Skip to content
FrontierDot

Reflection Beam: From Oct 5 Preview to Apache 2.0 Weights, and the Hardware It Needs

Live timeline · last updated · 1 entry in the update log

Reflection said the weights would come “later this month” and they did, eleven days later. The release is real. What it takes to run is the part most posts skip.

Reflection’s Beam went from a preview and waitlist on Oct 5 to public Apache 2.0 weights on Oct 16; it is a 501B-parameter, 23B-active text model for coding and agents, and its efficiency claims are Reflection’s own.

Read on 2026-10-17: Reflection’s announcement, model card and developer docs, and the Hugging Face launch post. Benchmark figures are Reflection-reported and we did not run the model. We cite official pages only and add nothing that is not on them.

Timeline

  1. Reflection introduces Beam as its first open-weight model: a sparse mixture-of-experts with 501 billion total and 23 billion active parameters, built for coding, reasoning and agentic work. The post says it was pretrained on 23.8 trillion tokens, and that a reinforcement-learning run generated over 100 million rollouts on 10.5K NVIDIA GB300 GPUs over four weeks. Beam is described as undergoing final red-teaming, with early access through a waitlist and weights promised for later that month.

    Reflection: Introducing Beam

  2. Reflection updates the announcement: it says it has updated Beam’s benchmark results, that the tables on the page reflect the latest numbers, and that all other results are unchanged.

    Reflection: Introducing Beam (update note)

  3. Weights go public. Hugging Face publishes its launch post for Beam under Apache 2.0, with BF16, FP8 and NVFP4 checkpoints, day-0 support in transformers, SGLang and vLLM, and access through Hugging Face Inference Providers. Reflection’s serving docs, last modified the same day, list the NVFP4 download at 273 GiB, FP8 at 468.72 GiB and BF16 at 933.81 GiB.

    Hugging Face: Welcome Beam

The promise, and the date it was kept

The October 5 announcement did two things at once. It showed what Beam could do, and it declined to hand over the model. The post says Beam was “undergoing final red-teaming and evaluations,” points readers to a waitlist for early access, and says Reflection would release the weights, technical report, model card and developer artifacts “later this month.” It also says the weights would come under an Apache 2.0 license, along with documentation and the full stack for running, evaluating and fine-tuning the model.

Eleven days later the Hugging Face launch post was live, dated October 16, and it states the license plainly: Apache 2.0. The developer docs page for downloading the weights carries the same last-modified date. For a company that had promised a month-end release, that is early and complete in the way that counts: the checkpoints exist and the license is the permissive one that was named.

The wording matters because a preview with a waitlist is easy to confuse with a release. Between October 5 and October 16 you could read about Beam but not download it. After October 16 you can. A timeline page is the right format for this kind of story, since the interesting fact is the gap between the two dates and the fact that it closed.

What Beam is, according to the model card

The model card describes Beam as a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active, built for coding, reasoning and agentic workloads. It takes text and produces text. The context window is 1 million tokens, the vocabulary is 152 thousand tokens, and the stated knowledge cutoff is June 30, 2026. The card lists 52 layers, three dense and 49 mixture-of-experts, with 256 experts, and gated grouped-query attention combined with sliding-window attention.

The Hugging Face post adds the detail that makes the layout concrete. Most layers use sliding-window attention over the previous 512 tokens, and every sixth layer plus the last one attend to the whole context, giving 43 sliding-window layers and 9 global ones. That is why the long context is affordable. The post works out that if all 52 layers kept every token, the key-value cache would grow by 208 KiB per token, but with only nine global layers it grows by 36 KiB per token, and it gives about 36 GiB for a full one-million-token sequence in BF16. Those are Hugging Face’s numbers for the transformers implementation, and they are per sequence.

Eight of the 256 experts run for each token, which is where the 23 billion active figure comes from. The post states the consequence directly: all 256 experts stay in memory, so deployment needs the memory of a 501B model while running at roughly the speed of a 23B one. Hold on to that sentence, because it explains the hardware section below.

The efficiency claim, and how to read it

Reflection’s pitch is not that Beam is the strongest open model. The announcement says the opposite in plain words: Beam is competitive with larger open models like GLM 5.2 and approaching Qwen 3.8-Max on coding and agentic tasks, while frontier open models like Kimi K3 remain ahead on raw capability. The advantage it claims is efficiency at inference time: scores comparable to GLM-5.2 on advanced reasoning benchmarks using three to four times less inference compute.

The post is open about how that comparison was built. It estimates compute as roughly two times the active parameter count times the mean generated tokens per attempt, using data from Artificial Analysis and DataCurve, and it says the estimate excludes prompt prefill, attention that depends on context, and serving overhead. In its own words it is an approximate compute comparison rather than measured inference cost. That candor is useful, and it also means the headline ratio is a model of cost, not a bill.

The model card softens the figure slightly: it says Beam reaches comparable solution quality to leading open models using approximately three times fewer inference-time reasoning tokens. Both statements come from Reflection. No independent evaluation is cited on any of the pages we read, so treat every benchmark number as provider-reported until a third party reproduces it.

The card lists scores for each of the three checkpoint formats. On SWE-Bench Verified, for example, it shows 81.6% for BF16, 80.2% for FP8 and 80.7% for NVFP4; on Terminal-Bench 2.1, 80.1%, 78.9% and 78.4%. The spread between formats is small on most rows, which is the point of that table: if you run the cheaper format you should expect similar, not identical, results.

What it takes to run

This is where a 23B-active model stops sounding small. Reflection’s own serving guide asks for four NVIDIA B300 GPUs connected by NVLink on a Linux amd64 host, with Docker and the NVIDIA Container Toolkit, and notes the checkpoints are large enough that you should check disk space first. The NVFP4 checkpoint is 273 GiB. The FP8 one is 468.72 GiB, and BF16 is 933.81 GiB.

The Hugging Face post gives the hardware reasoning in more general terms. NVFP4 shrinks the BF16 checkpoint from about 1 TB to roughly 280 GB and runs natively on Blackwell GPUs, but not on older ones such as the H200. FP8 is the format for Hopper and Ampere. It says two B200s at 180 GB each could hold the weights with little room left over, recommends planning for four B200s, and names a full 8×B200 node, two B300s at 288 GB each, or eight RTX Pro 6000 cards as other options. For people without that hardware it points to Inference Providers, listing Baseten, Fireworks and Together at launch.

So the honest summary for an individual developer is that Beam is open in the license sense and very much a datacenter model in the hardware sense. The people best placed to benefit from Apache 2.0 here are teams that want to host, fine-tune or distill a coding model themselves, and the model card says Reflection does not prohibit distillation, fine-tuning, quantization or merging.

What we could not verify

  • Whether Beam’s inference-efficiency ratio holds in someone else’s serving stack. Reflection describes its own calculation as an estimate, and we found no independent measurement on the official pages.
  • The full safety evaluation. The card reports refusal and compliance rates from Reflection’s own Iris Safety Evaluations and points to a separate safety report, which we did not read in full.
  • Any comparison against closed models. The pages we cite compare Beam to open models only.

Our take

Beam is a clean release by the standards of this year’s open-weight launches: a dated announcement, a stated deadline, a permissive license, three checkpoint formats and working serving recipes on day one. The weak spot is the framing. Efficiency claims are the easiest ones to make and the hardest to check, and the pages themselves say the comparison is approximate. We will keep this page as a live timeline. When an independent evaluation, a technical report or a licensing wrinkle appears, it gets a dated node and an entry in the update log, and the conclusion at the top changes if the facts do.

Related reading on this site: how Ling-3.1-flash’s 25B-active design maps to your hardware makes the same memory-versus-speed point for a different mixture-of-experts model, and the Muse Spark Open timeline covers the other big open-weight release of the same week.

Keep reading on Front

Every post, newest first, is on the Front blog index.

FAQ

Is Reflection Beam open source?

The weights are public under the Apache 2.0 license, per Reflection and Hugging Face. Training data and the full training recipe are not part of the release as far as the pages we read show.

When were Beam’s weights released?

Reflection announced Beam on October 5, 2026 and the weights were public by October 16, 2026, when Hugging Face published its launch post.

What hardware does Beam need?

Reflection’s serving guide lists four NVIDIA B300 GPUs for the NVFP4 checkpoint, which is 273 GiB. FP8 is 468.72 GiB and BF16 is 933.81 GiB.

Update log

  • — First published.

Sources