Thinking Machines Launches Inkling 2 Small: 594B/17B MoE, Vendor-Reported AA Intelligence Index ~31; Tinker Inference and Fine-Tuning Now, Apache 2.0 Weights Planned Next Week

Core Background & Industry Pain Points

Open and fine-tunable models still often trail closed frontiers on agentic coding, vision-in-the-loop work and long-horizon autonomy. On November 5 Thinking Machines launched Inkling 2 Small, the first second-generation model, keeping a closed-at-launch then Apache 2.0-after-safety-tests path.

Architecture Highlights & Internals

Official: 594B total / 17B active MoE with image input and variable thinking effort; no audio this generation. Training includes large-scale pretraining plus one joint RL run (~1,075 steps, 17M+ rollouts; coding ~38.5%). Vision uses a ~292M encoder; long tasks can use compaction.

Authoritative Benchmarks & Measured Scores

Vendor-reported from the launch blog (not AICoder re-runs): AA Intelligence Index ~31; FrontierCode v2.0 ~25.4%; DeepSWE v1.1 ~72.1%; DesignArena Website Elo ~1340. DesignArena’s account separately cited roughly 4th overall (third-party). Alignment suites improve versus older Inkling while still trailing some closed models.

Developer Hands-on Guide

On Tinker use thinkingmachines/Inkling-2-Small (64K) or the 256K :peft:262144 variant; list pricing about Prefill $0.30 / Sample $1.00 per 1M tokens. Weight downloads wait for the open release. Fit for long-horizon coding and vision-feedback prototypes; keep prior Inkling if you need audio.