Reflection AI trained its open model Beam to talk less and save compute

Usually, an AI is trained to find the right answer. Reflection AI also penalized it whenever it talked too much. “We simply penalize it on the number of tokens it can use,” co-founder Ioannis Antonoglou sums up in VentureBeat. The result is called Beam: the American start-up published its weights on Hugging Face on October 16, under the Apache 2.0 license.

It is Reflection's first open model. It is a mixture-of-experts (MoE) model with 501 billion parameters, 23 billion of which are active for each token. It is specialized in code and agents, accepts up to one million tokens of context and handles text only. The Apache 2.0 license allows commercial use.

Less compute for a comparable score, according to the developer

The promise is efficiency more than raw power. According to Reflection, Beam reaches scores comparable to GLM-5.2, Z.ai's model, with three to four times less compute on reasoning tests. The figures published on Hugging Face come from Reflection's own team: 81.6% on SWE-bench Verified and 80.1% on Terminal-Bench 2.1, at BF16 precision. Hugging Face's teams ran their own tests, on five tasks, but warn that these are impressions, not measurements (“vibe evals, not benchmarks”).

On Terminal-Bench 2.1, Beam trails DeepSeek V4.1 Flash (90.6%) and Moonshot's Kimi K3 (85%), according to Reflection's tables and Meta's model card, as cited by VentureBeat. It beats Meta's Muse Spark Open (76.4%), which had already billed itself as the best American open model. As for outside evaluations, Artificial Analysis says only, according to VentureBeat, that early indicators place Beam among the most token-efficient open models at an equivalent level.

Fewer tokens, a smaller bill?

Not necessarily, and Reflection acknowledges it. Its compute estimate is based on the number of active parameters and does not count the reading of instructions, which are often very long in agent and coding uses. Its comparison chart leaves out DeepSeek, known to be very efficient. Joe Spisak, vice president of product, admits that other models “could be much more efficient” thanks to their low-level optimization. He also acknowledges having “nothing concrete to share” on cost per task. Reflection's hosted service is free during its public testing phase, with no announced pricing.

What it takes to run it

An open model does not mean a lightweight model. While the compute is that of a 23-billion-parameter model, the memory is that of a 501-billion-parameter model. According to Hugging Face:

• the BF16 version weighs about 1 TB;
• the NVFP4 version weighs about 280 GB, with four B200 cards recommended. This format works natively only on NVIDIA's Blackwell-generation cards: older cards, such as the H200, must use the FP8 version;
• the model is supported at launch by transformers, vLLM and SGLang, and is available through several hosting providers (Baseten, Fireworks, Together).

This is a far cry from NVIDIA's Nemotron, which fits on a single PC graphics card. Beam targets companies that want to self-host a serious coding model under an unrestricted license. It remains to be seen whether independent evaluations will confirm the announced savings, and what a completed task will really cost.