Reflection's first model Beam lands second among US open weights: 30 on the Intelligence Index at half a rival's output tokens

A new second place has appeared among US open weights models, from Reflection, which had not previously released a model. That position had been held by models released through the week, with Meta's Muse Spark Open at 39 and NVIDIA's Nemotron 3.5 Super VL at 26. Beam scores 30 on the Artificial Analysis Intelligence Index, above Nemotron 3.5 Super VL at 26 and below Muse Spark Open at 39; outside China, only Muse Spark Open and the UAE's K2 Horizon (31) score higher.

Beam reaches its intelligence level with fewer output tokens, spending about half the generation to land near K2 Horizon. K2 Horizon sits one point ahead but uses about 52k output tokens per Intelligence Index task, twice Beam's figure. Beam uses about 26k output tokens per Intelligence Index task, roughly 19k for reasoning and 7k for answers, with final answers only 26% of output, the 28th-lowest share of 219 models tested on Intelligence Index v4.3.

Its capability profile is uneven, strong on automation tasks and clearly weaker in the terminal. On AutomationBench-AA, Beam leads K2 Horizon by 11 percentage points and is level with Qwen3.8 27B; on Terminal-Bench 4.0 it trails both. Beam scores 48.5% on AutomationBench-AA against 37.2% for K2 Horizon and 48.2% for Qwen3.8 27B; on Terminal-Bench 4.0 it scores 3.0% against 5.6% for Qwen3.8 27B and 8.1% for Muse Spark Open.

It answers knowledge questions correctly more often while fabricating answers more often, and its net score still leads comparable open models. K2 Horizon and Qwen3.8 27B answer correctly less often but hallucinate less. On AA-Omniscience Beam answers 33% correctly against 18% for K2 Horizon and 16% for Qwen3.8 27B; its hallucination rate is 47% against 26% and 30%, giving a net index score of 1.5 versus −3.0 and −10.0.

This report is for researchers and engineers tracking model evaluation and deployment cost, and covers published results on the Artificial Analysis Intelligence Index and its components. Beam is an open weights model with 501B total and 23B active parameters, a 1M-token context window and text modality. It completes 34% of tasks on the Artificial Analysis Cyber Index, the highest among US open weights models tested, and scores 28 on the Artificial Analysis Coding Agent Index when running in Mirror, Reflection's open-source CLI harness.

Beam's hallucination rate is 47%, above K2 Horizon's 26% and Qwen3.8 27B's 30%, and its net index score of 1.5 still leads after that deduction. Terminal tasks are a weak point, with 3.0% on Terminal-Bench 4.0 and 976 Elo on AA-Briefcase. On the Coding Agent Index it trails the next setup by 15 points and the leaders by 40. Proprietary models still set the frontier, with GPT-6 Astra (max) scoring 53 on the same token budget.