Reflection releases Beam: 501B parameters and 30 points on the Intelligence Index

Reflection has entered the market with a model that independent tests place near the top of American open systems. Beam is an open-weights model with 501 billion parameters in total and 23 billion active per token. It scored 30 on the Artificial Analysis Intelligence Index. Among American open-weights models, only Meta Muse Spark Open is higher, at 39. NVIDIA's Nemotron 3.5 Super VL, released a day earlier, scored 26. Among all open models, Beam ranks tenth, one point behind K2 Horizon from the UAE (31). Outside China, only these two models rank higher.

Few tokens per unit of intelligence

Beam's main distinction is not its absolute score but the price the model pays for it. On a single index task it spends about 26k output tokens: roughly 19k on reasoning and roughly 7k on the answer. K2 Horizon, which leads Beam by one point, uses about 52k, twice as many. Closed models still set the bar: GPT-6 Astra (max) scores 53 on a budget of about 27k tokens, nearly the same as Beam's.

The final answer makes up only 26% of Beam's output. That ranks 28th from the bottom among the 219 models tested on Intelligence Index v4.3. The model thinks a lot and says little. In the Cyber Index, where it solves 34% of tasks, that puts it on the "score vs. output tokens" frontier next to GPT-6 Sol (Daybreak Blue). Among the American open-weights models tested, Beam leads this index.

Strong agents, weak terminal

The picture across individual benchmarks is uneven. Key figures:

• AutomationBench-AA: 48.5%. K2 Horizon scores 37.2%, Qwen3.8 27B 48.2%.
• CritPt: 13.1%. That is more than double K2 Horizon (4.6%) and Qwen3.8 27B (5.4%).
• Terminal-Bench 4.0: 3.0%. Qwen3.8 27B gets 5.6%, Muse Spark Open 8.1%.
• AA-Briefcase: 976 Elo, another weak spot.
• Context window: 1M tokens. Text-only modality.

Coding results are more modest. Paired with Mirror, Reflection's own open CLI harness, Beam scores 28 on the Coding Agent Index. It trails the next configuration, Claude Code with Qwen3.8 Max (43), by 15 points, and the leaders (68) by 40. On token usage, though, the model is frugal: only Codex with GPT-6 Astra and Codex with GPT-6.1 Sol spend fewer.

Knows more, but also makes things up more often

In AA-Omniscience, Beam answers 33% of questions correctly. K2 Horizon scores 18% and Qwen3.8 27B 16%. The price of that broad knowledge is high: Beam's hallucination rate is 47%, versus 26% for K2 Horizon and 30% for Qwen3.8 27B. The model says "I don't know" less often and is more often confidently wrong. Its overall index score is 1.5, which is still better than its competitors' (−3.0 and −10.0). The gain in accuracy outweighs the penalty for fabrication, but the margin is small.

Who Beam suits in practice

Beam is a candidate for business-process automation and multi-step agentic scenarios where token cost and control over the weights matter. Spending half the output tokens of K2 Horizon directly lowers the inference bill at high volumes. With only 23 billion active parameters, serving load is lower than for dense models of comparable size, but you still have to store all 501 billion.

For terminal tasks and autonomous coding, Beam currently looks weaker than its rivals. If you are building an assistant that answers factual questions without checking sources, a 47% hallucination rate is a serious risk, and you need retrieval or answer verification. A sensible plan: use the model in agentic pipelines with external checks, and leave coding to stronger combinations. Mirror deserves a separate look: the open harness lets you reproduce Reflection's results on your own infrastructure and compare them with what the independent test showed.