Reflection AI
A US-based AI startup focused on developing open-weight large language models, with the goal of building competitive open foundation models within the Western AI ecosystem to rival models such as DeepSeek and Qwen.
Core Facts
Timeline (last 90 days)
Factory integrated Beam with Reflection AI's models and stress-tested it in real coding and agent workflows.
Beam's specific capability boundaries, the actual performance of Reflection AI's models in the tests, and whether Factory will turn the test conclusions into product features or pricing strategies are all currently undisclosed.
The core signal conveyed by the tweet is how much Factory values pre-release models, rather than an endorsement of the performance of Beam or Reflection AI's models.
NVIDIA has invested billions of dollars to support open-weight models and has acquired companies such as Hugging Face and Reflection AI.
Reflection AI says Beam was trained end-to-end from scratch rather than adapted from an existing open-source model, and the full weights will be released later this month.
Beam introduces a controllable length penalty into the reinforcement learning reward function, penalizing unnecessary token generation while rewarding successful task completion.
Reflection AI released the companion draft model DFlash2, with 1.93 billion parameters, designed specifically for speculative decoding.
Beam is explicitly optimized for three scenarios—software engineering, reasoning and long-running agentic tasks—rather than being a general-purpose chat model.
Reflection AI is entering the market with an "open weights + cloud hosting" combination rather than pure local deployment or a purely closed-source API.
Ollama has entered into a partnership with Reflection AI, and Reflection AI's first open-weight model Beam is now officially available on Ollama Cloud.
40 more timeline events
All Facts (20)
Reflection AI released an open-source model called Beam, with approximately 501 billion total parameters and only 23 billion activated parameters, using a MoE architecture, designed specifically for reasoning, coding, and agentic workflows, trained end-to-end from scratch, with weights to be made available this month.
70%UnverifiedBeam's specific capability boundaries, the actual performance of Reflection AI's models in the tests, and whether Factory will turn the test conclusions into product features or pricing strategies are all currently undisclosed.
50%UnverifiedThe core signal conveyed by the tweet is how much Factory values pre-release models, rather than an endorsement of the performance of Beam or Reflection AI's models.
50%UnverifiedNVIDIA has invested billions of dollars to support open-weight models and has acquired companies such as Hugging Face and Reflection AI.
50%UnverifiedReflection AI says Beam was trained end-to-end from scratch rather than adapted from an existing open-source model, and the full weights will be released later this month.
50%UnverifiedBeam introduces a controllable length penalty into the reinforcement learning reward function, penalizing unnecessary token generation while rewarding successful task completion.
50%UnverifiedBeam is explicitly optimized for three scenarios—software engineering, reasoning and long-running agentic tasks—rather than being a general-purpose chat model.
50%UnverifiedReflection AI is entering the market with an "open weights + cloud hosting" combination rather than pure local deployment or a purely closed-source API.
50%UnverifiedThe vLLM team worked with Reflection AI and Inferact to jointly optimize the serving path, including fused kernels and DFlash.
50%UnverifiedBeam is a low-latency U.S. open-weight reasoning model, mainly serving coding and agent-type tasks.
50%UnverifiedWhen integrating using the model name reflection/beam, after the free period ends it will automatically switch to normal billing with zero downtime; when integrating using reflection/beam-free, service will stop after the free period ends.
50%UnverifiedReflection AI has not yet published detailed benchmark data for Beam.
50%UnverifiedThe Beam model is positioned to let teams run and customize it on their own hardware, infrastructure, or in the cloud.
50%UnverifiedBeam is a token-efficient, frontier-scale model focused on coding and agentic use cases.
50%UnverifiedBeam is the first Reflection AI model integrated into the transformers ecosystem.
50%UnverifiedReflection AI is a Brooklyn-based startup founded by former Google DeepMind researchers Misha Laskin and Ioannis Antonoglou.
50%UnverifiedKepler's core team members come from companies such as Palantir, Meta, Goldman, and Reflection AI, among them Sarah, who was one of the engineering leads for Palantir Ontology, and Chris, who worked at Palantir for 15 years focusing on optimization problems.
50%UnverifiedReflection AI was founded by two former Google DeepMind researchers, its models focus on coding and agent directions, and are optimized for Nvidia chips.
50%UnverifiedReflection AI was founded by former DeepMind researchers.
50%UnverifiedReflection AI, as a newer participant, chose the open-source route rather than closed-source.
50%