Inkling 2 Small on BenchGecko · Thinking Machines Lab 594B MoE

What launched

Thinking Machines Lab released Inkling 2 Small on November 5, 2026 as the first model of its second Inkling generation. It is available on Tinker for inference and fine-tuning with closed weights. The lab intends to publish full weights under Apache 2.0 after external safety testing and red-teaming. BenchGecko adds the model and a new Thinking Machines provider.

Specs

From the Thinking Machines announcement and model card:

• Architecture · sparse mixture of experts, 594B total parameters, 17B active per token
• Experts · 384 routed, with 8 selected per token
• Layers · 48, hybrid local and global attention
• Context · up to 1,048,576 tokens
• Inputs · text and image (292M Perception Encoder vision tower); no audio
• Output · text
• Numerics · MXFP8 and NVFP4
• Licence · Tinker Terms of Use for now · Apache 2.0 weights planned after safety testing
• Serving · Tinker (training, sampling, serverless inference beta)

Pricing on BenchGecko

Official Tinker pricing lists Inkling 2 Small at $0.30 input / prefill, $1.00 sample / output and $0.03 cached input per million tokens (64K and 256K training tiers; serverless inference beta lists the same $0.30/$0.03/$1.00 rates at 256K). Artificial Analysis reports the same $0.30/$1.00 API rates and an Intelligence Index score of 31. BenchGecko stores the Tinker inference rates on the model row.

Vendor reported benchmarks

Thinking Machines and Artificial Analysis report these figures. BenchGecko records vendor and third-party claims as published.

• Artificial Analysis Intelligence Index v5.0 · 31
• FrontierCode v2.0 · 25.4
• DeepSWE v1.1 · 72.1
• SWEBench Pro v2 (Hard) · 86.3
• AA-Briefcase v1.1 · 1527
• HLE-Diamond · 28.7
• LongBench v2 · 70.6

Sources

• Thinking Machines announcement
• Inkling 2 Small model card
• Tinker pricing
• Artificial Analysis model page

Source: BenchGecko