Reflection AI Beam: 501B Open-Weight MoE, Apache 2.0 Weights Due This Month

Related: Hugging Face now hosts RL environments on the Hub, the kind of training task pool Beam used. See OpenEnv and Hub environments.

Update — October 16, 2026: The weights are out. See Reflection Beam is out: 501B open weights on Hugging Face.

Update — October 10, 2026: The FT reports Nvidia is in talks to acquire Reflection or invest more: details and what is unconfirmed.

Update — October 6, 2026: Reflection has now named the model. It is Beam, introduced October 5. The earlier version of this post, written from the Axios scoop, said no name, size or license was public; that is no longer true. Weights are still not out.

Reflection, the Nvidia-backed startup founded by two former Google DeepMind researchers, introduced Beam on October 5, 2026: its first open-weight model, a sparse mixture-of-experts system with 501 billion total parameters and 23 billion active, aimed at coding, reasoning and agents. The Axios scoop a day earlier had said a release was close. Reflection says the weights, a technical report and a model card arrive later this month under Apache 2.0.

The honest framing from Reflection's own post is modest: Beam "advances the Western open-weight frontier," is "competitive with larger open models like GLM 5.2 and approaching Qwen 3.8-Max" on coding and agentic tasks, and trails frontier open models like Kimi K3 on raw capability. Its argument is efficiency per token. This post covers the numbers, how much to trust them, and what to prepare.

An arrow splitting toward small boxes with one lit green.

TL;DR: what is confirmed and what is not

What Reflection announced

Beam is built on a numerically conservative architecture: interleaved local and global attention, fine-grained routed experts, 52 layers, and a controlled residual stream. Reflection says the pretrained base reached near-uniform expert use, with the busiest expert at 1.04 times the average load. It was pretrained on 23.8 trillion tokens of web, public and proprietary licensed data, in under four weeks on 6,144 GB300 GPUs, with nine semi-automatic rewinds and 92.3% goodput late in the run.

The bigger bet is reinforcement learning. Reflection reports over 100 million rollouts on 10.5K GB300 GPUs across four weeks, roughly one million curated environments, about 1.3 billion sandboxes, and an average of 110K concurrent rollouts. It trained with fully asynchronous policy gradients, stable even with a day of staleness, and added a controllable length penalty so users can set reasoning effort. Reflection says capabilities were still improving when the run ended.

Safety work used a separate alignment model merged with the RL model through multi-teacher on-policy distillation, with deliberative alignment for safety policy. Reflection says it will publish safety evaluations in the technical report and open-source the evaluation tools it used.

The benchmark numbers

These are from Reflection's table, with NR meaning not reported. Treat them as vendor claims until someone reruns them on the weights.

Read this plainly. Beam is level with GLM-5.2 and behind the newest open competitors, including DeepSeek V4.1 Flash, a much smaller and cheaper model, on DeepSWE and Terminal Bench. The case for Beam is not the leaderboard. It is the compute per solved task: Reflection estimates inference cost as 2 times active parameters times mean generated tokens, which flatters sparse models, and excludes prefill, attention and serving overhead. That is a reasonable first-order estimate and not a measured price. Compare it with real per-task costs, as in our Mistral Large 4 coverage, where Artificial Analysis cost per task changed the picture.

A demo worth noting, also self-reported: on a viral land-versus-water grid puzzle Beam scored 95.5%, between Claude Opus 5 at 92.5% and Fable 5 at 97.8%. One puzzle is an anecdote, not a benchmark.

The money behind it

The scale of spending is the reason this story traveled. Press reports put Reflection's compute commitments at more than $7 billion through 2029. That includes:

• a partnership with Nebius worth over $1 billion,
• roughly $150 million per month paid to SpaceX's Colossus cluster since July, and
• significant Nvidia server rentals.

On funding, Reflection raised $2 billion in October 2025 at an $8 billion valuation, and later reports cite valuations as high as $25 billion. These are reported numbers, not filings, so treat them as directional. The point for builders is not the valuation. It is that someone has committed enough compute to train a model that is meant to be competitive, and that it is being released as weights rather than sold only as an API. Colossus is also where Anthropic gets capacity, as we covered in the Anthropic and SpaceX Colossus partnership.

Why a US open-weight model matters

For the past year the practical open-weight stack has been mostly Chinese: DeepSeek, Qwen, Kimi, GLM. US-origin open weights exist, including Nvidia's Nemotron 3 Ultra, xAI's Grok weights on Hugging Face and Cohere's Command A+, but few have been the default pick for coding agents.

That gap matters for three kinds of buyers:

1. Regulated and government-adjacent teams that cannot use Chinese-origin models for policy or procurement reasons. A competitive US model widens their options.
2. Companies that want to fine-tune on private data without sending it to an API provider.
3. Anyone worried about concentration, in either direction. The ongoing debate over American closed models versus Chinese open weights is a policy fight with real procurement consequences.

Reflection is not alone in this lane, and Mistral's Large 4 landed the same week. Our coverage of Aleph Alpha's Kolibri shows the European version of the same push, and our explainer on what sovereign AI means maps the trade-offs.

What builders should do before the weights drop

You do not need to wait. A model release is only useful if you can test it quickly. A short readiness list:

1. Write your eval set now. Collect 30 to 100 real tasks from your own work (bug fixes, refactors, extraction, support replies) with pass/fail criteria. Public benchmarks are a starting point, not a decision.
2. Stand up a serving path with a current open model. Run a Qwen or DeepSeek variant through vLLM, SGLang, llama.cpp or your preferred runtime, behind an OpenAI-compatible endpoint. Then swapping in Reflection is a config change. Our guide to choosing open-weight versus closed models walks through the decision, and the DeepSeek V4 Pro benchmarks and pricing give you a current open baseline.
3. Price the hardware. Large mixture-of-experts models are cheaper per token than dense ones but still need memory. If you plan to self-host, compare against hosted open-weight APIs first.
4. Plan the license review. Read the license for commercial use, fine-tuning rights, redistribution and any usage restrictions before you build a product on it.
5. Keep a closed-model fallback. If the first version trails closed models, as reported, route easy and private tasks to open weights and hard ones to a hosted model.

What to watch when it ships

When Reflection publishes, check these before believing any headline:

• Name, size and architecture. Dense or mixture-of-experts, total versus active parameters, context length.
• License text. Not the blog summary, the actual license file.
• Independent benchmarks. Wait for third-party results, and run your own eval set.
• Tool use and long-horizon reliability. For agents, this matters more than a single leaderboard score. Our guide to fixing local models that loop covers the failure modes to test for.
• Quantized variants. Community engines move fast on new open weights, as the Strata thread showed for Qwen, but quantization quality varies by task.
• Availability on hosted providers. Many teams will use it through an inference host rather than self-hosting.

How to tell a report from a release (and why Beam is still a preview)

The gap between "launches" and "is preparing to launch" is where open-weight stories most often mislead. A real release has four artifacts you can verify yourself: a model card with the name and parameter counts, downloadable weights on a public host, a license file, and at least one reproducible benchmark or evaluation script. A scoop has none of them, only sourcing. Until Reflection publishes all four, the practical status is "announced intent."

It also helps to remember what "open-weight" does and does not promise. Open weights let you download and run the model, usually with fine-tuning rights, but training data, training code and sometimes commercial terms stay closed. Some licenses restrict very large deployments or competing products. For a startup planning to customize the model on private data, the license decides whether the whole plan is legal, which is why it belongs at the top of the checklist rather than the bottom.

Finally, consider timing. Open-weight models tend to improve fastest in the weeks after release, as the community ships quantized builds, inference fixes and fine-tunes. The first version you try is rarely the version you will run in production, so budget a short evaluation window and re-test after the first round of community patches.

What this means for what you build or pay

If the model arrives as described, near the top of Chinese open weights but behind the best closed models, the immediate impact is on negotiating power and procurement, not on your best-quality ceiling. You gain a credible US-origin option for sensitive workloads and one more reason for closed-model vendors to keep prices competitive. You do not gain a drop-in replacement for frontier reasoning.

The honest bottom line: this is a promising announcement of an announcement. Prepare your evals, hold your opinion until independent numbers appear, and avoid committing production workloads to a model nobody has been able to download.

Related reading

• Nvidia Nemotron 3 Ultra: 550B open-weight MoE
• American closed AI vs Chinese open weights: the strategy debate
• How to choose open-weight vs closed AI models
• DeepSeek V4 Pro benchmarks, pricing and agent coding
• Aleph Alpha Kolibri: sovereign open-weight MoE
• What is sovereign AI?
• Mistral Large 4 "Le Chonk": 1T open-weight preview
• Anthropic and SpaceX Colossus partnership
• Run a 125B Qwen model on a 12 GB GPU with Strata

Primary: Reflection's Introducing Beam post (October 5, 2026) · Axios scoop, October 4, 2026

Details are accurate as of October 6, 2026 and come from Reflection's announcement. Weights, license text, technical report and independent benchmarks are not yet public, and funding and compute-commitment figures come from press reports. We will update this post when weights ship.