Reflection AI pulled back the curtain on October 5, 2026, with the debut of Beam, its first open-weight AI model and the clearest answer yet to a question that has nagged US policymakers and enterprise buyers all year: can an American lab build an open model that competes with China’s on both performance and price? According to TechCrunch, Reflection AI Beam is a sparse mixture-of-experts system built for coding, reasoning, and agentic workloads, and the company says it matches top Chinese open models while using a fraction of the compute per query.
The timing is not an accident. Chinese labs including DeepSeek, Alibaba’s Qwen team, Moonshot AI (Kimi), MiniMax, and Z.ai (GLM) have spent 2025 and 2026 releasing open-weight models that top independent leaderboards and undercut US rivals on token price. Reflection AI Beam is Brooklyn, New York-based Reflection’s attempt to flip that script, and it arrives backed by a company valuation that has climbed fast enough to turn heads on its own.
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 Exactly Is Beam? Architecture and Training Breakdown
Beam is a sparse mixture-of-experts (MoE) model with 501 billion total parameters, of which only 23 billion are active for any given token, according to Reflection’s own launch blog post. That architecture, explained in more detail on Wikipedia’s mixture-of-experts entry, routes each input through only a subset of the model’s “expert” sub-networks rather than activating the full parameter count, which is exactly why Reflection AI Beam can claim frontier-class capability without frontier-class inference bills. Artificial Analysis also logged Beam launching with a 1 million-token context window, and Reflection’s developer documentation for the Beam-1.0-501B-A23B model lists that same 1-million-token window alongside a maximum documented output of 128,000 tokens and a June 30, 2026 knowledge cutoff, specs that put it in range of long-document and long-session agentic work that shorter-context open models struggle to handle in a single pass.
Reflection AI’s blog post states plainly: “Beam is a sparse Mixture-of-Experts model with 501 billion total parameters, 23 billion active, built for coding, reasoning, and agentic workloads.” The company says Beam was pretrained on 23.8 trillion tokens pulled from curated web data and proprietary licensed datasets. Training infrastructure details reported by Startup Fortune put the base-model pretraining run on 6,144 Nvidia GB300 NVL72 GPUs, completed in under four weeks at a reported 92.3% “goodput” rate, the share of GPU time spent on useful computation rather than restarts and failures.
A separate reinforcement-learning phase reportedly used 10,500 Nvidia GB300 GPUs over four weeks and generated more than 100 million rollouts, a post-training figure also tallied by AI Wiki in its October 2026 summary of Reflection’s training disclosures. Reflection credits that heavy RL investment, not sheer parameter count, as the real source of Beam’s efficiency: training the model to reach correct answers in fewer reasoning steps rather than brute-forcing its way there.
The Benchmark Numbers Reflection Is Touting
Reflection has published a specific set of self-reported scores for Beam: 80.9 on SWE-Bench Verified, 80.1 on Terminal-Bench 2.1, and 97.8 on AIME 2026. TechCrunch notes that these figures come from Reflection itself and had not been independently verified at launch, a standard caveat for any lab debuting a new model before third-party evaluators get hands-on access.
For context, Reflection positions Beam as roughly matching Z.ai’s GLM-5.2 on advanced reasoning tasks. One tracked comparison puts GLM-5.2 slightly ahead on Terminal-Bench 2.1, at 81.0 versus Beam’s 80.1, but Reflection argues the real story is efficiency: the company claims Beam needs three to four times less inference compute than GLM-5.2 to reach that comparable score. That is the heart of the Reflection AI Beam pitch, not “we beat everyone outright,” but “we get you 90-plus percent of the performance at a quarter of the cost.”
How Beam Stacks Up Against China’s Open-Weight Leaders
Reflection frames Beam as a direct alternative to DeepSeek, Qwen, Kimi K2, MiniMax, and GLM, the five Chinese model families that have dominated open-weight downloads over the past year. Public reporting shows Reflection positioning Beam as nearing Alibaba’s Qwen3.8-Max on some tasks, though a full side-by-side scorecard against DeepSeek’s latest release, Kimi K2, or MiniMax has not been published by any outlet covering the launch. The strongest documented comparison remains Beam versus GLM-5.2, where the efficiency gap, not the raw score gap, is Reflection’s central argument.
That nuance matters for anyone evaluating whether to trust the headline claim. Readers comparing frontier systems more broadly can see how that benchmark gap plays out across closed models in our breakdown of GPT-6 Astra vs Gemini 4 Argon vs DeepSeek, which shows just how fluid the leaderboard has become month to month.
Beam vs. Leading Open-Weight Models at a Glance
| Model | Developer | Total / Active Params | Terminal-Bench 2.1 | License Status |
|---|---|---|---|---|
| Beam | Reflection AI | 501B / 23B (MoE) | 80.1 (self-reported) | Apache 2.0, weights released |
| GLM-5.2 | Z.ai | Not disclosed in launch coverage | 81.0 (tracked comparison) | Open-weight, released |
| Qwen3.8-Max | Alibaba | Not disclosed in launch coverage | Not published in comparison | Open-weight, released |
| Kimi K2 | Moonshot AI | Not disclosed in launch coverage | Not published in comparison | Open-weight, released |
| MiniMax (current line) | MiniMax | Not disclosed in launch coverage | Not published in comparison | Open-weight, released |
The gaps in that table are deliberate. Independent, apples-to-apples numbers for Beam against DeepSeek, Qwen, Kimi K2, and MiniMax simply were not available in launch-day reporting, and inventing them would defeat the purpose of a fact-checked comparison. Readers tracking the broader Chinese open-model price war can find more documented figures in our look at MiniMax M3.1 vs GPT-6 Luna vs DeepSeek V4.1 Flash.
Why “3-4x Less Compute” Matters More Than Raw Scores
Inference cost, not training cost, is what actually shows up on a company’s cloud bill once a model ships to production. A model that scores a point or two lower on a benchmark but runs at a quarter of the token cost can still win the deployment decision for any enterprise running millions of daily agent calls. That is the bet underlying Reflection AI Beam: efficiency as the product, not just a side benefit.
Reflection’s own framing captures the pitch directly. “Beam combines strong agentic performance, efficient reasoning, and a 500B form factor to give enterprises, governments, and developers a true workhorse open model,” the company wrote in its launch thread on X. It is a pitch aimed squarely at the budget-conscious enterprise buyer who has watched Chinese labs undercut Western API pricing for the better part of two years.
Apache 2.0 License and the Waitlist-First Rollout
Beam’s rollout moved fast once red-teaming wrapped: an early version was first offered to a select group of users through a waitlist, and SiliconANGLE reported that Reflection has since released the full weights under an Apache 2.0 license in October 2026. The technical report is still listed as coming. The Apache 2.0 license permits commercial use, modification, and redistribution.
# License header for Beam's released weights
# SPDX-License-Identifier: Apache-2.0
# Full text: https://www.apache.org/licenses/LICENSE-2.0The staged rollout, early access first, open weights second, mirrors how several labs have handled high-stakes launches this year, giving Reflection a window to tune safety behavior before the broader developer community starts stress-testing Beam in production.
Inside Reflection AI: Founders, Funding, and the Valuation Climb
Reflection AI was founded in March 2024 by Misha Laskin, now the company’s CEO, and Ioannis Antonoglou, its president and chief technology officer. Both spent years at Google DeepMind before starting the company; Antonoglou is credited as a co-creator of AlphaGo. That DeepMind pedigree helped Reflection raise money almost immediately: the company emerged from stealth in March 2025 with $130 million in hand, split between a $25 million seed round and a $105 million Series A, according to the company’s own financing disclosures and Wikipedia’s record of the raise. From there, Reflection kept raising money at a pace few two-year-old startups match.
According to Reuters, Reflection raised $2 billion in October 2025 at an $8 billion valuation, with participation from Nvidia, former Google CEO Eric Schmidt, Citi, 1789 Capital, Lightspeed, and Sequoia. By spring 2026, reports citing CEO Misha Laskin put the company’s valuation closer to $25 billion, reflecting how quickly investor appetite for US-based open-weight alternatives has grown since Chinese labs began dominating open-model downloads.
Reflection AI Funding and Compute Timeline
| Date | Event | Figure | Source |
|---|---|---|---|
| March 2024 | Company founded by Misha Laskin and Ioannis Antonoglou | – | Startup Fortune |
| October 2025 | Funding round led with Nvidia, Sequoia, Lightspeed participation | $2B raised at $8B valuation | Reuters |
| June 2026 | Compute deal with SpaceX for GB300 access at Colossus 2 | $6.3B | CNBC (via Startup Fortune) |
| July 2026 | Compute deal with Nebius for GB300 access | $1B | Bloomberg (via Startup Fortune) |
| Spring 2026 | Valuation reported by CEO Misha Laskin | ~$25B | CNBC |
| October 5, 2026 | Beam unveiled; early access waitlist opens | 501B / 23B param MoE | TechCrunch, Reflection AI |
Those compute deals, worth more than $7 billion combined, secure Reflection access to Nvidia’s GB300 chips through 2029, locking in the hardware pipeline needed to both train successors to Beam and serve inference at scale now that the model has been released with public weights. For more on how that chip demand ripples through Nvidia’s broader roadmap, see our coverage of the Nvidia DGX Spark 64GB launch amid the current memory price surge.
The $7 Billion Compute Bet Behind Beam
It is worth sitting with how unusual this is for a company barely two and a half years old. Reflection signed a $6.3 billion compute agreement with SpaceX in June 2026 for access to Nvidia GB300 chips at SpaceX’s Colossus 2 data center, followed by a $1 billion cloud-compute deal with Nebius in July. Combined, those contracts run through 2029 and effectively guarantee Reflection the raw silicon needed to keep iterating past Beam without getting stuck in the GPU queue that has slowed smaller labs.
That kind of multi-year compute lock-in has become the real moat in frontier AI, arguably more decisive than any single benchmark score. Companies like AMD have made similar strategic bets this year, as covered in our piece on AMD’s $8.2 billion World Labs deal, underscoring how compute access, not just model architecture, now decides which labs stay competitive.
Historical Context: How the Open-Weight Race Got Here
Open-weight models were, for years, a secondary story in AI, with closed systems from OpenAI, Anthropic, and Google getting the headlines while Meta’s Llama family and a handful of European labs like Mistral held the open lane. That changed when DeepSeek’s early 2025 release demonstrated that a Chinese lab could ship an open model competitive with closed frontier systems at a sliver of the reported training cost, triggering a wave of follow-on releases from Qwen, Kimi, MiniMax, and GLM throughout 2025 and into 2026.
By mid-2026, independent trackers were showing Chinese labs holding a meaningful share of top spots on open-weight leaderboards, a dynamic we detailed in our earlier report on Reflection AI’s $25 billion bet to take on China’s open models. Beam is the practical output of that bet: not just capital raised, but an actual model shipped to challenge the assumption that the cheapest, most efficient open-weight systems would keep coming out of China by default.
Market Impact: What Beam Means for Nvidia, OpenAI, and Cloud Providers
For Nvidia, which has backed Reflection as an investor, Beam is a validation story as much as a product launch: another well-funded customer locking in GB300 demand for years, at a moment when GPU supply remains the tightest constraint in the industry. For OpenAI, Anthropic, and Google, the pressure is more indirect. None of those companies compete head-to-head with an open-weight model the way Beam does with GLM-5.2, but every efficient open alternative chips away at the argument that frontier capability requires a closed, metered API.
Cloud providers and enterprise buyers stand to benefit most directly. A credible, US-domiciled, Apache 2.0-licensed model that claims GLM-5.2-class reasoning at a fraction of the inference cost gives procurement teams a domestic option they can deploy without the data-sovereignty and export-control questions that come with running a Chinese open model in regulated environments. That is precisely the audience Reflection name-checked in its own launch materials: enterprises, governments, and developers.
What Reflection AI Is Saying About Beam’s Launch
“Introducing Beam: a highly efficient agentic open model with 501B total parameters and 23B active.”
Reflection AI, company account, via X
“Beam combines strong agentic performance, efficient reasoning, and a 500B form factor to give enterprises, governments, and developers a true workhorse open model.”
Reflection AI, company account, via X
“It’s our first step toward building frontier intelligence that is open and widely accessible to all.”
Reflection AI, company account, via X
“We are introducing Beam, Reflection’s first open-weight model.”
Reflection AI, company blog, via reflection.ai
Every one of those statements comes directly from Reflection’s own channels rather than a third party paraphrasing the company’s position, which is worth noting given how much of the Beam story still rests on self-reported figures awaiting independent confirmation.
Competitive Comparison: Open-Weight vs. Closed Frontier Models
Launch coverage of Reflection AI Beam is careful to note what the model is not claiming. Reflection has not published a benchmark table pitting Beam against OpenAI’s or Anthropic’s proprietary frontier systems, and the company’s three-to-four-times efficiency claim is explicitly aimed at other open-weight models, both Western releases like Meta’s Llama line and Chinese releases like GLM-5.2, rather than at closed systems. That distinction matters: Beam competes in the open-weight lane, where developers can download, fine-tune, and self-host the model, not in the metered-API lane where ChatGPT, Claude, and Gemini operate.
Readers weighing open-weight self-hosting against paying for a managed API can see the tradeoffs laid out in our guide to GLM-5.3-Flash setup, comparing API access versus self-hosting, a decision that now applies just as directly to Beam, whose weights were published on Hugging Face under the Apache 2.0 license on October 16, 2026. For a wider lens on how fast the leaderboard itself keeps shifting, our breakdown of LMArena vs Artificial Analysis explains why self-reported scores from any lab, Reflection included, need independent confirmation before enterprises bet production workloads on them.
What Comes Next: 5 Predictions for Beam and the Open-Weight Race
- Independent benchmark trackers such as LMArena and Artificial Analysis can now run and publish their own Beam scores, since the weights were released on October 16, 2026, and those numbers will likely diverge at least somewhat from Reflection’s self-reported figures.
- Expect at least one Chinese lab, most plausibly Z.ai or Moonshot AI, to respond within weeks with an updated GLM or Kimi release emphasizing inference efficiency rather than raw parameter count, directly countering Beam’s core pitch.
- Enterprise and government buyers in regulated industries will become the first serious adopters of Beam specifically because of its US domicile and Apache 2.0 license, not because it necessarily tops every benchmark.
- Reflection’s compute partners, SpaceX and Nebius, will likely get more public attention as AI infrastructure deals in their own right, given the combined $7 billion in GB300 access they unlock through 2029.
- Reflection’s valuation trajectory, from $8 billion in October 2025 to a reported $25 billion within six months, suggests another funding round tied directly to Beam’s reception is plausible before the end of 2026.
Frequently Asked Questions
What is Reflection AI Beam?
Beam is Reflection AI’s first open-weight model, a sparse mixture-of-experts system with 501 billion total parameters and 23 billion active parameters, built for coding, reasoning, and agentic tasks. It was unveiled on October 5, 2026.
When will Beam’s weights be publicly available?
SiliconANGLE reported in October 2026 that Reflection has now released the full weights under an Apache 2.0 license. The technical report is still listed as coming and has not been released. An early version had first gone out to a limited group through a waitlist while the model completed red-teaming.
How does Beam compare to Chinese open models like DeepSeek and Qwen?
Reflection positions Beam as a US-based alternative to DeepSeek, Qwen, Kimi K2, MiniMax, and GLM. The best-documented comparison is against Z.ai’s GLM-5.2, where Beam reports a slightly lower Terminal-Bench 2.1 score (80.1 versus 81.0) but claims three to four times lower inference compute. Full benchmark comparisons against DeepSeek, Qwen, Kimi K2, and MiniMax have not been published.
Who founded Reflection AI?
Reflection AI was founded in March 2024 by Misha Laskin, now CEO, and Ioannis Antonoglou, president and chief technology officer. Both previously worked at Google DeepMind, and Antonoglou is credited as a co-creator of AlphaGo.
How much funding has Reflection AI raised?
Reuters reported Reflection raised $2 billion in October 2025 at an $8 billion valuation, with Nvidia, Eric Schmidt, Citi, 1789 Capital, Lightspeed, and Sequoia participating. By spring 2026, reports citing CEO Misha Laskin put the company’s valuation at roughly $25 billion.
What compute deals support Beam’s training and deployment?
Reflection signed a $6.3 billion compute deal with SpaceX in June 2026 for Nvidia GB300 access at SpaceX’s Colossus 2 data center, followed by a $1 billion deal with Nebius in July 2026. Combined, these deals secure GB300 chip access through 2029.
Is Beam’s performance independently verified?
No. As of launch, Beam’s benchmark scores, including 80.9 on SWE-Bench Verified, 80.1 on Terminal-Bench 2.1, and 97.8 on AIME 2026, are self-reported by Reflection AI. TechCrunch and other outlets have noted these figures had not been independently verified at the time of the announcement.
What license will Beam use?
Reflection has released Beam’s weights under an Apache 2.0 license, as SiliconANGLE reported in October 2026, which permits commercial use, modification, and redistribution subject to the license’s standard terms.
