Image: Reflection AI

Reflection AI's Beam is finally out in the open, and the first independent report card is in. Benchmarking firm Artificial Analysis gives it 30 on its Intelligence Index, making it the second-smartest open model built in the US. That's good, but not the "Western open frontier" leap Reflection promised.

When Reflection unveiled Beam earlier this month, the weights weren't available yet and every number came from the company. Now anyone can download Beam on Hugging Face under the Apache 2.0 licence, which allows commercial use.

Beam is now available across platforms, with open weights under the Apache 2.0 license - in collaboration with an ecosystem of 40+ partners.

We're releasing:
- Weights, tokenizer, and reference implementation
- BF16, FP8 and NVFP4 weights and a DFlash2 drafter
- The open source… pic.twitter.com/tHYv4cuYIh

— Reflection (@reflection_ai) October 16, 2026

Beam is a 501-billion-parameter model, but only about 23 billion of those do the work for each word it writes. That design keeps running costs down. Reflection is shipping three versions, from full precision down to a compressed NVFP4 version that needs less memory. It's also releasing Mirror CLI, the coding tool it trained Beam with.

The company says more than 40 partners are on board at launch, including AWS, Google Cloud, Microsoft Azure, Nvidia, Ollama and LM Studio. It describes the bundle as "an open operating system for agents."

So where does Beam actually stand? Among US open models, only Meta's Muse Spark Open (39) scores higher, according to Artificial Analysis. Beam beats Nvidia's brand-new Nemotron 3.5 Super VL (26) and Thinking Machines' Inkling (25). Outside China, only Muse Spark Open and the UAE's K2 Horizon (31) are ahead.

Image: Artificial Analysis

China is a different story. The same chart puts Xiaomi's MiMo-V2.6-Pro at 46 and Z.ai's GLM-5.3 at 45. Even Alibaba's far smaller Qwen3.8 27B scores 34. Closed models such as Claude Opus 5.5 (58) and GPT-6 Astra (53) are in another league.

Beam does have real strengths. It is thrifty, using about 26,000 output tokens per test task, half of K2 Horizon's roughly 52,000. That fits Reflection's claim that it needs far fewer reasoning tokens than rivals. It also scored 48.5% on AutomationBench, a test of everyday business workflows, well ahead of K2 Horizon's 37.2%.

The weak spots are just as clear. On Terminal-Bench 4.0, a hard test of working in a command line, Beam managed 3.0%, behind Qwen3.8 27B (5.6%) and Muse Spark Open (8.1%). And while it answered 33% of knowledge questions correctly, when it didn't know an answer, it guessed instead of saying so 47% of the time, against 26% for K2 Horizon.

Here's what stands out when you put the two sets of numbers side by side. Reflection's own model card claims 80.1% on Terminal Bench 2.1, an older version of the test. Artificial Analysis's harder version 4.0 gives 3.0%. The two aren't directly comparable, but the gap shows how much a headline score depends on which test a company picks.

It cuts both ways. Reflection reports 39.3% on the public part of AutomationBench, lower than the 48.5% Artificial Analysis measured. Its card also scores Beam at 10.8 on Artificial Analysis's Omniscience knowledge index, using public questions only. The full independent run puts it at 1.5.

Two caveats. Reflection lists Artificial Analysis among the partners that tested Beam before release, though the scores are Artificial Analysis's own measurements. And Artificial Analysis hasn't measured Beam's speed or price yet because it found no public API, so real-world costs are still unknown.

For businesses that want a US-made model they can run on their own servers, Beam is now one of the best options. For anyone who just wants the strongest open model, China still holds the lead.

Reflection’s first model, Beam, scores 30 on the Artificial Analysis Intelligence Index, placing it #2 among US open weights models

Beam is an open weights model with 501B total and 23B active parameters, released on @huggingface

Congratulations to @reflection_ai on the… pic.twitter.com/s8cIi1UtMN

— Artificial Analysis (@ArtificialAnlys) October 16, 2026