Inkling 2 Small

• Nov 5, 2026

  Thinking Machines Lab releases Inkling 2 Small

Thinking Machines Lab's first second-generation model, a 594B-total/17B-active MoE announced 2026-11-05 (thinkingmachines.ai/news/inkling-2-small). Available on Tinker for serverless inference, sampling and fine-tuning; the lab says it intends to publish full weights under Apache 2.0 'next week' pending safety evaluations, so openWeights and license stay false/null until the weights are actually out. Pricing is Tinker's serverless inference rate ($0.30 input, $0.03 cached, $1.00 output per MTok at 256K context); Tinker training is $1.50/MTok at 64K and $3.00 at 256K. contextWindow 1M is Artificial Analysis' figure and matches the 1M trajectory limit the lab used for evals; Tinker serves 256K (serverless) and 64K or 256K (training). knowledgeCutoff is the '2026-06' data cutoff on the lab's pretraining chart. The lab reports none of the tracked benchmarks: no GPQA Diamond, SWE-bench Verified, ARC-AGI-2 or LMArena figure; its SWE-bench Pro is the v2 Hard split (86.3%), HLE is HLE-Diamond (28.7%), Terminal-Bench is only Terminal Bench Science v0.1 (5.7%), and GDPval-AA is v2.1 (1630), a different scale from the v2 this key records. Also reported at effort 0.99 (xhigh): DeepSWE v1.1 72.1%, FrontierCode v2.0 25.4%, Vibe Code Bench v1.1 81.6%, AA-LCR v1.1 83.3%, MMMU Pro 81.6%. Artificial Analysis scores it 31 on Intelligence Index v5.0.1 (the reasoning variant, no effort label) at $1.18 per task, 157 output tokens/s and 29.08 s to first answer token. predecessorId is null: the lab calls it its 'second model generation' but names no model it replaces.