Reflection AI released Beam on October 5. 501 billion parameters total, 23 billion active at inference, pretrained on 23.8 trillion tokens, with a 1-million token context window. Apache 2.0 open weights are promised for later this month. No exact date yet.

Two things set this apart from the usual model launch. First, who built it. Misha Laskin is CEO — he led reward modeling for DeepMind's Gemini project. His co-founder Ioannis Antonoglou co-created AlphaGo. They raised $2B at an $8B valuation in October 2025. Second, the pitch. Reflection isn't positioning Beam on raw benchmark leadership. Their stated claim is 3-4× less inference compute than comparable open-weight frontier models — meaning the same task output for a fraction of the hardware cost.

Per Reflection's own evaluation suite — all benchmarks below are first-party unless otherwise noted — Terminal Bench v2.1 sits at 80.1%. For reference: the September 2026 published leaderboard has Gemini 3.8 Flash at 90.8%, Fable 5.1 at 85.02%, Grok 4.5 at 83.3%, and DeepSeek V4 Flash at 82.7%. Beam's 80.1% puts it roughly fifth. Competitive — not leading.

Source spread

• TechCrunch — Reflection debuts Beam — [builder]. Primary news coverage; all benchmark numbers taken from Reflection's evaluation suite. Notes explicitly that "DeepSeek V4.1 Flash, Kimi K3, and GLM 5.3 beat Beam on most rows."
• The Next Web — Reflection's first open-weight model — [builder]. Focuses on the GLM-5.2 efficiency comparison and the 3-4× compute claim.
• Unite.ai — Reflection AI Unveils Beam — [hype]. Detailed architecture breakdown, training data specs, context window specs.
• SiliconAngle — Reflection AI $2B raise — [builder]. Context on the funding and Reflection's positioning as "America's open frontier lab."

Pros & cons

What's real:

• The founding team is not typical. Laskin and Antonoglou worked at the highest level of RL research for years. Their claimed fluency in efficient training runs isn't a marketing line — it comes from building Gemini and AlphaGo on constrained compute budgets.
• 23.8 trillion training tokens is one of the larger published training-data figures for any open-weight release. Scale of training data matters; publishing the number is more transparent than most.
• Apache 2.0 license commitment is significant if it holds. No commercial restrictions means this is deployable infrastructure from the moment the weights land.
• A 1M context window at 501B total / 23B active is a real capability. The token density story holds for long-document workloads.
• SWE-bench Verified at 80.9% and SWE Pro v2-Hard at 77.2% are strong if independently verified. Those are coding-agent numbers worth taking seriously.

What deserves a side-eye:

• Every benchmark in the launch is first-party from Reflection's own evaluation suite. TechCrunch explicitly notes that Kimi K3, DeepSeek V4.1 Flash, and GLM 5.3 beat Beam on most competitive rows. That's not the framing Reflection leads with.
• Terminal Bench v2.1 at 80.1% is fifth on the September 2026 published leaderboard. Competitive is the right word. "Rivals Chinese models" is doing extra work.
• No API pricing announced. No confirmed open-weight date. For production planning purposes, this model does not yet exist in a usable form.
• The 3-4× inference efficiency claim is the most interesting number in the release and the least sourced. It needs independent verification before it affects any infrastructure decision.
• SWE-bench Verified scores across labs have been converging upward; 80.9% may be reproducible on benchmark tasks but gap against production codebases is worth testing directly.

Samwise's take

What builders need to know

Further reading

• TechCrunch — Reflection debuts Beam — primary news coverage, most neutral framing
• The Next Web — Reflection's first open-weight model — efficiency comparison focus
• Unite.ai — Reflection AI Unveils Beam — detailed architecture and training data specs
• SiliconAngle — Reflection AI $2B raise — funding and founding team context