Thinking Machines Lab on November 5, 2026 introduced Inkling 2 Small, the first model in its second Inkling generation. The company describes a sparse Mixture-of-Experts system with 594 billion total parameters and 17 billion active parameters, native image input, variable reasoning effort, and a context window of up to one million tokens. The model is available today on Tinker for inference, fine-tuning and sampling with closed weights.
The lab says Inkling 2 Small scores 31 on the Artificial Analysis Intelligence Index, a large jump over its prior Inkling Small score of 13 on the same index. Gains are attributed to more efficient pre-training and a large unified reinforcement-learning run rather than separate reasoning and agentic runs. Training notes cite an initial NVFP4 pretrain on NVIDIA GB300 NVL72 racks, later high-quality phases, then a joint RL run of more than 17 million rollouts over nine days.
Availability follows the company's staged open-weights path. Weights stay closed on Tinker at launch. Thinking Machines intends to publish full weights under an Apache 2.0 license the following week after external safety testing and red-teaming. Fine-tuned checkpoints are expected to become downloadable with that weight release, with additional platforms planned afterward.
Benchmark tables in the announcement emphasize agentic coding, vision-in-the-loop creation, knowledge work and finance tasks. Reported highlights include 86.3 percent on SWE-Bench Pro v2 Hard, strong DesignArena website and game scores among compared open-weight peers, and top marks on several finance and enterprise benches in the lab's tables. The company also shows long-horizon demos such as procedural city simulators, a Python PC emulator that boots Linux, and an autonomous time-series post-training run on Tinker.
Limitations are stated plainly: the model does not support audio, can be verbose, sometimes chooses tools poorly in skill-heavy codebases, and is less optimized for chat presentation. Safety sections report lower CBRN risk than the open-weight frontier on the lab's measures, improved alignment versus prior Inkling models, and remaining gaps versus closed models such as GPT-6.1 Sol on internal misbehavior evals.
Independent outlets including MarkTechPost, Techmeme and Unite.AI covered the launch the same day, matching the parameter counts, Tinker availability and planned Apache 2.0 weight release. For coding and agents desks, Inkling 2 Small matters as a Mira Murati-led lab's second-generation open-weights candidate with strong multimodal agent claims, not merely a tip-feed headline. Claims about benchmarks, safety scores and training scale are Thinking Machines Lab's unless independently re-run.