Reflection AI launches Beam, a 501B open-weight model

Reflection AI announced Beam on October 5, 2026, an open-weight language model with 501 billion parameters, of which 23 billion are active for each token. The weights, a model card and a technical report will ship "later this month" under the Apache 2.0 license, Reflection's announcement says. The company claims Beam matches GLM-5.2 on reasoning while using three to four times less inference compute. If that holds up, developers get a large coding model with no custom license terms attached.

What Reflection released

Beam is a mixture-of-experts model, or MoE. An MoE model splits its parameters into many smaller "expert" networks and uses only a few of them for each token. That is why Beam stores 501 billion parameters but computes with 23 billion at a time.

Reflection trained it on 23.8 trillion tokens of web, public and licensed data. The context window started at 256,000 tokens and was extended to 1 million in a later training stage. The context window is how much text the model can read at once.

Only early access exists today, through platform.reflection.ai. Full weights come later in October under Apache 2.0, a permissive license that Crypto Briefing calls "one of the most permissive open-source licenses."

The benchmark numbers

Reflection published scores across coding, reasoning and agent tests. All of these are the company's own numbers. Crypto Briefing notes that independent testing has to wait until the weights are public.

Benchmark	What it tests	Beam
SWE-Bench Verified	Fixing real GitHub issues	80.9
SWE-Bench Pro v1	Harder software tasks	65.5
Terminal Bench v2.1	Command-line tasks	80.1
MCP Atlas	Tool calling through MCP	78.7
BrowseComp (with context management)	Web research	77.4
GPQA Diamond	Graduate-level science questions	90.5
AIME 2026	Competition math	97.8
Humanity's Last Exam (no tools)	Very hard expert questions	36.2

The comparisons depend on who is reporting. Reflection says Beam scores comparable to GLM-5.2 with three to four times less inference compute. SiliconANGLE writes that Beam approaches Alibaba's Qwen 3.8-Max and trails Anthropic's Claude Fable 5.1. Crypto Briefing goes further and says Beam outperforms both GLM-5.2 and Qwen 3.8-Max. The reports conflict on Qwen, so treat that claim as unsettled.

How Reflection trained it

The training run was short and large. Reflection says pretraining used 6,144 Nvidia GB300 GPUs for under four weeks, with 92.3% goodput. Goodput is the share of time the cluster spent on useful training rather than recovering from faults.

Reinforcement learning, where the model improves by trying tasks and getting scored, used 10,500 GB300 GPUs for four weeks. SiliconANGLE gives the figure as 10,000. Reflection says the stage ran 1.3 billion sandboxes and more than 100 million practice attempts. A sandbox is an isolated environment where the model can run code safely.

Reflection also reports even use of its experts. The busiest expert carried 1.04 times the average load, which means the work was spread almost evenly.

Who is behind Beam

Former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou founded Reflection in March 2024, according to SiliconANGLE and Crypto Briefing. Crypto Briefing reports the company has raised more than $4 billion at a $25 billion pre-money valuation. Its backers include Nvidia, Sequoia and Citigroup.

What this means for developers

There is nothing to download yet, so plan rather than migrate. Join the early-access program if you want to test Beam on your own coding tasks before the weights land.

Budget for big hardware. The MoE design cuts compute per token, not memory. All 501 billion parameters must still sit in memory, which is about 500GB at one byte per parameter. When this site covered GLM-5.3's open weights in August, that model needed at least eight high-end GPUs. Expect Beam to need a multi-GPU server too, or a hosted provider.

The license is the clearest advantage. Apache 2.0 allows commercial use without the custom terms that came with GLM-5.3. Have legal check the final license file when the weights ship, since only the announcement exists today.

Finally, rerun the benchmarks yourself. Reflection compared Beam with GLM-5.2, though Z.ai has since shipped GLM-5.3. Test both on your own repositories before you switch, and watch for the technical report later this month.