Reflection Beam: a 501B open-weight model for coding agents

What is Beam?

Beam is Reflection's first model. The company announced it as a preview on October 5 and has now released the weights (Reflection). It is text-in, text-out and was developed predominantly for English. Its knowledge cutoff is June 30, 2026 (model card).

According to the Hugging Face blog, the model has 52 layers: 3 dense and 49 MoE. Each MoE layer holds 256 experts, of which 8 run per token. Deployment therefore needs the memory of a 501B model, while compute per token is that of a 23B model (Hugging Face).

Long context is handled by hybrid attention. 43 layers attend only to the previous 512 tokens, while 9 layers see the full context. By the blog's calculation, the KV cache at 1M tokens is about 36 GiB per sequence in BF16, versus 208 GiB if every layer were global.

Benchmarks

Among the BF16 results provided by Reflection, the most relevant for developers are SWE-bench Verified 81.6%, Terminal Bench v2.1 80.1%, SWE Bench Pro v1 66.1% and LiveCodeBench 88.3%. On reasoning, Beam scores 90.6% on GPQA Diamond and 97.7% on AIME 2026. Humanity's Last Exam (HLE) comes in at 36.0% without tools and 51.1% with tools. For tool calling, MCP Atlas is 80.8% and BrowseComp with context management is 76.8%.

Source: Hugging Face / Reflection. Beam benchmark results in BF16, FP8 and NVFP4 precision; figures provided by the Reflection team.

The quantized versions lose little: the NVFP4 checkpoint scores 80.7% on SWE-bench Verified and 78.2% on Terminal Bench. These are the vendor's own measurements; there is no independent verification yet.

In its October 5 announcement, Reflection also compared Beam with other models, and it says itself that Kimi K3 remains ahead on raw capability. On Terminal Bench v2.1, for example, GLM 5.3 scored 88.2 and Kimi K3 88.3 (Reflection). The company's main argument is efficiency: it claims Beam matches GLM-5.2 on reasoning tasks with 3-4x less inference compute. We covered where the Chinese field stands in our earlier overview.

What hardware does it need?

The weights come in three formats. The BF16 checkpoint is about 1 TB. FP8 targets Hopper and Ampere GPUs. NVFP4 is about 280 GB and runs natively only on the Blackwell generation.

Hugging Face says two B200s (2×180 GB) can just hold the NVFP4 weights, but leave little room for the KV cache, so it recommends four. With SGLang, the FP8 version runs on 8 H200s, and the vLLM recipe spreads BF16 across 16 GPUs. This is a model for a server rack, not a desktop.

Without your own hardware there are several routes: Reflection's own OpenAI-compatible API, Hugging Face Inference Providers (Baseten, Fireworks, Together) and Inference Endpoints. In the cloud, AWS SageMaker and Google Cloud Model Garden already offer it; Reflection says Azure and Oracle support is coming soon (Reflection Docs). The sources we reviewed do not list prices.

Customization and fine-tuning

The open license is the point here. According to the model card, Reflection does not prohibit fine-tuning, quantization, distillation or even using the model's outputs to train other models.

Hugging Face published a LoRA SFT recipe with TRL that runs on a single node of 8 B300s at about 19,000 tokens per second. In the example, the model was taught to reason in the language the user asked for. After one epoch it did so on 16 of 16 held-out prompts, versus 3 of 16 for the base model.

For reinforcement learning, Reflection and RadixArk built a GRPO recipe for the Miles framework. Full-parameter GRPO needs 96 H200s; the LoRA version needs 32. Reflection says NVIDIA NeMo RL, Unsloth, Axolotl and SkyRL also support the model.

Alongside Beam, Reflection released Mirror CLI, a coding agent built on a Rust core that was trained together with the model. Beam also works with the Pi, OpenCode and Hermes harnesses. NVIDIA is moving in a similar direction: Nemotron 3.5 Super VL is also an open MoE model for agents, at a much smaller size.

Who should take a look?

Beam is interesting for teams that want to run a large coding model on their own infrastructure for privacy or sovereignty reasons, without license restrictions getting in the way. The entry bar is high, though: even the smallest NVFP4 version needs several Blackwell GPUs. For a first test, the API or an inference provider is simpler. Reflection is also offering $10,000 in Hub credits for the best projects built with Beam.