most recently at a $25 billion pre-money valuation, from investors including Nvidia, Sequoia Capital and Lightspeed Venture Partners, alongside separate multi-billion-dollar compute agreements with SpaceX and Nebius.

Sources

• TechCrunch: “Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost” (5 October 2026)
• Wikipedia: Reflection AI — company background and funding history
• MarkTechPost: “Reflection AI Introduces Beam: A 501B Open-Weight MoE Model With 23B Active Parameters” (5 October 2026)
• Our earlier coverage: Alibaba’s New Zhenwu V900 Chip and Its $53bn Bet on 20GW of AI Data Centres
• Our earlier coverage: Anthropic’s IPO Filing: A $42bn Loss, a $2 Trillion Ask, and an Existential-Risk Warning
• Our earlier coverage: OpenAI Paused Its Biggest AI Training Run Over Cyber Risk — Open Weights Are Two Weeks Behind
• Our earlier coverage: Anthropic’s $35bn Lambda Deal: Nvidia Becomes AI’s Landlord, Not Just Its Chip Supplier
• Implicator.ai: “DeepSeek Narrows US AI Benchmark Lead to 3% After September Release”, citing Bloomberg Intelligence analyst Robert Lea (4 October 2026)
• Axios: “Open-source AI gets more compute from SpaceX” (22 June 2026)

Please follow and like us: