Editorial illustration for Thinking Machines Releases Inkling 2 Small, a 594B MoE With 17B Active
Thinking Machines Releases Inkling 2 Small 594B MoE
Thinking Machines Lab put out Inkling 2 Small on Tuesday, the first entry in its second Inkling generation and a sizable jump from the original Inkling Small. The new model is a sparse Mixture-of-Experts setup with 594 billion total parameters, though only 17 billion activate per token, and it handles both text and images with a 1-million-token context window. On the Artificial Analysis Intelligence Index it scores 31, more than double the 13 posted by its predecessor, putting it fourth out of 29 models in its price bracket.
For now, Inkling 2 Small only runs through the Tinker API. There's no word yet on what hardware would be needed to self-host it, though the weights do support MXFP8 and NVFP4 formats. Thinking Machines has said weights will stay closed at launch, with an Apache 2.0 release planned for next week once external safety testing and red-teaming wrap up, in line with the lab's safe path to open weights policy.
The model's performance varies sharply by task. On SWE-Bench Pro v2's hard split it hits 86.3%, trailing only Kimi K3 in the lab's own comparison table. Terminal-Bench 4.0 tells a different story.
Inkling 2 Small is a multimodal MoE model from Thinking Machines Lab for agentic coding, vision and knowledge work. It takes text and image input and returns text.
Why this matters
A jump from 13 to 31 on the Artificial Analysis Intelligence Index in one generation is the kind of gain that gets attention, but it's worth asking what's driving it before getting too excited. Thinking Machines hasn't published training details here, just benchmark tables scored at effort 0.99 and temperature 1.0, mostly through third-party evaluators like Scale AI and Vals. That's useful for comparison, but it's not the same as independent reproduction.
For developers already on Tinker, the practical story is more interesting than the score: 594B total parameters with only 17B active means you get frontier-ish reasoning without frontier-ish inference costs, and the variable reasoning effort knob lets you trade latency for quality on a per-request basis. That's a real lever for anyone building products where cost per query actually matters. For researchers, the open question is whether this MoE architecture is what's producing the benchmark lift or whether it's just a bigger, better-tuned version of the same recipe.
Worth watching how it performs outside the curated task list once people start fine-tuning on their own data.
Related Reading
Open Source Study: 6.7% of Deepfake Requests on Hugging Face Targeted Children
Open Source D&B rebuilds 642 million‑business database after AI agents hit limits
Open Source Axiom Math raises USD 200M Series A led by Menlo to keep AI slop out of math
From Archives Airbnb develops AI travel agent to handle trip planning beyond routine tasks
From Archives Tencent releases Hy4 AI with vision and computer use, claims second in China
Common Questions Answered
What is the parameter structure of Inkling 2 Small and how does its Mixture-of-Experts design work?
Inkling 2 Small is a sparse Mixture-of-Experts model with 594 billion total parameters, but only 17 billion parameters activate per token. This sparse architecture allows the model to maintain a large parameter count while keeping computational efficiency by selectively activating only the necessary parameters for each input.
How much did Inkling 2 Small improve over the original Inkling Small on the Artificial Analysis Intelligence Index?
Inkling 2 Small scored 31 on the Artificial Analysis Intelligence Index, more than doubling the score of 13 achieved by its predecessor. This significant improvement placed the new model fourth out of the evaluated models, representing a substantial leap in performance between the two generations.
What multimodal capabilities does Inkling 2 Small support and what is its context window size?
Inkling 2 Small is a multimodal model that accepts both text and image inputs and returns text outputs, making it suitable for agentic coding, vision, and knowledge work tasks. The model features a 1-million-token context window, enabling it to process and understand significantly longer sequences of information.
What are the limitations of the benchmark scores reported for Inkling 2 Small?
Thinking Machines has not published detailed training information for Inkling 2 Small, providing only benchmark tables scored at effort 0.99 and temperature 1.0 through third-party evaluators like Scale AI and Vals. While these scores are useful for comparison purposes, they do not represent independent reproduction of results and should be considered within that context.
Further Reading
- Inkling 2 Small Launches on Tinker With Weights Next Week - ShiftDelete
- Introducing Inkling 2 Small - Thinking Machines Lab
- Inkling 2 Small Model Card - Thinking Machines Lab
- Inkling: Our Open-Weights Model - Thinking Machines Lab
