Beam’s model technical report is now available. Read the full report to learn about the methodology our team used to train Beam end-to-end in-house, here: reflection.ai/documents/beam…
We developed the Iris Safety Evaluations (ISE) to test whether Beam’s responses align with and uphold our safety policy in realistic contexts, such as in adversarial and agentic settings. Evaluated across cyber offense, chemical, biological, and radiological/nuclear categories, Beam holds higher refusal rates than leading Western and Chinese open models.
Sharing more about how we trained and evaluated Beam, our first model, for safety: reflection.ai/blog/open-mode…
In the coming weeks, we'll release additional safety research, evaluations, and tools under Project Iris, Reflection's open AI safety initiative.
We are committed to prioritizing safety in every model we build—and sharing what we develop and learn so that researchers, developers, policymakers, and enterprises can build on our work, advancing safety across the broader AI ecosystem.
Exclusive @thetimes interview with @reflection_ai co-founders @MishaLaskin@real_ioannis
We cover:
- How their new open model Beam could catch up with the capabilities of closed models in the next year
- Why renting intelligence from a closed model provider is like renting from a landlord who could “hike up your rates or evict you”
- AI is not a "winner-takes-all" market
- Plans for a big expansion in London
- Would they restrict governments from using their models for autonomous weapons and mass surveillance?
- Is there an AI bubble?
thetimes.com/business/techno…
Beam 1.0 did especially well on EMB, which tests agents on building Excel financial models. Beam ranks #14 among all open-weight models and #1 among US open-weight models, just ahead of Muse Spark Open.
Just watched the @netflix Instadocs about the @OpenAI cyberattack on @huggingface. Super disappointing, as they went for maximum anthropomorphism of the technology and full doomer mode.
I spent more than two hours with them trying to correct all their mistakes and ground this in reality, explaining that agents are technological systems, not aliens. But it looks like the drive to scare people for maximum views was too strong.
Also, they didn’t mention the lack of containment and monitoring of the agents, how closed-model safeguards blocked our defensive work, or how we defended ourselves with AI using an open model. And that we still lack a lot of information and transparency, including full agent traces and that companies need to remain accountable for what their agents do.
It ends up being a fun sci-fi movie though!
Having stared at a fair amount of graphs over the last two weeks, I had somehow missed this one. Thanks @EthanCastr for pointing it out.
Beam is optimized for token efficiency and this graph shows this. Keep building!
Beam is now available across platforms, with open weights under the Apache 2.0 license - in collaboration with an ecosystem of 40+ partners.
We're releasing:
- Weights, tokenizer, and reference implementation
- BF16, FP8 and NVFP4 weights and a DFlash2 drafter
- The open source harness we built and trained Beam on: Mirror CLI
Weights and model card on Hugging Face: huggingface.co/reflection
Beam weights are out today (Apache 2.0 license).
We integrated Beam across 40+ different partners.
And are also open sourcing Mirror CLI - the harness Beam was trained against.
Happy building.
Beam is now available across platforms, with open weights under the Apache 2.0 license - in collaboration with an ecosystem of 40+ partners.
We're releasing:
- Weights, tokenizer, and reference implementation
- BF16, FP8 and NVFP4 weights and a DFlash2 drafter
- The open source harness we built and trained Beam on: Mirror CLI
Weights and model card on Hugging Face: huggingface.co/reflection
Beam is now available across platforms, with open weights under the Apache 2.0 license - in collaboration with an ecosystem of 40+ partners.
We're releasing:
- Weights, tokenizer, and reference implementation
- BF16, FP8 and NVFP4 weights and a DFlash2 drafter
- The open source harness we built and trained Beam on: Mirror CLI
Weights and model card on Hugging Face: huggingface.co/reflection
A fun first community event last night at NeueHouse to kick off the post-Beam announcement era. Met people ranging from compiler engineers to investment bankers (former SWE), and a mix of researchers, devrel-ers, and FDEs in between. Looking forward to continuing some of the conversations from last night on the Reflection Discord.
And a big shout out to @zkwentz for crushing it on the panel. Dropped some real-talk tales of the scientific rigor it took for @reflection_ai to build a frontier model, and the realities of what massive scale RL post-training looks like.
NYC showed up!
Together AI × NVIDIA × Reflection. A full room of builders, great conversations, and a shared excitement for what’s next in open AI.
Thanks to everyone who joined us tonight.
Berlin and San Francisco up next. This community keeps growing, and we’re just getting started.
Find your city. Come build with us: luma.com/user/together_ai