ReflectionOpen SourceCodingReasoning
Beam 1.0
Reflection’s Beam 1.0 — Apache-2.0 text-only sparse MoE (501B total, 23B active, 1M context) for agentic coding and reasoning. Weights shipped 2026-10-16 with day-0 support in Hugging Face Transformers, vLLM, and SGLang. This is not the beam.cloud GPU tool.
Tool calling · Thinking · Coding
Last reviewed: 18 October 2026
Overview
Reflection’s Beam 1.0 — Apache-2.0 text-only sparse MoE (501B total, 23B active, 1M context) for agentic coding and reasoning. Weights shipped 2026-10-16 with day-0 support in Hugging Face Transformers, vLLM, and SGLang. This is not the beam.cloud GPU tool.
Capabilities
- Vision: No
- Audio input: No
- Tool calling: Yes
- Thinking: Yes
- MCP: No
- Coding: Yes
- Structured output: No
Technical specifications
- Provider
- Reflection
- License
- Apache-2.0
- Context window
- 1M
- Parameters
- 501B total / 23B active
- Architecture
- Sparse MoE, text only
- Release
- 2026-10-16
- Modalities
- Text
- Vision
- No
- Audio input
- No
- Tool calling
- Yes
- Thinking
- Yes
- MCP
- No
- Open weights
- Yes
- API
- No
Supported modalities
Text
Context window
1M (1,000,000 tokens)
Pricing
No first-party token price was published with the weight release.
Availability
API: No
Chat UI: No
Open weights: Yes
Regions: Global
Apache-2.0 weights. Hugging Face, vLLM, and SGLang support landed with the release. The slug reflection-beam is used so it does not collide with the beam-cloud GPU tool.
Use cases
- Self-hosted coding agents
- Long-context text workloads
- Teams that need an Apache-2.0 MoE
Strengths
- Apache-2.0 weights
- 1M context
- 23B active parameters
- Day-0 serving in Transformers, vLLM, and SGLang
Limitations
- Text only
- New lab with a thin ecosystem
- No first-party token price published with the weights
Related guides
Related benchmarks
Related research
- Introducing Beam
Original Paper
- Beam: an open large language model
Original Paper
Related tools
Companies
Explore more models
- DeepSeek R1DeepSeekDeepSeek’s reasoning-focused model trained with reinforcement learning for multi-step math, science, and coding problem solving.
- Sarvam 105BSarvamSarvam’s flagship open-weight reasoning model. A Mixture-of-Experts transformer with 10.3B active parameters, trained from scratch in India, and used to power the Indus assistant.
- Sarvam 30BSarvamSarvam’s efficient open-weight reasoning MoE. About 2.4B active parameters, built for real-time conversational agents on the Samvaad platform and for local or GPU-constrained deployment.
- DeepSeek V3DeepSeekDeepSeek’s MoE general model — strong open-weight performance on coding and knowledge tasks with competitive API pricing.
- DeepSeek V4DeepSeekDeepSeek’s V4 generation — live Flash is V4.1-Flash-1017 (API id deepseek-flash, upgraded 2026-10-17): Causal Encoder–Decoder MoE, 552B backbone (8B active prefill / 16B decode), native image+text, and 1M context. API id and prices are unchanged from the 2026-09-10 Flash list. MIT weights on Hugging Face. Legacy deepseek-v4-flash and deepseek-v4-flash-vision-exp aliases temporarily route here. deepseek-v4-pro still serves V4-Pro-0813 after 2026-09-14 at unchanged Pro rates (DeepSeek withdrew the planned Flash reroute).
- Kimi K3Moonshot AIMoonshot’s Kimi K3 — 2.8T MoE (104B active) open-weight multimodal agentic model with 1M context, native vision, and strong long-horizon coding. Weights on Hugging Face under the Kimi K3 License.