Reflection AI releases Beam, a 501B-parameter open-weight model, under Apache 2.0

Reflection AI has published the full weights of Beam, a 501-billion-parameter mixture-of-experts model, on Hugging Face under the Apache 2.0 licence. That makes it one of the largest Western models that anyone can download, modify and use commercially without a custom licence. The company says it is aimed at coding, reasoning and agent work, but every performance figure so far comes from Reflection itself.

What happened

Reflection announced Beam on 5 October 2026, saying the model was in final red-teaming and that weights would follow "later this month". Early access then ran through a waitlist. On 16 October the company said Beam was "now available across platforms, with open weights under the Apache 2.0 license", according to coverage from Unite.AI and Shattered.io. The model card on Hugging Face (reflection/Beam-1.0-501B-A23B) lists the Apache 2.0 licence, with the FP8, NVFP4 and DFlash2 builds in separate repositories beside it.

The release includes:

• the weights in three precisions: BF16, FP8 and NVFP4;
• a tokenizer and a reference implementation;
• DFlash2, a small "drafter" model for speculative decoding, in which a cheap model proposes several tokens and the large model verifies them in one pass;
• Mirror CLI, an open-source agent harness, and a public beta of Reflection's own API.

Beam is text-only. Reflection says mid-training extends its effective context to 1 million tokens, while its reinforcement-learning runs used contexts of up to 256,000 tokens. Unite.AI lists day-zero support in Transformers, vLLM, SGLang, Unsloth, Axolotl and NVIDIA's Dynamo, and a long list of hosting platforms including AWS, Google Cloud, Azure, Cloudflare, Databricks, Together AI, OpenRouter, Ollama and LM Studio. Shattered.io says Reflection cites "40+" partners but has not published the full list.

Background

Reflection was founded in 2024 by Misha Laskin and Ioannis Antonoglou, both formerly of Google DeepMind. In October 2025 it raised $2 billion in a round led by NVIDIA, which invested $800 million at an $8 billion valuation. Laskin told CNBC in April that the latest round closed at a $25 billion pre-money valuation. The company's stated goal, from the start, was frontier-grade open models from a US lab.

That goal matters because the strongest open-weight models of the past year have mostly come from Chinese labs. Reflection's own announcement compares Beam with GLM 5.2 and says it approaches Qwen 3.8-Max on coding and agentic tasks, while conceding that Kimi K3 remains ahead on raw capability.

There is also a corporate sub-plot. On 10 October the Financial Times reported that NVIDIA is in early talks to deepen its investment in Reflection or acquire it, possibly through a larger stake or an acqui-hire licence. Reuters said it could not immediately verify the report, Bloomberg reportedly said no terms had been agreed, and neither company commented. Treat that as an unconfirmed report.

How it works

Beam is a sparse mixture-of-experts model: 501 billion parameters in total, but only about 23 billion are active for each token, because a router sends each token to a small subset of "experts". Reflection describes 52 layers with interleaved local and global attention and fine-grained routed experts. A user-set "reasoning effort" parameter trades answer length against quality.

Reflection's training figures, all self-reported, give a sense of scale. Pretraining covered 23.8 trillion tokens and ran in under four weeks on 6,144 NVIDIA GB300 GPUs, with 92.3% "goodput" and nine semi-automatic rewinds after faults. Reinforcement learning used about 10,500 GB300 GPUs for four weeks, with more than 100 million rollouts across roughly a million environments. Alignment combines multi-teacher on-policy distillation with deliberative alignment.

One practical point: with 23 billion active parameters, Beam is cheaper per token than a dense 500B model, but you still need memory for all 501 billion. The NVFP4 and FP8 builds exist to make that more manageable on current NVIDIA hardware.

The numbers

These are Reflection's own scores, last updated by the company on 8 October. No independent evaluator had published results for Beam at the time of writing.

SWE-Bench Verified80.9

Terminal Bench v2.180.1

SWE-Bench Multilingual78.0

MCP Atlas78.7

SWE-Bench Pro v2-Hard77.2

SWE-Bench Pro v165.5

DeepSWE v1.144.4

Beam's self-reported scores on coding and agent benchmarks (score)

	score
SWE-Bench Verified	80.9
Terminal Bench v2.1	80.1
SWE-Bench Multilingual	78.0
MCP Atlas	78.7
SWE-Bench Pro v2-Hard	77.2
SWE-Bench Pro v1	65.5
DeepSWE v1.1	44.4

Source: Reflection AI, "Introducing Beam" (company-reported)

On reasoning, Reflection reports 97.8 on AIME 2026, 90.5 on GPQA Diamond and 36.2 on Humanity's Last Exam without tools. Among its weaker results are 38.0 on tau3 banking and 37.0 on AutomationBench, and an AA Omniscience index of 13.0. Reflection also claims GLM-5.2-level reasoning scores with three to four times less inference compute, and itself calls that comparison approximate.

For context, Mistral's Large 4, announced on 6 October, is a 1-trillion-parameter model with 52 billion active parameters. Mistral lists 61.7% on DeepSWE v1.1, against Beam's 44.4, but the two companies use different harnesses and settings, so we do not treat those numbers as comparable. Mistral has promised weights "by the end of the month" and has not named a licence, so Beam is currently the only one of the two you can actually download.

What it means

For developers and companies. Apache 2.0 includes an explicit patent grant and no usage restrictions, which makes legal review far simpler than for models with custom community licences. A US-origin, permissively licensed model of this size is attractive to organisations that cannot or will not run Chinese-origin weights, and the launch hosting list means many can try it through an API they already use.

For competitors. Open weights at this scale put price pressure on closed API providers for coding and agent work, particularly if the efficiency claims hold up. Shattered.io notes that once weights are open, Reflection cannot restrict how they are deployed.

For everyone else. A 501B model is not something most individuals will run at home. The practical benefit is indirect: hosts compete on price, researchers can inspect and fine-tune it, and smaller models can be distilled from it where the licence allows.

What we don't know yet

• Independent results. Every benchmark above is Reflection's. Until third parties run Beam on their own harnesses, treat the ranking against GLM, Qwen and Kimi as a claim.
• The technical report. Reflection said a report and model card would accompany release and that safety evaluations would be published. Shattered.io says no detailed report came with the weights, so training data provenance and safety results cannot yet be checked.
• Pricing. Hosting prices are in flux. One launch write-up lists Baseten at $1.30 per million input tokens, but its own FAQ and another report swap the cache and output figures, so we are not quoting them.
• The NVIDIA talks. Whether a deal happens, and what it would do to Beam's open licence, is unknown.

What to watch next

Independent evaluations from groups such as Artificial Analysis and academic benchmark maintainers; Mistral's Large 4 weights and licence, due by the end of October; any confirmation or denial of the NVIDIA talks; and whether the promised technical report lands. If you plan to build on Beam, test it on your own tasks before trusting leaderboard numbers.

Sources

• Reflection AI: Introducing Beam (primary source)
• Hugging Face: reflection/Beam-1.0-501B-A23B model card
• Unite.AI: Reflection releases Beam weights
• Shattered.io: Reflection Beam open-source, Apache 2.0
• Let's Data Science: NVIDIA may buy Reflection (summarising FT, Reuters, Bloomberg)
• Mistral AI: Introducing Mistral Large 4