On November 5, 2026, Thinking Machines Lab launched Inkling 2 Small, a 594-billion-parameter mixture-of-experts model designed for agentic workflows. Founded by former OpenAI CTO Mira Murati, the lab bets that large, sparse models with open weights will outcompete both proprietary giants and lean open-source models on real coding and research tasks — and Inkling 2 Small is the strongest evidence yet.
What's Under the Hood
Inkling 2 Small uses a sparse MoE architecture: 594B total parameters, but only 17B active on any given forward pass. That keeps inference cost comparable to a dense small model while giving the network vast capacity for specialized tasks. Training consumed 8.4 trillion tokens, ran natively in NVFP4 on NVIDIA GB300 NVL72 racks, and included a joint reinforcement-learning phase of 1,075 steps across more than 17 million rollouts — nine days of compute on 2,048 GPUs.
The model supports a 1-million-token context window and accepts both text and image inputs through a 292-million-parameter vision encoder. Thinking Machines positions it squarely at "agentic loops" — tasks where an AI must plan, write code, call tools, and iterate without constant human supervision.
Benchmarks That Actually Matter
According to the official model card, Inkling 2 Small scores:
- 86.3% on SWEBench Pro v2 (Hard) — the coding-agent benchmark that asks models to fix real GitHub issues end-to-end
- 72.1% on DeepSWE v1.1
- 81.6% on MMMU Pro (multimodal reasoning)
- 31 on the Artificial Analysis Intelligence Index, up from 13 for its predecessor Inkling Small
The company also ran an Autoresearch case study where the model autonomously fine-tuned a smaller open-source model for time-series forecasting — spending $426 over 4.5 hours with no human direction and improving downstream performance by roughly 50%. It's an early but concrete demonstration of what "agent-first" actually looks like in practice.
Open Weights — But Not Quite Yet
The model currently runs only on Thinking Machines' own Tinker platform, priced at $0.30 per million input tokens and $1.00 per million output tokens. Full Apache 2.0 weights are promised "next week," pending the results of external safety testing and red-teaming. It's an odd move for a lab that markets openness as its identity, but the company says the safety review is non-negotiable before distribution.
When the weights drop, Inkling 2 Small will rank among the most capable open models available — especially for teams running private inference on code generation or research pipelines. At 17B active parameters, the hardware requirements are within reach of well-resourced organizations who can't or won't send data to a third-party API. That combination of raw capability and eventual openness is exactly what the market has been waiting for.
Thinking Machines is simultaneously reportedly closing a $5 billion funding round, which would rank it among the best-capitalized AI labs outside the established giants. If the weights perform as advertised on release, that valuation will make a lot more sense to skeptics.