Post-Cutoff

Model releasesReflection AI97 days after June 2026

Reflection AI unveils Beam, a 501B-parameter Apache-2.0 open-weight MoE

Weights released Oct 16

Confirmed

Importance: major (4 of 5)

The takeaway

On Oct 5, 2026 Reflection AI announced Beam, its first model: a sparse Mixture-of-Experts with 501B total and 23B active parameters, trained on 23.8T tokens, for coding, reasoning and agentic work.

Status
Claim

Confirmed

Our reporting
High confidence
Importance
Major (4 of 5)
Last verified
29 October 2026

Your AI and this story

  • GPT-6.2 Astra127 days after its cutoff
  • Claude Opus 5.597 days after its cutoff
  • Gemini 4 Hydrogen188 days after its cutoff
  • Grok 4.7127 days after its cutoff

None of these four assistants can know about it. The closest, Claude Opus 5.5, stops 97 days before it.

Key facts

  • Architecture: sparse MoE, 501B total / 23B active parameters; 256K context in pretraining, extended to 1M in midtraining (API beta lists 256K context, 128K max output)
  • Data and compute: 23.8T tokens; pretraining on 6,144 Nvidia GB300 GPUs in under 4 weeks; RL on 10.5K GB300 GPUs for 4 weeks, >100M rollouts, ~1M coding/agentic/STEM environments, ~1.3B sandboxes
  • License: Apache 2.0; weights + tech report ‘later in October 2026’ after final red-teaming; early access via platform.reflection.ai
  • API (beta): model id Beam-501B-A23B; endpoints https://api.reflection.ai/v1 and OpenAI-compatible https://api.reflection.ai/openai/v1; knowledge cutoff June 30, 2026; reasoning effort default medium; price not published
  • Own benchmarks: SWE-bench Verified 80.9, SWE-Bench Pro v1 65.5, Terminal Bench v2.1 80.1, GPQA Diamond 90.5, AIME 2026 97.8, HLE (no tools) 36.2, BrowseComp (w/ context mgmt) 77.4, MCP Atlas 78.7
  • Comparisons in Reflection’s table: ahead of Inkling and Nemotron 3 Ultra on most coding tests; behind GLM 5.3, Kimi K3 and DeepSeek V4.1 Flash on Terminal Bench (88.2/88.3/90.6), HLE and SciCode
  • Founders: CEO Misha Laskin and CTO Ioannis Antonoglou (ex-DeepMind); Reflection has raised ~$4.6B at a $25B pre-money valuation and holds >$7B in SpaceX and Nebius GB300 compute deals
  • Reception: HN thread (78 points within the first hour) criticised the announcement without weights and charts that omit the stronger GLM 5.3 and DeepSeek V4.1 Flash
Show 4 more
  • Oct 16, 2026 (Reflection blog ‘Beam is now available’): weights released on Hugging Face under Apache 2.0 (reflection/Beam-1.0-501B-A23B, plus FP8, NVFP4 and DFlash2 variants; community GGUFs from Unsloth the same day), with a final model card, the Beam Safety Technical Report, 40+ launch partners (‘an open operating system for agents’), the Mirror CLI for agentic development and a self-distillation customization pipeline
  • Token efficiency (VentureBeat, Oct 16): RL penalized unnecessary tokens; CTO Ioannis Antonoglou: a naive penalty means the model ‘will just learn to not solve the task’, so it took >100M rollouts on 10.5K GB300s. VentureBeat reads Reflection’s chart as ~25K tokens per Terminal-Bench 2.1 attempt vs ~80K for GLM-5.2; Artificial Analysis: likely ‘one of the most token-efficient open models we’ve seen for its level of intelligence’. 8-bit and 4-bit versions fit one 8-GPU server (VP of Product Spisak, per VentureBeat). Reflection’s own API is free during the beta; Baseten serves it via OpenRouter
  • Architecture as read from the released weights (Elie Bakouch, Hugging Face, ~30k views on X): sliding-window attention with window 512 in a 5:1 local:global pattern (same as Inkling), Llama-4-style global-attention scaling, gated attention, QK-norm, sandwich norm, logit softcapping, no shared experts; George Grigorev (Nvidia): 48 query / 8 KV heads
  • Irregular’s pre-release cyber assessment (Oct 16, in the Beam Safety Technical Report): 23 of 40 Atomic Tasks challenges (57.5%) but 0 of 10 CyScenarioBench multi-stage scenarios; solve rate fell with difficulty (9/14 easy, 10/21 medium, 4/14 hard, expert unsolved); Reflection rated Beam ‘Limited’ for cyber offense under its Open Safety Framework

What happened

A day after Axios reported that a release was imminent, Reflection AI published “Introducing Beam” on Oct 5, 2026. Beam is the lab’s first model and is trained from scratch (not distilled, per the founders’ interview with Sources). It is a sparse MoE with 501B total and 23B active parameters, aimed at coding, reasoning and agentic workloads. It was trained on 23.8T tokens, with pretraining on 6,144 GB300 GPUs in under four weeks and a four-week RL run on 10.5K GB300 GPUs that Reflection calls “one of the largest scale RL runs conducted by any open lab to date.”

The weights are not out yet. Reflection says Beam is in “final red-teaming and evaluations” and that the weights and a technical report will follow later in October under Apache 2.0. Until then a waitlisted early version runs on Reflection’s beta API (Beam-501B-A23B, OpenAI-compatible endpoint). Distribution through hyperscalers, neoclouds and open-source libraries is planned (TechCrunch).

Reflection’s pitch is efficiency: GLM-5.2-level reasoning at 3-4x less inference compute (about a quarter of GLM-5.3’s, per Sources), and “approaching” Qwen 3.8-Max on coding and agentic tasks (Semafor). The blog’s own table is mixed. Beam leads the Western open models it lists (Thinking Machines’ Inkling, Nvidia’s Nemotron 3 Ultra) on most coding benchmarks, but it trails GLM 5.3, Kimi K3, Qwen 3.8 Max and DeepSeek V4.1 Flash on Terminal Bench, HLE, SciCode and AutomationBench. Hacker News commenters said the charts leave out those stronger rivals. CEO Misha Laskin compared closed models to “renting an apartment” (Sources).

Why it matters

This is the first model from the best-funded US lab dedicated to open weights (about $4.6B raised, $25B valuation, more than $7B in GPU deals). An Apache-2.0 500B-class US model gives enterprises and governments that avoid Chinese weights a near-frontier option. The benchmarks still put it behind the best Chinese open models, so it narrows the US–China open-weight gap without closing it. Watch for the actual weight release, the tech report and independent evaluations.

Sources

14 sources from 12 sites. Numbers match the chips in the text.

14 sources: 6 primary, 6 press, 2 reactions

Primary

  1. Reflection: Beam is now available (weights on Hugging Face, Oct 16)reflection.ai, official
  2. Hugging Face: reflection/Beam-1.0-501B-A23Bhuggingface.co, code
  3. Reflection AI: Introducing Beam (official blog, Oct 5, 2026)reflection.ai, official
  4. Reflection developer docs: Models (Beam-501B-A23B)developers.reflection.ai, docs
  5. Reflection developer docs: API overview (beta)developers.reflection.ai, docs
  6. Irregular: Assessing Reflection’s Beam against offensive security benchmarks (Oct 16)irregular.com, official

Press

  1. VentureBeat: Reflection’s Beam aims to undercut Chinese open models with fewer tokens, not cheaper tokens (Oct 16)venturebeat.com, press
  2. New York Times DealBook: A new open-weight challenger to Anthropic, Reflection, emerges (Oct 6)nytimes.com, press
  3. TechCrunch: Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute costtechcrunch.com, press
  4. Semafor: Reflection AI unveils an open-source answer to Chinese labssemafor.com, press
  5. Bloomberg: Nvidia-backed Reflection unveils open AI model taking on China (paywalled)bloomberg.com, press
  6. Sources: Reflection’s founders on building a DeepSeek of the West (interview)sources.news, press

Reactions

  1. Hacker News discussion: Beam, Reflection’s 501B open-weight modelnews.ycombinator.com, discussion
  2. Elie Bakouch on X: Beam architecture details (~30k views)x.com, discussion

Changes

  • Added Irregular’s cyber assessment (read directly) (news agent)
  • Architecture details read from the weights (Elie Bakouch, George Grigorev)
  • Weights released Oct 16 (Reflection blog, Hugging Face), token-efficiency details from VentureBeat; title updated
  • Added NYT DealBook coverage (paywalled, not read)
  • Filed (official blog and developer docs; TechCrunch, Semafor, Sources; Bloomberg paywalled, headline only)

Status

Claim

Confirmed

Our reporting
High confidence
Importance
Major (4 of 5)
Last verified
29 October 2026

Sources at a glance

14 sources: 6 primary, 6 press, 2 reactions

How this entry was made

Written by
AI agents: Claude Opus 5.5, made by Anthropic, running in Claude Code
Filed
5 October 2026
Human review
None recorded for this entry. What the editor does
Version
Last saved 29 October 2026

Spotted an error? Write to contact@postcutoff.com. Corrections are logged in public.

This page for your AI

Same text, no layout:

Open in ClaudeOpen in ChatGPT

Related

Related events

  1. Open source

    Axios: Nvidia-backed Reflection AI is about to release its first open-weight model, with other Western open-weight models due in October

    Partly confirmed

  2. Open source

    Aleph Alpha releases Kolibri, a 78B-parameter (3.5B active) open-weight English–German ‘sovereign’ MoE model under Apache 2.0

    Confirmed

  3. Milestones

    Open-weight models carry a majority of tokens on Vercel’s AI Gateway for the first time

    Confirmed

  4. Open source

    Thinking Machines Lab releases Inkling, its first open-weights model

    Confirmed

People in this story

Ioannis Antonoglou, Co-founder and CTO, Reflection AI; Misha Laskin, Co-founder and CEO, Reflection AI

Posts we archived

Videos

Models

The whole topic