Inkling 2 Mini

Inkling 2 Mini is an open-weight model from Thinking Machines, released November 20, 2026. It ranks #38 of 43 models on the Vals Index with 46.46%. Its best result is #10 of 49 on CyberBench v1.1.

Release Date: Nov 20, 2026

Developer Thinking Machines ย ๐Ÿ‡บ๐Ÿ‡ธ
Context Window 1M
Max Output Tokens 1M
Token Costs (in/out) $0.15/0.50
Weights Open
Input Modalities

Accuracy

46.46 % ยฑ 1.11

Cost / Test (Vals Index)

$ 3.197

Latency

47 min 15 s

Vals Index
BenchmarksAccuracyRankings

46.46%

ยฑ1.11
38/43

50.05%

ยฑ1.35
40/46

55.77%

ยฑ1.08
42/44

23.58%

ยฑ1.10
28/29

23.16%

ยฑ3.87
26/31

74.58%

ยฑ5.35
10/49

66.02%

ยฑ2.49
37/48

56.54%

ยฑ2.13
39/46

17.79%

ยฑ2.66
77/100

25.71%

ยฑ1.90
105/109

86.37%

ยฑ1.86
25/111

22.97%

ยฑ2.83
24/35

18.00%

ยฑ3.86
44/54

85.89%

ยฑ1.78
14/30

0.00%

ยฑ0.00
38/38

14.12%

ยฑ2.20
71/91

77.92%

ยฑ3.76
44/134

47.78%

ยฑ1.70
34/35

5.00%

ยฑ1.72
36/44

11.61%

ยฑ4.96
45/47

0.50%

ยฑ0.50
30/67

43.24%

ยฑ4.24
28/36

37.92%

ยฑ2.22
20/23

6.57%

ยฑ0.51
57/70

0.00%

ยฑ0.00
46/49
Proprietary BenchmarksAcademic BenchmarksIndustry Partners
Vals
Default Provider : Thinkingmachines
Temperature: 1
Top P: 1
Top K: Default
Max Output Tokens: 1,024,000
Reasoning Effort: 0.99

Updates

Nov 20, 2026

Thinking Machinesโ€™ new open-weight Inkling 2 Mini ties Claude Opus 5.5 (CVP) on CyberBench at 74.58% (#10 of 49), at $0.93 per test versus $4.17. Itโ€™s small enough to run on a Mac laptop or DGX Spark.

  • On the Vals Index, Mini scores 46.46%, up 17.5 points from the original Inkling. It improves on the original across 13 of 18 shared benchmarks.
  • On Vibe Code Bench, it scores 77.92%, up from the original Inklingโ€™s 19.21%. It scores 56.54% on Finance Agent v2.1 at about 5 minutes per task, and 37.92% on SWE-bench Multimodal.
  • Mini reaches 96% of Inkling 2 Smallโ€˜s Vibe Code Bench score (77.92% versus 81.55%) and 97% of its SAGE score (85.89% versus 88.57%).

Thinking Machines says Mini has 26B parameters, with 5.4B active on each token, and twelve of its 24 layers run three times on each token, using more computation per token instead of more tokens. It has a 1M-token context window and supports 1M output tokens. The weights are available on Hugging Face and Ollama under Apache 2.0.

Congrats to the Thinking Machines team on the release!