A New York startup founded by two former Google DeepMind researchers has delivered the West’s answer to China’s open-source AI dominance, and the timing could not be more pointed. Axios reported on October 4, 2026 that Reflection AI, backed by Nvidia and a roster of marquee investors, was readying its first open-weight foundation model, a system explicitly designed to challenge Chinese labs like DeepSeek and Alibaba’s Qwen for control of the open-model ecosystem that much of the world’s AI tooling now runs on. The company followed through, unveiling the model as Beam on October 5, 2026 and releasing its weights on October 16, 2026 under the name Beam 1.0, a 501 billion-parameter system under the Apache 2.0 license that confirmed the launch window reporters had anticipated.
The announcement lands at a moment when the gap between Chinese and Western open-weight models has become one of the most-discussed fault lines in AI strategy circles. For more than a year, developers building everything from chatbots to coding agents have reached for DeepSeek and Qwen releases not because of national preference but because, until now, nobody in the US or Europe has shipped anything as capable and as free to use. Reflection AI wants to close that gap, and it has raised billions of dollars to try.
Don't miss new tech stories on Google
Add Tech Insider once in the Google app and our stories appear in your news suggestions.
What Reflection AI actually announced
Reflection AI unveiled Beam, its first open-weight model, on October 5, 2026, and released its weights on October 16, in an effort to pressure Chinese open-weight developers. TechCrunch had already flagged the company’s ambitions earlier, when it covered the funding round that made the project possible. Reflection presented Beam as a U.S. open-weight competitor to Chinese models including Z.ai’s GLM-5.2, particularly for coding and agentic workloads. It said Beam’s results are comparable to GLM-5.2 on advanced reasoning benchmarks, though complete technical benchmark details had not yet been published.
What the reporting did not confirm in early October is now far clearer. Reflection AI unveiled Beam on October 5, 2026 and released its weights on October 16, 2026 under the name Beam 1.0 and the Apache 2.0 license, a sparse mixture-of-experts system with 501 billion total parameters, according to BenchLM, though only about 23 billion of those are active per token, and the model ships with a 1 million-token context window, per OpenRouter’s October 2026 coverage, alongside BF16, FP8, and NVFP4 checkpoints. Multiple outlets, including BigGo and Startup Fortune, had already noted that other Western AI companies were expected to ship their own open-weight models around the same window, and Beam 1.0’s arrival suggests October 2026 is indeed becoming that turning point for open AI in the US and Europe simultaneously.
Reflection AI’s strategy has been consistent for more than a year. Co-founder and CEO Misha Laskin told TechCrunch that “today the best open models are coming out of China,” a blunt framing of the problem the company says it exists to solve. Laskin has also been more pointed about the practical consequences of that gap. In comments reported by Gizmodo, he said that “global AI adoption is actually very much being built on top of Chinese models, including [by] American companies.”
Who is behind Reflection AI
Reflection AI was founded in 2024 by Misha Laskin and Ioannis Antonoglou, both exits from Google DeepMind. Laskin, who holds a PhD in quantum physics, reportedly led reward-modeling work tied to Gemini before leaving to start the company. Antonoglou’s pedigree runs deeper into DeepMind lore: he joined the lab in 2012, reportedly as its 25th employee, and contributed to AlphaGo, AlphaZero, and MuZero, the reinforcement-learning systems that first made DeepMind a household name in AI research circles. Reuters has described Antonoglou and Laskin as former DeepMind researchers who left to build what they call a Western counterweight to Chinese open models.
The company’s first publicly released model is Beam 1.0, an open-weight model whose weights came out on October 16, 2026, alongside the Reflection API and Mirror CLI (beta). That release gives Reflection AI something developers can download and run themselves, rather than a pitch built only on funding. The strategic bet, as Antonoglou described it in comments reported by Forbes, is straightforward: ensure there is a competitive frontier open lab in the Western world.
The money behind the mission
Reflection AI’s funding trajectory is itself a story about how seriously investors are taking the open-model gap. The company first raised $130 million in March 2025, a $25 million seed round combined with a $105 million Series A, which valued it at roughly $545 million, according to Wikipedia. Reuters reported that the company then raised a $2 billion Series B in October 2025, valuing it at roughly $8 billion, a jump TechCrunch pegged at 15 times that March 2025 valuation in barely six months. Nvidia was the anchor investor in that October round, reportedly contributing around $800 million, alongside participation from Eric Schmidt, Citigroup, 1789 Capital, Lightspeed, and Sequoia Capital, according to reporting summarized by BigGo and Startup Fortune.
The valuation did not sit still. Startup Fortune reported that by June 2026, Reflection AI had completed additional financing at an approximately $25 billion pre-money valuation. A separate account from Sacra puts a March 2026 round at $2.5 billion and a resulting $27.5 billion valuation, a figure that does not perfectly reconcile with the Startup Fortune number. Given the discrepancy, both figures should be read as reported estimates rather than confirmed, audited valuations. What is consistent across every account is the direction of travel: in under a year, Reflection AI’s valuation went from roughly $8 billion to somewhere in the $25 billion to $27.5 billion range.
Forbes also reported in July 2026 that Reflection AI signed a $1 billion compute deal, with Antonoglou telling the outlet he expected an open-source model to ship later in the year, a plan that traced back to October 2025, when TechCrunch reported the company was already targeting a 2026 frontier model trained on tens of trillions of tokens, with the model’s weights made public while its underlying datasets and training pipelines stayed proprietary. That timeline lines up with the October reporting from Axios and the subsequent Beam 1.0 release, suggesting the compute deal and the model launch were two parts of the same plan rather than separate initiatives.
| Funding Round | Date Reported | Amount | Valuation | Source |
|---|---|---|---|---|
| Series B | October 2025 | $2 billion | ~$8 billion | Reuters |
| Nvidia investment (part of Series B) | October 2025 | ~$800 million | Included above | BigGo / Startup Fortune |
| Compute deal | July 2026 | $1 billion | Not applicable | Forbes |
| Follow-on round (disputed figure) | June 2026 | $2.5 billion (Sacra) / unspecified (Startup Fortune) | $27.5 billion (Sacra) / ~$25 billion pre-money (Startup Fortune) | Sacra / Startup Fortune |
Why the China open-model gap matters now
To understand why this launch is getting so much attention, it helps to look at how quickly Chinese labs seized the open-weight lead. DeepSeek’s releases, whose open weights are hosted on Hugging Face, forced a wave of reassessment across the industry about how much raw capital US labs actually need to spend to stay competitive, and Alibaba’s Qwen family has become a default choice for developers who want a capable, freely downloadable model without a usage fee. Moonshot AI’s Kimi line and Zhipu’s GLM models, including Z.ai’s GLM-5.3, released in August 2026, and the GLM-5.2 that Reflection named when positioning Beam, added further competition from China, each pushing open benchmarks forward while US labs mostly kept their best systems behind an API paywall.
Laskin’s framing, delivered in an interview with TechJournal UK, captures the anxiety driving Reflection AI’s entire business plan: today there is a very large gap between open-model capabilities in the West and the ones in China. That gap has real downstream consequences. Enterprises in regulated industries, governments wary of data residency tied to foreign-controlled models, and developers who simply want to self-host without relying on a closed API have had limited high-quality Western alternatives. Reflection AI’s bet is that this unmet demand is large enough to justify a multibillion-dollar valuation for a company that, until it released Beam, its first open-weight model, on October 16, 2026, had generated most of its public attention through funding announcements rather than product benchmarks.
Laskin has also framed the company’s ambitions in product terms, telling TechJournal UK plainly that “we’re building frontier Western open models,” a statement that positions Reflection AI not as a fast-follower but as a lab aiming for the top tier of open releases, not just a credible alternative.
How Reflection AI compares to the Chinese open-weight field
Because Reflection AI had not published complete technical benchmark details when it unveiled Beam, any head-to-head performance comparison with DeepSeek, Qwen, Kimi, or GLM would still be speculative, even though the company said Beam achieved results comparable to GLM-5.2 on advanced reasoning benchmarks. For a sense of how fast this field moves, recent testing already pits GPT-6 Astra against Gemini 4 Argon and DeepSeek, and Moonshot AI’s own Kimi line recently landed on Amazon Bedrock as a 1.56TB release, underlining how crowded the open and semi-open field already is even before Reflection AI’s Beam arrived. What can be compared, based on available reporting, is market position, funding, and strategic intent.
| Lab | Region | Known Model Family | Reported Funding / Backing | Reflection AI Positioning |
|---|---|---|---|---|
| Reflection AI | United States | Beam 1.0 (open weights, Apache 2.0) | $2B+ raised, Nvidia-backed, ~$25B+ valuation reported | Aims to match leading Chinese open-weight systems first |
| DeepSeek | China | DeepSeek open-weight releases | Not disclosed in this reporting | Named repeatedly as the benchmark Reflection is chasing |
| Qwen (Alibaba) | China | Qwen open-weight releases | Backed by Alibaba’s broader cloud and AI budget | Named alongside DeepSeek as the open-weight leader Reflection targets |
| Kimi (Moonshot AI) | China | Kimi model family | Not addressed in current Reflection AI reporting | Not directly compared in available reporting |
| GLM (Zhipu AI) | China | GLM model family | Not addressed in current Reflection AI reporting | Not directly compared in available reporting |
That last column matters. The reporting available right now supports a genuine rivalry narrative between Reflection AI and DeepSeek or Qwen specifically, but it does not support claims that Reflection AI’s Beam already outperforms any of these systems: the GLM-5.2 comparison is a company claim made on October 5, 2026, and the available reporting says complete technical benchmark details had not yet been published. Readers should treat “American DeepSeek” framing, which has circulated in some secondary coverage, as a media characterization of intent rather than a confirmed technical result.
Nvidia’s stake in the outcome
Nvidia’s roughly $800 million contribution to Reflection AI’s Series B is not an isolated bet. The chipmaker has spent much of 2026 funding infrastructure and software plays that keep demand for its GPUs growing beyond the current wave of frontier-model training runs, a strategy visible in its other recent moves on compute-hungry AI tooling. A credible Western open-weight lab gives Nvidia another reason for enterprises to buy GPUs for self-hosted inference rather than renting capacity exclusively through closed-model APIs, reinforcing the hardware demand story that underpins Nvidia’s valuation regardless of which specific lab wins the open-model race.
That dynamic also explains why Nvidia has been willing to back a company whose flagship product, prior to this announcement, was a code-comprehension tool rather than a foundation model. The Asimov product generates revenue and a user base today; the open-weight model, now that it has shipped, generates the kind of developer mindshare and self-hosted deployment volume that drives long-term chip demand. Nvidia’s DGX Spark hardware line, which launched its 64GB configuration at $4,999 amid a broader memory-price surge, stands to benefit directly if Reflection AI’s open model becomes a popular choice for on-premises deployment. Reports dated October 22, 2026 say Nvidia is in early discussions to invest further in, or acquire, Reflection AI, though the acquisition possibility remains only a rumor.
Market impact: why investors are paying attention
The jump from an $8 billion valuation in October 2025 to a reported $25 billion-plus by the following spring reflects more than normal AI-market froth. It signals that late-stage investors believe the open-weight segment of the AI market, long treated as a secondary concern next to frontier closed models from OpenAI, Anthropic, and Google, is becoming a venue where real enterprise dollars will move. Enterprises that want to avoid vendor lock-in, meet data-sovereignty requirements, or simply cut per-token costs have strong incentives to adopt open-weight models they can run on their own infrastructure, and right now a disproportionate share of the best options available to them come from Chinese labs.
With Beam 1.0 released as open weights under the Apache 2.0 license on October 16, 2026, Reflection AI could reshape vendor selection conversations at large enterprises that have been hesitant to build core workflows on top of Chinese-origin weights for compliance or geopolitical reasons, even when those models test well on public benchmarks. That hesitancy has been a persistent, if rarely quantified, tailwind for any credible Western open alternative, and it is likely a meaningful part of why investors assigned Reflection AI a reported $25 billion pre-money valuation.
Historical context: how we got here
The open-weight AI race did not start with Reflection AI, and it will not end with it either. Meta’s Llama family was the first widely adopted Western open-weight effort at real scale, giving developers an alternative to closed APIs years before Chinese labs became dominant in the category. But as Chinese labs iterated faster on efficiency and cost, and as some Western labs pulled back from fully open releases in favor of tightly licensed or partially open models, the center of gravity in open AI shifted east. DeepSeek’s emergence in particular forced a broad re-examination in Western boardrooms of how much capital advantage actually matters when a smaller, leaner team can match frontier-adjacent performance at a fraction of the training cost.
Reflection AI’s own history fits neatly into that arc. Laskin and Antonoglou left DeepMind in 2024, a period when DeepMind and OpenAI were both doubling down on closed, commercialized frontier models rather than open releases. Their decision to found a company explicitly built around open weights was, even at the time, a bet against the prevailing strategy of the labs they came from. The $2 billion raised in October 2025 at an $8 billion valuation suggests investors were willing to back that contrarian bet well before any product existed to prove it out, and the subsequent valuation jump suggests that confidence has only grown as Chinese open models kept winning developer mindshare throughout 2025 and into 2026.
What remains unconfirmed
It is worth being explicit about the boundaries of what had actually been reported before launch, given how much speculation tends to accumulate around a release like this one. As of October 4, 2026, the following details had not yet been confirmed by Reflection AI or verified in primary reporting: the model’s official name, its exact parameter count, its underlying architecture, the composition of its training data, its license terms, a published model card, and any independently verified benchmark scores, and the company had not locked in a firm release date beyond describing the window as “soon” or “later in October.” That changed on October 5, 2026, when Reflection AI unveiled Beam, and on October 16, 2026, when it released the weights as Beam 1.0 under the Apache 2.0 license, disclosing 501 billion total parameters, 23 billion active parameters per token, and a 1 million-token context window, though independently verified head-to-head benchmark scores against DeepSeek and Qwen remain limited.
What has been reported with stronger sourcing includes the company’s founding year and founders, the leadership team’s DeepMind backgrounds, the existence and function of the Asimov product, the October 2025 funding round and its $8 billion valuation, Nvidia’s role as an anchor investor, and the general strategic framing that Reflection AI is building an open-weight model meant to compete with Chinese labs. Treat any claim beyond that list, including direct performance comparisons with specific DeepSeek or Qwen model versions, as unverified until Reflection AI publishes its own technical documentation.
Competitive comparison: the broader open-model landscape
Reflection AI is not entering an empty field, and its success or failure will be measured against a crowded set of existing options that developers already use every day, a landscape that extends well beyond the Nvidia-and-AMD duopoly on the hardware side and includes head-to-head model battles like Claude Sonnet 5.5 against GPT-6 Luna and Gemini 3.8 on the model side. Alibaba’s Qwen family has broad adoption because it spans a wide range of model sizes, which makes it workable for both lightweight and heavy inference jobs. DeepSeek built its reputation on cost efficiency relative to apparent capability, a narrative that moved markets when it first surfaced and that still shapes how investors value AI training efficiency today. Meta’s Llama models retain a loyal developer base built over several release generations, giving them an installed-base advantage that a brand-new entrant like Reflection AI will need time to match, regardless of how its first model actually performs on benchmarks.
Reflection AI’s advantage, if it has one, is not a head start on model quality, since Beam is the company’s first model. Its advantage is capital, a founding team with direct frontier-lab research pedigree, and a narrow, well-articulated mission that has proven easy for investors like Sequoia, Lightspeed, and Nvidia to rally behind. Whether that translates into a model developers actually choose over Qwen or DeepSeek will depend on independent benchmarks, since Reflection has not yet published complete technical benchmark details beyond saying Beam is comparable to GLM-5.2 on advanced reasoning.
What developers should watch for next
For engineering teams evaluating whether to adopt Beam or continue building on existing open-weight options, the practical questions center on hardware requirements and real-world benchmark performance, since the Apache 2.0 license terms are already settled. A model that requires enterprise-grade GPU clusters to run at reasonable latency will have a very different adoption curve than one tuned to run efficiently on consumer or prosumer hardware, the kind of distinction that has already shaped adoption patterns for other recent open releases in the market.
Licensing will matter just as much as raw capability. Chinese labs have generally released their flagship open-weight models under permissive licenses that allow commercial use with minimal restriction, a major factor in their rapid enterprise adoption. Beam’s Apache 2.0 license matches that permissive baseline, so the question is now whether benchmark results are strong enough to pull developers away from already-entrenched Chinese alternatives like the ones already tracked in pricing comparisons against MiniMax M3.1 and DeepSeek V4.1 Flash, or from leaderboard standings tallied across LMArena and Artificial Analysis.
How Beam 1.0 Is Being Distributed
Beam is a text-in/text-out sparse mixture-of-experts model with 501 billion total parameters and 23 billion active parameters, a 1-million-token context window, and BF16, FP8, and NVFP4 checkpoints. Reflection announced it on October 5, 2026 as its first frontier open-weight model, targeting coding, reasoning, software maintenance, tool use, and long-horizon agentic tasks. Before the October 16 release, Beam was in selective preview and final red-teaming, with early access through Reflection’s API.
On October 16, 2026, Reflection released Beam’s weights on Hugging Face under the Apache 2.0 license. Public availability expanded from hosted preview access to downloadable BF16, FP8, and NVFP4 checkpoints, with a tokenizer, reference implementation, and developer harness. Beam also became available through launch partners and cloud platforms, including CoreWeave’s serverless inference offering, where it can be accessed through an OpenAI-compatible API. Reflection’s own page states: “Beam 1.0 is out, with open weights, the Reflection API, and Mirror CLI (beta).”
For teams deciding how to try the model, the practical choice is between downloading the checkpoints and calling a hosted endpoint. Apache 2.0 weights do not by themselves make inference free, because hosting, API, GPU, or cloud charges may apply, so check each provider’s own pricing page before committing. On quality, Reflection said Beam achieved results comparable to GLM-5.2 on advanced reasoning benchmarks, but that remains a company claim, so run your own evaluations on your coding and agent workloads before switching anything in production.
Predictions: where this goes from here
Based on the trajectory visible in current reporting, several outcomes look plausible over the next six to twelve months. First, the near-term call has already played out: Reflection AI unveiled Beam on October 5, 2026 and released its weights on October 16, 2026, in line with the outlets that tied the launch to the company’s recent compute investments. Second, the company led with transparency around licensing terms, publishing the weights under Apache 2.0, since permissive licensing is the single most direct lever Reflection AI has to win developers away from Qwen and DeepSeek quickly.
Third, expect other Western AI companies to accelerate their own open-weight announcements in the same window, following the pattern BigGo and Startup Fortune both flagged, which would turn October 2026 into a genuine inflection point for Western open AI rather than a single-company story. Fourth, expect Nvidia to continue positioning its hardware, including the DGX Spark line, around whichever open-weight models gain the most developer traction, regardless of country of origin, since the company’s core incentive is GPU demand rather than any particular model’s success. Fifth, expect scrutiny of Reflection AI’s valuation to intensify now that its model has shipped, since a $25 billion-plus valuation for a company whose main public product before this launch was a code-comprehension tool will face direct comparison against its first foundation model’s real-world reception.
The bigger picture for Western AI strategy
Reflection AI’s launch of Beam is being read by much of the industry as a test case for whether capital and research pedigree alone can close a technical and adoption gap that Chinese labs built over roughly two years of rapid, cost-efficient releases. The company’s founders have research credentials that few startups can match, Antonoglou’s AlphaGo-era DeepMind tenure chief among them, but credentials do not automatically translate into a model developers choose to adopt over free, proven alternatives already embedded in their workflows.
What makes this moment different from earlier attempts to challenge Chinese open-weight dominance is the scale of capital now backing the effort and the direct, repeated public framing from Reflection AI’s own leadership that this is a geopolitical and commercial priority, not just a product launch. Laskin’s description of global AI adoption already running on top of Chinese models, as reported by Gizmodo, frames the stakes in terms that go well beyond a single company’s product roadmap. Whether Beam proves capable enough to change that picture will become clear as independent benchmark results arrive, since Reflection had not yet published complete technical benchmark details when it announced the model.
Frequently Asked Questions
What is Reflection AI?
Reflection AI is a New York-based startup founded in 2024 by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou. The company’s public product is Asimov, a code-comprehension agent, and it has released its first open-weight foundation model, Beam.
Has Reflection AI’s open-weight model been released yet?
Yes. Reflection AI introduced Beam on October 5, 2026 and released the weights, under the name Beam 1.0 and the Apache 2.0 license, on October 16, 2026, disclosing 501 billion total parameters, 23 billion active parameters per token, and a 1 million-token context window, according to BenchLM and OpenRouter.
How much funding has Reflection AI raised?
Reflection AI first raised $130 million in March 2025, a $25 million seed round plus a $105 million Series A, at roughly a $545 million valuation, per Wikipedia. Reuters then reported a $2 billion Series B in October 2025 at an $8 billion valuation, about 15 times that March figure, with Nvidia contributing roughly $800 million. Later reporting from Startup Fortune and Sacra describes a follow-on round in 2026 that pushed the valuation to somewhere between $25 billion and $27.5 billion, though the exact figures differ between sources.
Why is Reflection AI’s model being compared to DeepSeek and Qwen?
Reflection AI’s leadership has repeatedly framed the company’s mission around closing the gap between Chinese and Western open-weight model capability, specifically naming DeepSeek and Alibaba’s Qwen as the systems it aims to match or beat.
Who founded Reflection AI, and what is their background?
Misha Laskin, who holds a PhD in quantum physics, and Ioannis Antonoglou, an early Google DeepMind employee who contributed to AlphaGo, AlphaZero, and MuZero, co-founded the company in 2024 after leaving DeepMind.
Is Nvidia an investor in Reflection AI?
Yes. Nvidia reportedly led or anchored Reflection AI’s October 2025 Series B with an investment of roughly $800 million, according to reporting from BigGo and Startup Fortune.
What other investors have backed Reflection AI?
Reported backers include Eric Schmidt, Citigroup, 1789 Capital, Lightspeed Venture Partners, and Sequoia Capital, with additional investor listings naming CRV, Disruptive Technology Advisers, Shaw and Partners, and SV Angel.
Will Reflection AI’s model outperform DeepSeek or Qwen?
That is not yet known. Reflection said Beam achieved results comparable to GLM-5.2 on advanced reasoning benchmarks, but that is a company claim and complete technical benchmark details had not been published, so any claim that it already beats specific Chinese open-weight models should be treated as speculation until independent results arrive.
