SHELFCRITTER AI Digest

MODEL

Beam

modeltopic-noteopen-weight

Overview

Beam is Reflection‘s flagship open-weight model — a 501B-parameter sparse Mixture-of-Experts with 23B active per token, aimed at coding, reasoning, and agentic workloads. Weights ship under Apache 2.0 later in October 2026, with a technical report and the full fine-tuning / eval stack. Reflection’s own (approximate) comparison claims ~3–4× less inference compute than Z.ai‘s GLM 5.2 at comparable scores on their internal evals; the architectural source of the delta is the 23B-active count vs GLM 5.2’s ~40B-active on a 753B MoE. The launch marks a Western lab shipping Apache-2.0 weights at Chinese-model compute efficiency, not a cadence shift — the Jul–Sep 2026 500B+ open-weight cohort (Kimi K3 2.8T, Qwen 3.8 Max 2.4T, DeepSeek-V4-Flash 284B) already carries the thread.

Timeline

  • 2026-10-06-AI-Digest — Reflection unveils Beam at 501B/23B active. Weights release under Apache 2.0 later October with full fine-tuning and eval stack. Reflection claims ~3–4× less inference compute than GLM 5.2 on their internal comparison (vendor-approximate, excludes prompt processing, attention operations that scale with context, and serving overhead; no independent benchmark yet).
  • 2026-10-13-AI-Digest — Beam surfaces as context for the Oct 10 FT-sourced NVIDIA–Reflection talks: Reflection’s announcement five days earlier is the proximate trigger for NVIDIA’s reported move to deepen its position in the lab. Company-reported benchmarks reiterated in the digest: SWE Bench Pro v2-Hard 77.2, Terminal Bench v2.1 80.1, SWE Bench Verified 80.9; positioning roughly GLM 5.2 parity (beats GLM-5.2 on 8/13 shared rows), trailing GLM 5.3 and Qwen 3.8 Max. Weights are still pending under Apache 2.0 after red-teaming; the 3–4× less inference compute claim remains self-reported and excludes serving costs. Reframe worth carrying: Beam is the first US frontier-scale reasoning MoE at open weights, not the first US open-weights release.

Key Developments

  1. 501B/23B active sparse MoE: Beam’s active-parameter count (23B on a 501B total) is roughly half of GLM 5.2‘s 40B active on a 753B total — the architectural source of the “cheaper inference” claim.
  2. Apache 2.0 open-weight release: full weights plus the fine-tuning and eval stack, not an API-only model release.
  3. Capital shape: NVIDIA-led $2B round in Oct 2025 at $8B valuation (Sequoia, Lightspeed, DST, 1789, Eric Schmidt as co-investors); a ~$25B pre-money round reportedly closed in April 2026. The SpaceX Colossus 2 GPU-compute contract is a separate operating-expense arrangement, not an equity relationship.