Relevant if you want a Western open-weight option for coding agents. Beam takes text only, has a 1 million token context window and ships in three precisions, alongside an open-source terminal agent called Mirror. Reflection says Kimi K3 still leads on raw capability, and claims its edge is lower inference cost.
Reflection AI is a New York AI lab founded in 2024 by Misha Laskin and Ioannis Antonoglou, both formerly of Google DeepMind. It builds open-weight models for coding and agentic work. Its first release, Beam, is a 501-billion-parameter mixture-of-experts model whose weights were published in October 2026.
Founded: 2024 · HQ: New York, NY, USA · Team: 50-100 · CEO: Misha Laskin · Funding: $2B raised (company post); investors incl. NVIDIA, Lightspeed, Sequoia, DST, CRV, B Capital, Citi, 1789 · Valuation: $8B (October 2025, reported by Reuters)
About Reflection AI
Reflection AI, Inc. dates to 2024, when Misha Laskin and Ioannis Antonoglou left Google DeepMind to start it. Press coverage credits Laskin with leading reward modeling for Gemini and Antonoglou with co-creating AlphaGo. Laskin is chief executive and Antonoglou is chief technology officer. The company is based in New York, and its careers page lists roles in New York, San Francisco, London and Washington D.C. It first applied large-scale reinforcement learning to autonomous coding, then widened the goal to general agentic reasoning. Its stated aim is frontier open intelligence, set against closed labs such as OpenAI and Anthropic.
The product line has four parts, per the company site. Beam is the open-weight model. The Reflection API platform serves Beam as a hosted endpoint. Mirror is an open-source command-line agent framework for running Beam as a coding assistant, written as a Python frontend over a Rust agent core. A forward-deployed solutions team works with enterprises and governments on custom deployments, including private cloud, on-premises, air-gapped and edge setups.
Beam is the first model. Reflection announced it on October 5, 2026 and made it available on October 16, 2026. It is a sparse mixture-of-experts model with 501 billion total parameters and 23 billion active per token, pretrained on 23.8 trillion tokens, with a 1 million token context window and text-only input. Reflection says its reinforcement learning run used 10.5 thousand NVIDIA GB300 GPUs over four weeks and generated more than 100 million rollouts. The weights are published on Hugging Face in BF16, FP8 and NVFP4, with a speculative decoding drafter. Day-zero serving support covers Hugging Face Transformers, vLLM, SGLang, Unsloth, Axolotl and NVIDIA Dynamo, and Reflection says it tested Beam with OpenCode and Hermes Agent as well as Mirror.
Funding comes from a $2 billion raise that Reflection describes in its own post, titled Building Frontier Open Intelligence. Reuters reported that round in October 2025 at an $8 billion valuation, led by NVIDIA. Reflection names B Capital, Citi, CRV, Disruptive, DST, Eric Schmidt, Zoom Ventures, Lightspeed, Sequoia and 1789 among its investors. Data provider PitchBook, cited by press, put an earlier raise at $130 million when the company left stealth. Press reports in March 2026 described talks on a much larger round at a higher valuation; Reflection has not confirmed a close, so no later figure is recorded here.
On capability, the figures below are Reflection's own and have not been reproduced by HokAI. Its launch table lists SWE-bench Verified at 80.9 and GPQA Diamond at 90.5 for Beam. Reflection describes Beam as competitive with GLM-5.2 and approaching Qwen3.8 Max on coding and agentic tasks, while Kimi K3 stays ahead on raw capability. Its pitch is efficiency: it claims scores comparable to GLM-5.2 on advanced reasoning with three to four times less inference compute. The nearest open-weight peers by company are Moonshot AI, Z.ai, DeepSeek, Mistral AI and Meta AI.
Safety is handled through a published Open Safety Framework. It says Reflection's approach is informed by the NIST AI Risk Management Framework, ISO 42001 and voluntary commitments, and that a standing governance group covering research, engineering, security, product, legal and policy owns the process. Reflection says Beam went through final red-teaming and evaluation before the weights were released. The company argues openness lets outside researchers probe risks, and it also acknowledges that widely available capable models raise misuse concerns.
The limits are worth stating. Beam is one model released weeks ago, so there is little independent track record. It takes text only and was built mainly for English. HokAI found no SOC 2 or ISO 27001 report or trust center for Reflection's hosted API, so no certifications are listed. Open weights also mean the safeguards can be removed by anyone who fine-tunes the model, which is the central objection critics raise against this approach. And the valuation reports are press figures, not company disclosures.
Mission
Make intelligence open and accessible to all by building open models that anyone can use, inspect and build on.
Products
- Beam (Open-weight model): https://huggingface.co/collections/reflection/beam-10
- Reflection API Platform (Developer API): https://platform.reflection.ai/
- Mirror CLI (Open-source agent framework): https://reflection.ai/blog/beam-is-available
- Reflection Solutions (Forward-deployed services): https://reflection.ai
Reflection AI Models on HokAI
Links
Frequently Asked Questions
How much funding has Reflection AI raised?
Reflection says it raised $2 billion, as stated on its own blog. Reuters put the October 2025 round at an $8 billion valuation, with NVIDIA leading. Backers include NVIDIA, Lightspeed and Sequoia, plus several other firms and individuals the company names. Press reports from March 2026 described talks on a larger round, which Reflection has not confirmed.
What products does Reflection AI make in 2026?
It makes four things: Beam, an open-weight coding and agent model; a hosted API platform that serves Beam; Mirror, an open-source command-line agent framework; and a forward-deployed solutions service for enterprises and governments. HokAI has not recorded API prices, so check the platform page for current rates before budgeting.
How does Reflection AI handle data privacy and compliance?
HokAI could not find a trust center or audit report for the hosted API, so no certifications are listed. Its Open Safety Framework says its approach draws on the NIST AI Risk Management Framework and ISO 42001, which is a guidance statement and not a certification. Because Beam can be self-hosted, teams that need data to stay in their own environment can run the weights themselves.
Who are Reflection AI's main competitors?
The closest rivals are open-weight labs: Moonshot AI with Kimi K3, Z.ai with GLM, DeepSeek and Alibaba with Qwen. Reflection says Beam beats none of them on raw capability but uses less inference compute, so it wins on serving cost and as a Western-built option. It loses on track record, since Beam is its first release against rivals with several generations out.
How do you start using Reflection AI's products?
The fastest route is an API key from the Reflection platform, which takes minutes. To self-host, pull the Beam weights from the Hugging Face collection and serve them with vLLM, SGLang or Transformers; the FP8 and NVFP4 versions cut memory needs for a model this size. For a terminal agent, install the open-source Mirror CLI and point it at either endpoint.