Head to head

Beam vs Inkling

13:2

Beam wins 13 of 15 shared benchmarks

Verdict

Inkling, from Thinking Machines Lab, is the other big US open-weight model of 2026 and already ships under Apache 2.0. Beam beats it on 13 of 15 shared benchmarks, with large margins on SWE-Bench Pro, Terminal-Bench and CritPt. Inkling is marginally ahead on IFBench and the AA Omniscience index. Reflection notes Inkling's RL used 30 million rollouts against more than 100 million for Beam.

Benchmark by benchmark

Benchmark	Beam	Inkling	Difference
SWE-Bench Pro v2-Hard	77.2	56.9	+20.3
SWE-Bench Pro v1	65.5	54.3	+11.2
Terminal-Bench v2.1	80.1	63.8	+16.3
SWE-Bench Verified	80.9	77.6	+3.3
AIME 2026	97.8	97.1	+0.7
Humanity's Last Exam (no tools)	36.2	29.7	+6.5
SciCode	49.7	46.1	+3.6
CritPt (AA)	16.3	5.4	+10.9
GPQA Diamond	90.5	87.2	+3.3
MCP Atlas	78.7	76.0	+2.7
τ³ banking	38.0	25.0	+13.0
BrowseComp (context mgmt)	77.4	77.1	+0.3
AA-LCR (long context)	79.3	77.3	+2.0
IFBench	79.7	79.8	-0.1
AA Omniscience index (public split)	13.0	14.2	-1.2

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

The two models

	Beam	Inkling
Developer	Reflection AI	Thinking Machines Lab
Based in	United States	United States
Parameters (total / active)	501B / 23B	975B / 41B
Weights	Apache 2.0, promised for later in October 2026	Apache 2.0
Released	October 5, 2026	July 15, 2026
How to access	Waitlisted beta API; weights coming	Hugging Face, Tinker API