Inkling 2 Small Benchmarks, Fully Mapped: AA 31, Topping ARC, and Terminal-Bench 4.0's 13.1%

Inkling 2 Small has no shortage of scores, but they are very unevenly spread: the same official table holds an 86.3% highlight and a 13.1% weak spot. This article separates every checkable score by source — vendor-measured, independent boards, third-party tests — and states the test conditions behind each. The headline: its strengths are vision, finance and abstract reasoning; its weak spot is terminal-agent work; and the official scores were all measured at the highest thinking level, so everyday use will score lower.

The official table: strengths and weak spots equally clear

Every score in the official model card was measured at effort 0.99 (xhigh), temperature 1.0, with a 1M-token trajectory cap:

Two caveats on method: first, everything is an xhigh-effort score — default-effort scores will be lower, and the lab published no default-effort comparison; second, several scores are vendor-measured, with rivals' figures taken from their own reports rather than head-to-head runs.

AA index 31: how to read the number

The Artificial Analysis Intelligence Index v5.0 unrounded is 30.6, rounded to 31 in the announcement; v5.0.1 also reads 31. Three reference frames: against its predecessor Inkling Small at 13, it is +18; within its price class, 4th of 29; against the field, below DeepSeek V4.1 Flash (32.6), Claude Haiku 5.5 (36.5) and slightly below Gemini 3.8 Flash (30.8). AA's words were that on weights release it would be "the highest-scoring open weights model developed outside China" — the qualifiers matter: AA index only, outside China only, open weights only. All three must hold for the "first" claim.

Topping ARC: abstract reasoning is a genuine strength

On ARC Prize's tests, Inkling 2 Small scored 81.7% on ARC-AGI-2 and 96.5% on ARC-AGI-1, the best open-weights results. ARC measures abstract reasoning and pattern generalization, not knowledge recall — and this result corroborates the high PerceptionBench and MMMU Pro scores, suggesting vision and abstract reasoning are real strengths rather than a single table's accident.

Terminal-Bench 4.0 at 13.1%: how bad is the weak spot

It is the weakest item in the official card: 13.1%, under half of GLM-5.3-Flash's 32.8% and under a third of GLM 5.3's 41.9%. Terminal-Bench tests terminal-agent work — completing multi-step tasks autonomously in a command-line environment, which needs tool selection, state tracking and long-horizon planning. The lab has not explained the low score; one of the model card's own listed weaknesses matches it exactly: "it doesn't always pick the best tools for large codebases." Read against SWE-Bench Pro's 86.3%: give it a well-defined code-patch task and it is strong; ask it to find its own way around a terminal environment and it is weak. Test it on your own tasks before buying it for terminal-agent work.

Independent measurements: the cost and speed side

AA's independent measurements add the dimension the official table lacks: $1.18, 15.7 minutes and about 150K output tokens per task, against $0.20, 7.0 minutes and about 37K for the predecessor. Scores rose, and so did the bill and the clock. Plotly's data-analysis benchmark (Nov 12) is another independent data point: 9 points above the predecessor at similar cost and twice the speed — in data analysis, at least, the progress came without a cost increase. thoughtfullab's finance-diligence test was likewise positive: a full memo for $0.69, spotting a $1.2B misstatement.