Thinking Machines Launches Inkling 2 Small with 594B MoE Architecture and Apache 2.0 Weights Coming

Quick Facts

• Inkling 2 Small is a 594B Mixture-of-Experts model that activates only 17B parameters per forward pass and supports a 1 million token context window.
• The model scored 31 on the Artificial Analysis Intelligence Index, an 18-point jump over its predecessor, with the biggest gains in agentic coding and enterprise tasks.
• Weights will be released under an Apache 2.0 license after external safety testing finishes, with pricing set at $0.30 per million input tokens and $1.00 per million output tokens.

Thinking Machines Lab released Inkling 2 Small on Nov. 5, 2026, its second-generation model and its most capable yet. The launch arrives less than four months after the original Inkling lineup debuted in July.

The model is a Mixture-of-Experts architecture with 594 billion total parameters but routes each token through just 17 billion active parameters per forward pass. It includes a 292-million-parameter vision encoder and supports text, image, and tool-calling inputs across a 1 million token context window.

How It Was Built

Thinking Machines trained the model on a cluster of NVIDIA GB300 NVL72 racks, each containing 72 Blackwell Ultra GPUs. Initial pretraining covered 8.4 trillion tokens of natural-language, code, and image data and completed in 68 hours using native NVFP4 training.

Later phases added 3.8 trillion tokens of newer, higher-quality data, roughly 320 billion tokens of reasoning and agent trajectories, and about 34 billion tokens of selected agent trajectories. The company then ran a nine-day reinforcement learning workflow to refine the model.

The 48-layer decoder-only transformer routes each token to eight of 384 experts and combines local and global attention, according to the official model card.

Benchmark Results

Artificial Analysis scored Inkling 2 Small at 31 on its Intelligence Index, up 18 points from the previous version. Its AutomationBench-AA score jumped from 5% to 46%, and its FrontierCode 2.0 score rose from 5% to 25%.

On enterprise-specific tasks, the model scored 46.8% on EnterpriseOps-Gym, a benchmark testing agentic performance across Customer Service, HR, and IT workflows. That result beats Claude Sonnet 5, Anthropic's mid-tier model, and several open-weight rivals including GLM-5.3.

Software engineering scores include 72.1% on DeepSWE v1.1, 81.6% on Vibe Code Bench v1.1, and 86.3% on SWE-bench Pro v2 (Hard). Vision benchmarks came in at 81.6% on MMMU-Pro Standard 10 and 64.7% on PerceptionBench.

Vals AI called it the highest-scoring US open-weight model by a wide margin and ranked it seventh among all open-weight models globally at launch, according to the company's announcement. When the weights ship, Thinking Machines says Inkling 2 Small will rank above Mistral Large 4, which scored 26 on the same index.

The model's weakest result was 13.1% on Terminal-Bench 4.0, well behind DeepSeek-V4.1-Flash-1017's 33.8% on that test.

Pricing and Availability

The model is currently available through Thinking Machines' Tinker API with closed weights. Pricing is $0.30 per million input tokens and $1.00 per million output tokens at the 64K context tier. Cached tokens cost $0.03 per million.

That is roughly six times more expensive than its predecessor on a per-Intelligence-Index-task basis, at approximately $1.18 versus $0.20. But the price stays well below comparable closed models: Gemini 3.8 Flash runs about $7.77 and Claude Sonnet 5.5 about $10.02 on the same basis.

Open weights under an Apache 2.0 license are expected within the week of launch, pending completion of external safety testing and red-teaming. The company has not set a specific release date.

Business Context

The launch comes as Thinking Machines is reportedly pursuing a $5 billion fundraising round, according to coverage from the time of the announcement. The company has not confirmed the fundraising publicly.

The model targets developers building agents for software engineering, financial document analysis, image interpretation, and company-specific workflow automation. Its Apache 2.0 license, once active, will allow commercial use and modification without royalties, a meaningful advantage for enterprise buyers evaluating self-hosted deployments.