Reflection releases Beam's weights: 501B parameters, Apache 2.0, and benchmarks it graded itself

Reflection published the weights of Beam on Hugging Face on October 16, eleven days after announcing the model with a promise to release them later in the month. The model card lists an Apache 2.0 license, and checkpoints in BF16, FP8 and NVFP4. It is the first model from the Brooklyn-based lab, which TechCrunch reports has raised about $4.7 billion from backers including Nvidia.

The release matters less as another big open model than for its terms. Beam is a frontier-scale U.S. model under a standard permissive license, with no revenue threshold and no custom use agreement. The model card says Reflection does not prohibit distillation, training on outputs, fine-tuning, quantization or merging.

What was released

• Architecture: a sparse mixture-of-experts model with 501.3B total and 23.4B active parameters: 52 layers (3 dense, 49 MoE), 256 experts, 8 routed per token.
• Input and context: text in, text out, with a 1M-token context window. The model card says it was developed predominantly for English, with a knowledge cutoff of June 30, 2026.
• Formats: BF16 (about 1 TB of weights), FP8 for Hopper and Ampere GPUs, and NVFP4 (about 280 GB) for Blackwell, plus a DFlash2 speculative-decoding drafter.
• Tooling: day-0 support in Transformers, vLLM and SGLang, according to Hugging Face's launch post, along with SFT and RL recipes and an open-source Rust harness from Reflection.
• Where to run it: Hugging Face Inference Providers (Baseten, Fireworks and Together), Inference Endpoints, and Amazon SageMaker JumpStart, where AWS says the model was also ported to Trainium3.

The benchmark numbers are Reflection's own

The model card reports, for the BF16 checkpoint, 80.1% on Terminal Bench v2.1, 81.6% on SWE-bench Verified, 66.1% on SWE-Bench Pro v1, 80.8% on MCP Atlas and 90.6% on GPQA Diamond. All of these come from Reflection's own evaluations, run at temperature 0.7. No independent leaderboard has published a full evaluation yet.

Context from other vendors' own figures keeps those numbers in proportion. VentureBeat notes that on Terminal Bench 2.1, DeepSeek V4.1 Flash reports 90.6% and Moonshot's Kimi K3 85.0%, while Meta's Muse Spark Open, released the day before, reports 76.4%. On that self-reported basis, Beam leads Meta's open model on this test but trails the strongest Chinese ones. Reflection itself has said it is not trying to replace frontier models on every task.

Fewer tokens, not cheaper tokens

Reflection's pitch is efficiency per finished task. The model card says Beam reaches comparable solution quality to leading open models with roughly 3x fewer reasoning tokens. Reading Reflection's chart, VentureBeat puts Beam at about 25,000 tokens per Terminal Bench attempt against about 80,000 for Z.ai's GLM-5.2, which scores slightly higher. According to VentureBeat, Artificial Analysis said early indicators point to Beam being among the most token-efficient open models for its level.

Two caveats apply. Reflection has not published a dollar cost per completed task, and its compute estimate counts generation work but leaves out prompt processing, which dominates the long prompts typical of coding agents. Fewer output tokens does not automatically mean a smaller bill.

What changes for developers

For teams choosing an open model for coding agents, Beam adds a permissively licensed option with explicit effort control. The chat template accepts reasoning effort levels from minimal to max, and the model card documentation notes that dropping the model's earlier reasoning from multi-turn history degrades output. Teams that strip reasoning between turns to save context should test that before deploying.

Tool calls use a modified Harmony format with XML-style argument bodies. Hugging Face's post recommends passing the tool schemas to the tokenizer's response parser, so a bare value like 7 is cast to the declared type. Code built around JSON tool-call parsing will need that adapter.

On long context, Hugging Face calculates that the hybrid attention design (43 sliding-window layers of 512 tokens and 9 global layers) grows the KV cache by about 36 KiB per token in BF16, roughly 36 GiB for a full 1M-token sequence. That is about a sixth of what a fully global design would need, which makes very long sessions more practical on owned hardware.

Hardware is still the main barrier. Hugging Face recommends four B200 GPUs for the NVFP4 checkpoint, and VentureBeat reports that full precision needs tens of GPUs. Most teams will try it first through hosted providers. VentureBeat reports that Reflection's own hosted service is free during public beta, with no price set yet.

Open questions

The technical report is listed as coming but has not been published. The Financial Times has reported that Nvidia held talks about buying or investing further in Reflection. VentureBeat reports that Reflection's CTO called this 'baseless rumors', without explicitly confirming or denying talks. The Apache 2.0 grant on weights already published cannot be withdrawn, whatever happens to the company.