Reflection Beam is a mixture-of-experts open-weight model
Reflection Beam is a large language model whose trained weights Reflection AI plans to publish, so developers can download and run it themselves. Reflection announced it in a blog post on October 5, 2026, while the model was still in final red-teaming and evaluation.
Reflection Beam at a glance (Reflection's announcement)
Only 23B active parameters run per token in the MoE model
An MoE model splits its network into many expert blocks and routes each token through a few of them, so Beam stores 501B parameters but computes with 23B for every token it generates. The active count sets how much arithmetic each token costs, and Reflection's efficiency claim is built on that number.
For scale, TechCrunch puts Z.ai's GLM 5.2 at about 744B total and 40B active. Beam is smaller on both counts.
We are introducing Beam, Reflection’s first open-weight model. Beam is a sparse Mixture-of-Experts model with 501 billion total parameters, 23 billion active, built for coding, reasoning, and agentic workloads. — From the opening paragraph that states the architecture, size and intended workloadsView official source →
Built for AI coding and AI agents
Reflection built Beam for coding, reasoning and agent tasks. It was pretrained on 23.8 trillion tokens, and a midtraining stage stretched its effective context to 1M tokens so it can hold a large codebase or a long agent session at once.
Beam reads and writes text only. Reflection says it can still use images or other data if they are first turned into text.
Founded by DeepMind alumni and backed by Nvidia
Reflection AI is a Brooklyn lab founded in 2024 by two former Google DeepMind researchers, and Nvidia is among its investors. TechCrunch, citing PitchBook, reports about $4.7 billion raised from Nvidia, Sequoia Capital, Lightspeed Venture Partners and others, with the last round at a $25 billion pre-money valuation. Over the summer of 2026 the company signed deals worth more than $7 billion with SpaceX and Nebius that secure Nvidia GB300 chips through 2029.
This summer, Reflection signed deals collectively worth more than $7 billion with SpaceX and Nebius to secure access to Nvidia’s GB300 chips through 2029. — From the paragraph on Reflection's compute contractsView official source →
Nvidia in talks to buy or invest further, per the FT
On October 10, 2026, the Financial Times reported that Nvidia is in talks to acquire Reflection AI or deepen its investment. According to Reuters, the talks are early and could take several shapes. One option is an acqui-hire, where Nvidia hires the staff and licenses the technology instead of buying the company outright, which could sidestep a long regulatory review.
A deal could come within weeks, but the talks could also collapse. Reflection declined to comment, and Nvidia did not respond right away. Nvidia has already put $800 million into the startup.
Neither company has said whether a deal would change the plan to release Beam's weights under Apache 2.0 this month. Until the weights are actually published, treat the open release as a stated intent rather than a given.
Talks are at an early stage and a deal could take several forms, including a so-called acqui-hire arrangement where Nvidia would hire staff and license technology rather than pursue a full acquisition, potentially avoiding a lengthy regulatory review, the newspaper said. ／ An agreement could be reached in the coming weeks, according to the report, which cited multiple people familiar with the matter, while adding that the discussions could still fall apart. — From the bullet points summarizing the FT reportView official source →
Benchmark scores against Kimi K3, GLM 5.2 and DeepSeek V4
On the table Reflection published, Beam lands level with GLM 5.2 and behind Kimi K3, Qwen 3.8 Max and DeepSeek V4.1 Flash. All the numbers below are Reflection's own measurements, and no third party has checked them yet.
SWE Bench and Terminal Bench coding results
DeepSWE and SWE-Bench Pro give a model real bugs from real repositories and count how many it fixes. Terminal Bench scores multi-step work in a command line, and MCP Atlas scores tasks that require calling outside tools. Higher is better on all four, and Beam's best showing against the field is on SWE-Bench Pro and MCP Atlas, where it edges past GLM 5.2.
Beam vs Chinese open-weight models on coding and agent benchmarks (Reflection's figures)
Beam and GLM 5.2 are within half a point. DeepSeek V4.1 Flash and Kimi K3 are far ahead.
Beam
GLM 5.2
Kimi K3
Qwen 3.8 Max
DeepSeek V4.1 FlashBeam is last of the five, but its gap to DeepSeek V4.1 Flash is smaller here than on DeepSWE.
BeamGLM 5.2Kimi K3Qwen 3.8 MaxDeepSeek V4.1 FlashBeam beats GLM 5.2 and falls just short of Qwen 3.8 Max. Reflection lists no score for the other models.
BeamGLM 5.2Qwen 3.8 MaxBeam is ahead of GLM 5.2 and behind Kimi K3 and Qwen 3.8 Max. The table has no DeepSeek V4.1 Flash score.
BeamGLM 5.2Kimi K3Qwen 3.8 MaxReflection also reports 80.9 on SWE-Bench Verified, ahead of Inkling at 77.6. Inkling is the open model Thinking Machines Lab released in July 2026, so the comparison covers a Western rival as well as Chinese ones.
Where Beam trails Kimi, Qwen 3.8 Max and DeepSeek
Beam scores below Qwen 3.8 Max on all four charts and well below DeepSeek V4.1 Flash wherever both appear. The widest gap is DeepSWE, 44.4 against 74.2.
Kimi K3 also leads on the knowledge tests: 46.9 to 36.2 on Humanity's Last Exam without tools, and 93.5 to 90.5 on GPQA Diamond. Reflection does not dispute any of this. It writes that Kimi K3 and similar models stay ahead on raw capability and rests Beam's case on efficiency.
Where frontier open models like Kimi K3 remain ahead on raw capability, Beam's advantage is efficiency at inference time. — From the sentence that follows the benchmark tableView official source →
Inference cost claim of 3-4x less compute than GLM 5.2
Reflection says Beam matches GLM 5.2 on advanced reasoning benchmarks while using 3-4x less inference compute. It adds that the gap grows against 2T-plus parameter models such as Qwen 3.8 Max.
How Reflection arrived at the efficiency figure (Reflection's announcement)
FLOPs give an estimate, not measured inference compute
FLOPs count the floating-point operations a model performs to produce an answer. The 3-4x figure comes from a formula, not from running the models on the same hardware and measuring the cost. The formula rewards a model with fewer active parameters and shorter answers, and it skips prompt prefill, attention cost that grows with context, and serving overhead.
The figure also says nothing about memory. Only 23B parameters do the math for each token, but a server still has to hold all 501B, so hosting costs will not shrink by the same ratio.
These estimates exclude prompt prefill, context-dependent attention operations, and serving overhead, so they represent an approximate compute comparison rather than measured inference cost. — From the method note under the efficiency comparisonView official source →
Reinforcement learning on 10,500 Nvidia GB300 GPUs
Reflection credits Beam's efficiency to large-scale reinforcement learning, which kept 10,500 Nvidia GB300 GPUs busy for 4 weeks and produced more than 100 million rollouts. The runs drew on roughly 1 million environments covering coding, agent and STEM tasks. For comparison, Reflection cites 30 million rollouts for Inkling and 753K for MiMo.
Safety tuning through distillation and red teaming
Reflection trained a second model only on safety and alignment, then merged it with the large RL model through multi-teacher on-policy distillation (MOPD). Distillation here means one model learning to reproduce the outputs of teacher models.
At launch Beam was still in final red-teaming, where testers try to make the model misbehave. Reflection plans to publish the safety results in its technical report and to open-source its internal safety evaluations.
Apache 2.0 license, commercial use and weight release date
Reflection plans to release Beam's weights in October 2026 under Apache 2.0, together with a technical report, a model card and tools for running, evaluating and fine-tuning the model. Until then, access is limited to a waitlist.
Beam release plan (Reflection and TechCrunch)
Running Beam as a local LLM on your own GPU
You cannot run Beam as a local LLM until the weights ship, and at 501B parameters it is not sized for a laptop. The full 501B parameters have to sit in memory even though only 23B are active per token.
Once the weights are out, Apache 2.0 lets you copy, modify and redistribute the model for free, with a patent license included. Nothing in the license limits commercial use, so Beam can go into internal tools or paid products.
This month, we will release the weights under an Apache 2.0 license, along with documentation and the full stack for running, evaluating, and fine-tuning the model. — From the paragraph announcing the releaseView official source →
Fine tuning LLM tools and open source AI library support
The release is meant to include a fine-tuning stack, and TechCrunch reports that Reflection plans integrations with open-source libraries at launch. Distribution will run through hyperscalers (the largest cloud platforms) and neoclouds (cloud providers built around GPUs).
The blog also shows Beam writing a fine-tuning notebook for the smallest Gemma-4 model after reading Unsloth's documentation. Reflection has not named which libraries or clouds will carry Beam, so check the release notes before planning a deployment.
Reflection says it will release Beam’s weights and full technical details this month, with distribution through hyperscalers and neoclouds and integrations across open source libraries at launch. — From the paragraph on release timing and distributionView official source →
Beam's case rests on Reflection's own numbers
Reflection Beam is a 501B mixture-of-experts model from a US lab, sold on efficiency rather than top scores. On Reflection's own numbers it ties GLM 5.2 and trails Kimi K3, Qwen 3.8 Max and DeepSeek V4.1 Flash, and its 3-4x compute saving is a FLOPs estimate.
The weights and technical report are due in October 2026, and the FT has reported that Nvidia is in talks to buy the company. Only independent benchmarks after that release will allow a fair comparison. For the models it was measured against, see:
FAQ
- Can I download Reflection Beam today?
- Not yet. Reflection says the weights, a technical report and a model card will come out in October 2026 under the Apache 2.0 license. Until then, the only way in is the early-access waitlist for a select group of users.
- Is Reflection Beam better than Kimi K3?
- No, not on raw scores, and Reflection says so itself. On its own table Kimi K3 leads Terminal Bench v2.1 with 88.3 to Beam's 80.1, and DeepSWE v1.1 with 68.0 to 44.4. Reflection's case for Beam is that it reaches GLM 5.2-level reasoning scores with less inference compute.
- Can Reflection Beam be used commercially?
- Yes, once the weights are out. Apache 2.0 grants a free, worldwide, irrevocable license to copy, modify and redistribute the model, plus a patent license, and it has no clause that limits commercial use.
Related articles
Kimi K3.1 Pro benchmarks, pricing and open weightsKimi K3.1 Pro is Moonshot AI's post-trained Kimi K3. It nearly matches Claude Opus 5.5 on coding, costs a third less than K3 and now has open weights.
What is mixture of experts (MoE)? How it worksMixture of experts (MoE) splits an AI model into many experts and runs only a few per token. See how routing works, what active parameters mean and the catch.
Agentic memory: IBM finds a dose, not a switchAgentic memory is a dose, not a switch: IBM Research tested eight models on AppWorld, and curated retrieval gave gpt-oss-120b +16.1 points for +5% tokens.
Apple Intelligence launches in China with QwenApple Intelligence comes to iPhones in China with iOS 27.3 in December. Alibaba's Qwen handles some Siri and Writing Tools requests, but Siri AI is left out.