Beam by Reflection AI: status and timeline

Beam

A 501B sparse mixture-of-experts open-weight model for coding, reasoning and agents. Only 23B parameters active per token, 3–4x more efficient than dense models of comparable ability.

DeployStage 4 of 5

Announced October 2026; 501B parameters with 23B active. Weights releasing October 2026.

Updated Oct 2026Checked 10 Oct0 updates this week

Milestones

No announced next step

Reflection AI founded by Laskin and AntonoglouMar 2024Complete.
Emerged from stealth with $130M seed and Series A fundingMar 2025Complete.
Series B $2B funding at $8B valuationOct 2025Complete.
Beam announced with weights releasing under Apache 2.0 in October 2026Oct 2026Complete.

Most important, last 3 months

Reflection AI announces Beam, 501B-parameter open-weight model for coding and reasoningOct 2026

Current obstacles

Matching proprietary frontier modelsOpen-weight models must compete on reasoning and coding ability against OpenAI, Anthropic and other proprietary labs.
Specialized hardware for sparse modelsSparse mixture-of-experts models require custom inference kernels and hardware support to realize efficiency gains.

Physics limits

Sparse models require routing overheadDetermining which expert to activate for each token adds latency and memory overhead compared to dense models, offsetting some efficiency gains.
Training heterogeneous experts is harderTeaching different experts to specialize without all activating on every token requires careful load balancing and training techniques.

How it works

3 parts

Mixture of experts

Sparse activation

Different experts handle different token types. Only one expert (23B params) activates per token, not all 501B.

Reasoning and coding

Agentic capability

Beam targets software engineering, complex reasoning, and multi-step tasks that power AI agents.

Open weights

Full transparency

Weights, technical report, model card and developer tooling released under Apache 2.0, enabling deployment without licensing restrictions.

Spec sheet

Spec	Beam
Total parameters	501 billion^R (reported)
Active parameters per token	23 billion^R (reported)
Inference efficiency vs. larger models	3–4x reduction^R (reported)
Training: RL rollouts	100M+ rollouts on 10.5K NVIDIA GB300 GPUs^R (reported)

^R reported by the company

Update log

1 update

Oct 2026

Major: Reflection AI announces Beam, 501B-parameter open-weight model for coding and reasoningPressSoftware

About Reflection AI

The team behind Beam

Reflection AI

Beam open-weight model, sparse experts, reasoning

Reflection AI, founded in March 2024 by former DeepMind researchers Misha Laskin and Ioannis Antonoglou, builds open-weight frontier models for coding and reasoning. Backed by NVIDIA, Sequoia, Lightspeed and others.

Founded by Misha Laskin and Ioannis Antonoglou in March 2024, former DeepMind researchers on Gemini and AlphaGo.
Series B $2B in October 2025 at $8B valuation, led by NVIDIA and Lightspeed.
Beam, 501B parameters with 23B active per token, announced October 5, 2026 with weights to release under Apache 2.0 in October 2026.

reflection.ai (opens reflection.ai)Company profile →

Founded: 20242 yrs
Headquarters: United States
Status: Private
Valuation: $8BprivateOct 2025
Raised: $2.1B2 rounds
Last round: Series B · $2BOct 2025$8B post
Works in: AIOpen-weight models
Coverage: 1 program · 1 updatelatest 10d agochecked 10 Oct
Lead investors: NVIDIALightspeed Venture PartnersSequoia CapitalCRV
Investors: NVIDIA (opens nvidia.com)Sequoia Capital (opens sequoiacap.com)Lightspeed Venture Partners (opens lsvp.com)
People: Misha Laskin (opens scholar.google.com)Co-founderIoannis Antonoglou (opens scholar.google.com)Co-founder

Spec

Beam

Total parameters

501 billion^R (reported)

Active parameters per token

23 billion^R (reported)

Inference efficiency vs. larger models

3–4x reduction^R (reported)

Training: RL rollouts

100M+ rollouts on 10.5K NVIDIA GB300 GPUs^R (reported)