Head to head

Beam vs Kimi K3

1:13

Beam wins 1 of 14 shared benchmarks

Verdict

Reflection itself names Kimi K3 as the model that remains ahead on raw capability. Kimi K3 wins 13 of 14 shared benchmarks, often by 10 points or more on agentic coding and search. Beam's only win is τ³ banking. The trade-off is size: Kimi K3 has 2.8 trillion parameters with 104 billion active per token, about 4.5 times Beam's 23 billion, and its weights take roughly 1.4 TB even in 4-bit form.

Benchmark by benchmark

Benchmark	Beam	Kimi K3	Difference
DeepSWE v1.1	44.4	68.0	-23.6
SWE-Bench Pro v2-Hard	77.2	88.2	-11.0
Terminal-Bench v2.1	80.1	88.3	-8.2
SWE Atlas Codebase QnA	34.6	68.0	-33.4
Humanity's Last Exam (no tools)	36.2	46.9	-10.7
SciCode	49.7	58.7	-9.0
CritPt (AA)	16.3	23.4	-7.1
GPQA Diamond	90.5	93.5	-3.0
AutomationBench (public)	37.0	46.7	-9.7
MCP Atlas	78.7	82.3	-3.6
τ³ banking	38.0	37.1	+0.9
BrowseComp (context mgmt)	77.4	91.2	-13.8
DeepSearchQA (context mgmt)	80.1	95.0	-14.9
AA-LCR (long context)	79.3	88.7	-9.4

All scores are from Reflection's announcement (October 8, 2026 revision). Scores for Kimi K3 are as cited by Reflection from Artificial Analysis and DataCurve.

The two models

	Beam	Kimi K3
Developer	Reflection AI	Moonshot AI
Based in	United States	China
Parameters (total / active)	501B / 23B	2.8T / 104B
Weights	Apache 2.0, promised for later in October 2026	Kimi K3 License (custom)
Released	October 5, 2026	July 16, 2026
How to access	Waitlisted beta API; weights coming	Hugging Face, Moonshot API, OpenRouter