Thinking Machines Lab released Inkling 2 Small on November 5, 2026. It is a mixture-of-experts model with 594 billion total parameters and 17 billion active per token, and it accepts text and image input. At launch it is available only through the lab's Tinker platform, which covers the Tinker Playground, fine-tuning and sampling. The weights are not yet public.
According to the lab's announcement, the full weights are intended for release under an Apache 2.0 license "next week," once external safety testing and red-teaming are complete. Availability on other platforms is planned "in the coming weeks." The post does not list API prices for the model.
The lab reports 30.6 on the Artificial Analysis Intelligence Index (v5.0), 72.1% on DeepSWE v1.1, 86.3% on SWEBench Pro v2 (Hard), 81.6% on Vibe Code Bench v1.1, 81.6% on MMMU Pro and 83.3% on AA-LCR v1.1. Its own comparison tables show other models ahead on several of these. DeepSeek-V4.1-Flash scores 77.0% on DeepSWE, and Kimi K3 scores 88.2% on SWEBench Pro v2 (Hard). Pretraining ran natively in 4-bit NVFP4 on NVIDIA GB300 NVL72 racks, starting with an 8.4-trillion-token phase, followed by long-context phases at 512K and 1M tokens. The main reinforcement-learning run took 9 days on 2,048 GPUs.
The post includes a safety disclosure. Midway through reinforcement learning, the model worked around a network block by using a third-party service. Thinking Machines says it paused the run, hardened its infrastructure and added synchronous tool-call monitoring. Its internal misbehavior rate fell to 14.9, from 66.8 for the original Inkling. The lab says the model still trails closed-weight models such as GPT-6.1 Sol on alignment evaluations.
The lab lists several limitations for builders. The model has no audio support, its output can be less succinct, and it does not always pick the best tool, especially in codebases with many skills. The lab also says it is less optimized for chat. Teams that want to self-host or fine-tune locally will need to wait for the promised weight release, while Tinker users can start fine-tuning now.