The whole dayFri 16 Oct 2026
AI Compass
ArchiveAbout

Open sourceReflectionConfirmed

Story 2 of 10 in the Friday 16 October edition

Reflection releases Beam, a 501B open-weight coding model, under Apache 2.0

Reflection’s first open model, built for coding, reasoning and long agent tasks, is free to download, fine-tune and build on, and launched with more than 40 partners.

Edited by Ashutosh Shrivastava
17 Oct 2026, 10:30 IST · 1 min read

A large dark green dot on the left followed by dotted and solid rings that grow larger towards the right on a cream background
Image: ReflectionReflection’s share image for the announcement

Reflection released Beam’s weights on Hugging Face on 16 October 2026 under the Apache 2.0 licence. Beam is a mixture-of-experts model, which uses only part of its network for each token: 23B of its 501B parameters are active at a time. It works with text only, mainly in English, and has a 1M-token context window.

The weights come in three number formats, BF16, FP8 and NVFP4, with DFlash2, a drafter model that speeds up generation. Reflection trained Beam together with Mirror CLI, its terminal coding agent, now open source and in beta, and has tested it with OpenCode, Hermes and Pi.

Reflection’s own API is in a public beta that is free to start, and Reflection hasn’t published per-token prices. More than 40 partners offer Beam, including Amazon Web Services, Google Cloud, Microsoft Foundry, OpenRouter and Together AI, while Dell and HPE install it on premises.

Artificial Analysis gives Beam 30 on its Intelligence Index, second among US open-weight models. In Reflection’s own comparison, Kimi K3 stays ahead on raw capability.

Beam scores 80.1 on Terminal Bench v2.1, against 81.0 for GLM 5.2, 88.3 for Kimi K3 and 90.6 for DeepSeek V4.1 Flash Max, and 44.4 on DeepSWE v1.1, against 44.0 for GLM 5.2 and 68.0 for Kimi K3
Reflection’s coding comparison of 5 October, revised 8 October; final model-card scores differ slightlyImage: Reflection

Reflection’s case is efficiency: it says Beam scores close to GLM-5.2 on advanced reasoning tests while using 3 to 4 times less inference compute.

Beam reaches about 44% on DeepSWE v1.1 with about 43k generated tokens, where GLM-5.2 needs about 78k, and about 80% on Terminal Bench 2.1 with about 24k, against GLM-5.2’s 78% at about 82k
From Figure 2 of Reflection’s 5 October post: scores against mean generated tokensImage: Reflection

Sources

  1. PrimaryBeam-1.0-501B-A23B model cardhuggingface.co/reflection · Reflection’s weights, licence and model card on Hugging Face
  2. OfficialBeam is now availablereflection.ai · Reflection announcement, 16 October 2026
  3. OfficialBeam model cardreflection.ai · Final benchmark scores and usage terms
  4. OfficialIntroducing Beam: Reflection’s 501B open-weight modelreflection.ai · Reflection, 5 October 2026, with results revised 8 October
  5. On X@reflection_ai on X: Beam is available with open weights under Apache 2.0x.com · Reflection’s launch post
  6. On X@ArtificialAnlys on X: Beam scores 30 on the Artificial Analysis Intelligence Indexx.com · Source of the Intelligence Index score
  7. On X@arena on X: Beam lands in Agent Arena and Code Arena: WebDevx.com
  8. On X@arena on X: Reflection enters the Agent Arena top 15 labsx.com

The rest of the day

Read the whole day

Spotted an error? Email aiforsuccess.in+compass@gmail.com or tag @ai_for_success on X