# Update: Reflection Shows Beam — 501B / 23B Active, Apache 2.0 Weights Later This Month

Times of AI Desk · 2026-10-04 · Models

[https://timesof.ai/2026/10/reflection-open-weight-imminent](https://timesof.ai/2026/10/reflection-open-weight-imminent)

> Reflection published Beam on Oct. 5: a 501-billion-parameter sparse MoE (23B active) open-weight text model it says approaches GLM-5.2 on coding and reasoning with 3–4× less estimated inference compute. Weights, a technical report and a model card are promised later this month under Apache 2.0. Benchmarks are Reflection’s own; Kimi K3 and GLM-5.3 still lead on most shared tasks.

**Corrected October 14, 2026:** Removed claims that Beam was not on Artificial Analysis and an unsupported 'most serious U.S. attempt yet' superlative; noted that the comparison table is vendor-compiled.

**Updated October 6, 2026:** Reflection showed its cards. On **October 5** the company published **Beam** — confirming the [Axios scoop](https://www.axios.com/2026/10/04/reflection-open-weight-ai) that sources said an open-weight release was “soon,” without a name, size, license or date.

**Reflection AI**’s first open-weight model is a sparse mixture-of-experts text system: **501 billion** total parameters, **23 billion** active, pretrained on **23.8 trillion** tokens, with a **1M-token** context from midtraining. Reflection pitches efficiency, not a capability crown. On its own tables Beam sits near **GLM-5.2** on several coding and agentic tests and behind **GLM-5.3** and **Kimi K3** on most shared tasks. Weights, a technical report and a model card are due **later this month** under **Apache 2.0**. Early access is waitlist-only while final red-teaming finishes.

## What Reflection shipped

Pretraining finished in under four weeks on **6,144 Nvidia GB300** GPUs. The high-compute RL campaign used **10.5K GB300s** for four weeks — Reflection says **over 100 million** rollouts and roughly **1.3 billion** sandboxes. Beam remains **text-only**; Thinking Machines Lab’s **Inkling** is multimodal, so head-to-head coding wins in Reflection’s tables are not a full product comparison.

Sample self-reported scores (Reflection’s table; competitor numbers attributed via Artificial Analysis / DataCurve where Reflection cites them; vendor-compiled, so harnesses may differ):

| Bench | Beam | GLM-5.2 | GLM-5.3 | Kimi K3 |
| --- | ---: | ---: | ---: | ---: |
| Terminal-Bench 2.1 | 80.1 | 81.0 | 88.2 | 88.3 |
| SWE-Bench Pro v1 | 65.5 | 62.1 | — | — |
| GPQA Diamond | 90.5 | 91.2 | 91.7 | 93.5 |
| HLE (no tools) | 36.2 | 40.5 | 42.3 | 46.9 |

Reflection says Beam is “approaching **Qwen 3.8-Max**” and that **Kimi K3** remains ahead on raw capability. The **3–4× less inference compute** claim versus GLM-5.2-class reasoning is Reflection’s estimate: **FLOPs ≈ 2 × active parameters × mean generated tokens**, counting multiply-add as two ops. The company notes the estimate **excludes prompt prefill, context-dependent attention and serving overhead** — approximate compute comparison, not measured serving cost.

## Money and the factory pitch

**TechCrunch** (Rebecca Bellan, Oct. 5) says Reflection has raised about **$4.7 billion** (PitchBook) at a **$25 billion** pre-money valuation, and has **over $7 billion** in compute deals with **SpaceX** and **Nebius** for Nvidia GB300 access through 2029. That supersedes Axios’s adjective-only “massive” Nebius/SpaceX compute line from Oct. 4. The commercial frame is unchanged: sovereign **AI factories** that train Reflection weights on a buyer’s data and compute — with a **Shinsegae** pilot already announced in South Korea.

## Claims vs checks

All benches above are **vendor-reported**. TechCrunch notes the claims are **not independently verified**. Until weights and third-party evals land, this is a scorecard with a release promise, sold on efficiency rather than a lead.

## Limits

- Benchmarks and the 3–4× compute claim are **Reflection’s**. Competitor scores in Reflection’s tables are sourced from AA/DataCurve per Reflection, not re-run here.
- Weights are **not public**. Apache 2.0 release is “later this month,” not a dated GA.
- Funding and compute-deal figures follow **TechCrunch / PitchBook** wording; outside dollar figures that conflict with that wording are not used.
- Beam is text-only; multimodal peers are not like-for-like.
- AA Intelligence Index: Beam not listed as of Oct. 6.

## Sources

- [Reflection AI: Introducing Beam: Reflection’s 501B open-weight model (October 5, 2026)](https://reflection.ai/blog/introducing-beam)
- [TechCrunch: Reflection debuts Beam, an open-weight AI model to rival Chinese models at lower compute cost (October 5, 2026)](https://techcrunch.com/2026/10/05/reflection-debuts-beam-a-open-weight-ai-model-to-rival-chinese-models-at-lower-compute-cost/)
- [Axios: Scoop: Powerful open model is set to shake up AI race (Bradley Olson, October 4, 2026)](https://www.axios.com/2026/10/04/reflection-open-weight-ai)