The Aggregate Digest — Tuesday, October 20, 2026

DeepSeek V4.1 Flash (1017) replaces Kimi K3 as the best open-weights model

DeepSeek V4.1 Flash (1017), the October 17 version of DeepSeek V4.1 Flash, is now the best open-weights model. It scores 56.58% on the Vals Index, which averages Vals AI's agent tests by industry, against 50.3% for Kimi K3. Vals AI measured its cost at $0.64 per test, against $6.44 for Kimi K3.

A 23-point lead, held for three days. DeepSeek V4.1 Flash (1017) has an Aggregate Rating of 1874 and ranks #20. Kimi K3 is at 1851 and ranks #30. The lead was 59 points on October 18, when the new version had 42 benchmarks, and it is 23 now, with 230.

A clear lead in terminal work. Terminal-Bench 4.0 gives agents long command-line tasks, from software engineering to hardware design. DeepSeek V4.1 Flash (1017) completes 31.82% of them in Vals AI's run against Kimi K3's 17.17%, and 33.84% against 12.63% in Artificial Analysis' run at max effort. Vals AI's corporate tax tasks show 44.35% against 33.37%.

Artificial Analysis and Vals AI, October 20, 2026. Artificial Analysis rows use max effort for both models; Vals AI rows state no setting. Points on each board's 0–100 scale; different tasks, not a combined score.

Kimi K3 still recalls more facts. Kimi K3 answers 47.58% of Artificial Analysis' factual-recall questions correctly at max effort, against 42.85% for DeepSeek V4.1 Flash (1017). On Vals AI's Excel modeling tasks it also leads, 66.68% against 61.8%.

About twice the September version's cost. Vals AI measured $0.33 per test for the September DeepSeek V4.1 Flash, which scores 51.32% on its index. MiMo-V2.6-Pro, the next open-weights model there, scores 55.2% at $0.41.

Vals AI, October 20, 2026. Cost per test and Vals Index score as measured by Vals AI; selected open-weights models.

For terminal and agent work on open weights, test DeepSeek V4.1 Flash (1017) first. Where answers depend on recalled facts, keep Kimi K3 in the comparison.

Read the full digest on The Aggregate →