Thinking Machines Lab Inkling 2 Small MoE Model Released as Firm Pursues $5 Billion Round

Thinking Machines Lab released Inkling 2 Small on November 5, 2026. The 594 billion parameter multimodal sparse mixture-of-experts model reached 31 on the Artificial Analysis Intelligence Index, more than doubling the score of earlier Inkling releases. The company simultaneously confirmed it is finalizing a $5 billion funding round at a $40 billion pre-money valuation.

Model Architecture and Benchmark Results

Inkling 2 Small uses a 48-layer decoder-only transformer with 594 billion total parameters and 17 billion active parameters per token. It incorporates hybrid local and global attention plus a 292 million parameter vision encoder that handles text and image inputs up to a 1 million token context window. The model was trained natively in NVFP4 precision on NVIDIA GB300 NVL72 systems through a single unified reinforcement learning run.

On standard benchmarks the model records 86.3 percent on SWE-Bench Pro v2 Hard, 81.6 percent on MMMU Pro, and 83.8 percent on ScreenSpot Pro. These vision and agentic scores exceed most prior open-weights systems outside China, although certain coding tasks remain below the leading closed models. Variable reasoning effort allows users to trade compute for accuracy during inference.

Access currently runs through the company Tinker platform at $0.30 per million tokens for prefill and $1.00 per million for sampling. The same interface supports fine-tuning. Full Apache 2.0 weights are scheduled for release around November 12 after completion of external safety testing and red-teaming.

Funding Round and Operational Context

Accel is leading the new $5 billion round, which would more than triple the firm’s prior $12 billion post-money valuation set after its July 2025 seed round. That earlier round raised $2 billion from a16z, Nvidia, and Accel. The fresh capital arrives as Tinker platform revenue exceeds a $100 million annualized run rate.

Founding CEO Mira Murati established Thinking Machines Lab in February 2025 after leaving OpenAI. Earlier family members include the July 15, 2026 release of the 975 billion parameter Inkling model and the July 30, 2026 release of the 276 billion parameter Inkling-Small variant. Several founding-team departures have occurred since the seed round, yet the company maintains active compute partnerships with Nvidia.

Closing the round this week would supply runway for continued infrastructure expansion and safety work ahead of the open-weights launch. The financing also positions the firm to compete directly with other large-scale model developers that have already opened their weights.