InvestAlso reported elsewhere2 publishers2 min readPublished
Reflection AI says Beam reaches GLM-5.2 reasoning scores on 3-4x less inference compute
Reflection AI said it will release the weights for Beam, a 501-billion-parameter open model with 23 billion active, later this month. Red-teaming is unfinished, so every score so far is the company's own.
The Investor · Invest desk
Figures Reflection reported for Beam: 501 billion total parameters, 23 billion of them active; pretraining on 23.8 trillion tokens; an RL run on 10.5K Nvidia GB300 GPUs over 4 weeks that generated over 100 million rollouts; and approximately 1.3 billion sandboxes used for training and grading.
| Figure | Value | Note | Claim |
|---|---|---|---|
| Total parameters | 501 billion | 23 billion active | 1 |
| Pretraining tokens | 23.8 trillion | 2 | |
| GPUs in the RL run | 10,500 Nvidia GB300s | over 4 weeks, over 100 million rollouts | 3 |
| Training and grading sandboxes | 1.3 billion | approximately | 4 |
What happened
- Reflection describes Beam as a sparse mixture-of-experts model built for coding, reasoning and agentic workloads, and as the company's first open-weight release.
- Reflection's reinforcement-learning run generated more than 100 million rollouts on 10.5K Nvidia GB300 GPUs over four weeks.
- Reflection says Beam is competitive with larger open models like GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks.
- Daily Upside reported that Nvidia was among the investors in a $2 billion round last year that valued Reflection at $8 billion.
Why it matters
- constraint In the text of Reflection's post, every named comparison is to an open model, so a builder weighing Beam against a closed frontier API has to run that test once the weights ship.
- cost The RL run alone was about 7.06 million GPU-hours, so each $1 of hourly GPU price adds about $7 million to its cost. Reflection did not disclose the rate it paid.
- exposure The $25 billion round the Wall Street Journal reported Reflection discussing would be half of Moonshot's $50 billion valuation, and Moonshot and DeepSeek are both heading to IPOs.
- decision Teams that want a non-Chinese open-weight option, the buyer Daily Upside describes, now have a candidate to schedule evaluations for, but any commitment waits on the weights.
We think the efficiency claim is the one to test and the capability claim is the one to take at Reflection's word, because the company concedes the capability point itself. Its post says frontier open models like Kimi K3 remain ahead on raw capability and that Beam's advantage is efficiency at inference time [18]. The design agrees. Beam has 501 billion parameters [1] against a 2T-plus family for Qwen 3.8-Max [20], so it is about a quarter of that size or less [22], and only 23 billion are active, or 4.6% of the total [12].
Taken at face value, 3-4x less inference compute than GLM-5.2 at comparable advanced-reasoning scores [19] puts Beam at a quarter to a third of the compute [23], or rather puts it there on the benchmarks Reflection reports. The weights are still unreleased [16]. The post shows per-benchmark results in a figure, with NR marking scores that have not been reported [6].
Pretraining points the same way. Reflection used 23.8 trillion tokens [2], about 1,035 for each active parameter [13]. Training cost is paid once and serving cost on every request. In our view, that ratio assumes total serving spend will exceed training spend.
The counter-case is cost per task. Daily Upside cites a study of 6,800 tasks from researchers at Stanford, Carnegie Mellon, Berkeley and Microsoft in which cheaper models cost more than expensive ones on a third of tasks, because they took longer and sometimes failed [21]. Reflection's design answers part of that: it trained with a controllable length penalty and offers a reasoning effort setting where lower values favor shorter responses [5].
Either outside runs reproduce the compute ratio, and builders get a cheaper reasoning option from a US lab. Or the saving narrows once completions lengthen. Reflection says they fell early in reinforcement learning and grew again as agentic skills developed [7]. The post's text ties the 3-4x figure to advanced reasoning benchmarks [19]. Or the release slips past the month Reflection named [16]. We put most weight on the second.
What to watch
- Whether Reflection posts the weights, technical report and model card by the end of this month, as its post promises.
- Third-party results for Beam against GLM-5.2 at several reasoning effort settings, measured in cost per solved task.
- Whether the reported $25 billion round closes, and at what valuation against Moonshot's $50 billion.
Clarity's read
What the record supports and how the coverage leans. The claims behind it follow.
Reality
- Evidence35
- Adoption
- Insufficient
- Hype gap+30
- Incentives75
- Confidence55
Claim ledger
Ranked by verification strength, evidence, and original report placement.
- [1]
Beam is Reflection's first open-weight model, a sparse Mixture-of-Experts model with 501 billion total parameters and 23 billion active, built for coding, reasoning, and agentic workloads.
- [2]
Reflection pretrained Beam on 23.8 trillion tokens from the web and proprietary licensed datasets.
- [3]
Reflection's high-compute RL run generated over 100 million rollouts on 10.5K NVIDIA GB300 GPUs over 4 weeks of training.
- [4]
Beam's training and grading used approximately 1.3 billion sandboxes, drawn from one million coding, agentic and STEM environments.
- [5]
Reflection trained Beam with a controllable length penalty, and users control the capability-versus-tokens tradeoff through a reasoning effort parameter where lower settings favor shorter responses.
- [6]
A figure in the post shows Beam's performance across coding, agentic, reasoning and STEM benchmarks, with NR denoting scores that have not been reported.
- [7]
Early in RL, performance improved as completion lengths fell; later, as Beam developed stronger agentic capabilities, completion lengths grew again, with the additional tokens supporting further gains.
- [8]
Nvidia was among the investors in a $2 billion funding round last year that valued Reflection at $8 billion.
- [9]
The Wall Street Journal previously reported that Reflection has held talks this year for another fundraising round that would value it at $25 billion.
- [10]
Moonshot closed a fundraising round that valued it at $50 billion, according to Bloomberg, and is heading to an IPO in early 2027; DeepSeek is nearing a new round of at least 80 billion yuan ahead of a planned IPO next year.
- [11]
Daily Upside frames Beam as news for domestic AI users who do not want to pay for costly closed frontier models and do not want to rely on a Chinese open-weight ecosystem.
- [12]
Beam's active parameters are about 4.6% of its total.
- [13]
Pretraining used about 1,035 tokens per active parameter.
- [14]
The RL run was about 7.06 million GPU-hours, so each $1 of hourly GPU price adds about $7 million.
- [15]
The $25 billion valuation Reflection reportedly discussed is half of Moonshot's $50 billion.
- [16]
Beam is undergoing final red-teaming and evaluations, and Reflection will release the weights, technical report, model card and developer artifacts later this month.
ReportedInsufficientSource: Reflection AI blog post, reflection.ai2 sources— create a free account to open themView cited source - [17]
Reflection says Beam is competitive with larger open models like GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks.
ReportedInsufficientSource: Reflection AI blog post, reflection.ai2 sources— create a free account to open themView cited source - [18]
Reflection says frontier open models like Kimi K3 remain ahead on raw capability, and that Beam's advantage is efficiency at inference time.
ReportedInsufficientSource: Reflection AI blog post, reflection.ai2 sources— create a free account to open themView cited source - [19]
On advanced reasoning benchmarks, Reflection says Beam achieves scores comparable to GLM-5.2 while using 3-4x less inference compute.
ReportedInsufficientSource: Reflection AI blog post, reflection.ai2 sources— create a free account to open themView cited source - [20]
Models in the 2T+ parameter family, like Qwen 3.8-Max, require significantly more inference compute per token, according to Reflection.
ReportedInsufficientSource: Reflection AI blog post, reflection.ai2 sources— create a free account to open themView cited source - [21]
A study spanning 6,800 tasks by researchers at Stanford, Carnegie Mellon, Berkeley and Microsoft found cheaper models ended up costing more than more expensive models for one-third of tasks because they took longer and sometimes ended in failure.
- [22]
Beam's 501 billion parameters are about a quarter or less of a 2 trillion-parameter model.
- [23]
3-4x less inference compute means Beam uses a quarter to a third of GLM-5.2's compute on the reported benchmarks.
Sources
2 independent publishers whose own reporting we read for this story.
- Reflection AI
reflection.ai
1 article · October 11, 2026
- thedailyupside.comReflection Wants to Deliver America’s Most Powerful Open-Weight AI
1 article · October 11, 2026
Topics and entities
Follow any of these and your For You feed starts watching them — no settings page required.
Topics
- Open-Weight Model PublishingFollow
- AI Lab FundingFollow
- Large-scale reinforcement learning for LLMsFollow
- Inference Efficiency and Memory FootprintFollow