Head to head

Beam vs Qwen 3.8 Max

0:13

Beam wins 0 of 13 shared benchmarks

Verdict

Qwen 3.8 Max leads on every one of the 13 shared benchmarks, though several gaps are small (GPQA 92.6 vs 90.5, AA-LCR 80.3 vs 79.3, LongBench 66.3 vs 65.5). Qwen 3.8 Max is a 2.4-trillion-parameter model with 95 billion active per token, roughly four times Beam's active size, which is the compute gap Reflection points to.

Benchmark by benchmark

Benchmark	Beam	Qwen 3.8 Max	Difference
DeepSWE v1.1	44.4	51.0	-6.6
SWE-Bench Pro v1	65.5	67.7	-2.2
Terminal-Bench v2.1	80.1	86.6	-6.5
Humanity's Last Exam (no tools)	36.2	43.6	-7.4
SciCode	49.7	52.1	-2.4
CritPt (AA)	16.3	20.0	-3.7
GPQA Diamond	90.5	92.6	-2.1
AutomationBench (public)	37.0	39.8	-2.8
MCP Atlas	78.7	84.5	-5.8
τ³ banking	38.0	55.2	-17.2
AA-LCR (long context)	79.3	80.3	-1.0
LongBench v2	65.5	66.3	-0.8
IFBench	79.7	82.8	-3.1

All scores are from Reflection's announcement (October 8, 2026 revision). Scores for Qwen 3.8 Max are as cited by Reflection from Artificial Analysis and DataCurve.

The two models

	Beam	Qwen 3.8 Max
Developer	Reflection AI	Alibaba (Qwen)
Based in	United States	China
Parameters (total / active)	501B / 23B	2.4T / 95B
Weights	Apache 2.0, promised for later in October 2026	Qwen3.8-Max License (custom)
Released	October 5, 2026	August 3, 2026
How to access	Waitlisted beta API; weights coming	Hugging Face, Alibaba Cloud Model Studio, OpenRouter