GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
GLM-52 897 —
GPT-56SC 873 —
CL-OP5X 865 -0.9%
GROK-46H 865 -0.9%
GEM-37FH 865 -0.9%
GPT-56T 861 —
GLM-5 856 —
MUSE-SPK 841 —
QWEN-38X 824 -2.3%
GPT-6A 820 —
KIMI-K3X 810 -1%
CL-FAB5H 787 -0.9%
CL-OP5H 764 -0.9%
CL-OP46H 742 -0.9%
CL-OP47H 733 -1.1%
GEM-38FH 676 -1%
CL-OP47 585 -0.7%
INKL 531 —
CL-OP46 496 -0.2%
CL-OP48 490 -0.2%
← Back to feed

Thinking Machines Inkling Debuts as US Open-Weights Leader at AA Index 41 — Beats Kimi K2.6 on Agentic Benchmarks

Thinking Machines has shipped Inkling, its first production language model after a series of research previews. The model debuts at 41 on the Artificial Analysis Intelligence Index — the highest score among open-weight models released by a US lab — 3 points clear of NVIDIA Nemotron 3 Ultra (38), and well above Gemma 4 31B (29) and gpt-oss-120b (24).

Architecture

Inkling is a 975B-parameter mixture-of-experts model with 41B active parameters per forward pass. It accepts text, image, and audio input with text-only output. Images and video are encoded via a hierarchical patch encoder; audio uses discrete token encoding. All modalities project into a shared hidden space before decoder processing.

Context: 256K tokens via Thinking Machines’ Tinker platform API, 1M tokens via the HuggingFace open-weight release.

Benchmark Numbers

BenchmarkInklingKimi K2.6DeepSeek V4 Flash max
AA Intelligence Index414440
GDPval-AA v2 ELO1,2381,1901,189
τ³-Banking24%21%23%
Avg output tokens / task25K38K—

Inkling leads Kimi K2.6 on both agentic benchmarks despite sitting 3 points below it on the overall Intelligence Index. That gap suggests specific optimisation for agent and tool-use workflows over benchmark saturation.

AA-Omniscience: +2 accuracy score, 40% factual accuracy. Hallucination reliability is a known limitation at this tier.

Pricing

At 64K context: $1.87/M input, $4.68/M output ($0.374/M cached).
At 256K context: $3.74/M input, $9.36/M output.

This puts Inkling comfortably below frontier proprietary models on per-token cost. Claude Fable 5 sits 24 points higher on the Intelligence Index (64.9) but costs roughly 3x more per output token before reaching equivalent context windows.

Token Efficiency

Inkling averages 25K output tokens per Intelligence Index task — substantially lower than GLM-5.2 max (43K), Kimi K2.6 (38K), and DeepSeek V4 Pro max (37K). At Inkling’s pricing, that translates to materially lower cost-per-task in agent deployments where output length is the primary cost driver.

What It Means

The US open-weights tier has been ceded almost entirely to Chinese labs since Llama 4. Nemotron 3 Ultra (38) was the previous US ceiling. Inkling clears that by 3 points and beats two frontier Chinese open-weight models on agentic performance metrics — the benchmarks that matter most for production agent deployments.

The 1M-context HuggingFace release is notable. Most US labs at this capability tier ship API-only or impose heavy access restrictions on weights. Thinking Machines is going the opposite direction: full weight access, long context, multimodal. The pricing is competitive without being subsidised to unsustainable levels.

The agentic benchmark lead over Kimi K2.6 is the result to watch. GDPval-AA and τ³-Banking both measure multi-step agent task completion rather than static QA, and leading those benchmarks at Index 41 positions Inkling as a credible open-weight option for agent infrastructure where cost and deployability matter more than raw intelligence.