Artificial Analysis scored Inkling 41 on its Intelligence Index, the top US open-weights release, ahead of Nemotron 3 Ultra (38), Gemma 4 31B (29) and gpt-oss-120b (24). It still trails GLM-5.2 on agentic benchmarks and Kimi K2.6 on multimodal.
Inkling spends about 25K output tokens per Intelligence Index task, versus 43K for GLM-5.2, 38K for Kimi K2.6 and 37K for DeepSeek v4 Pro. Thinking less at comparable quality is a direct cost and latency win in agent loops.
pretrainingThe first, biggest phase of building a model: training it on enormous amounts of text so it learns language, facts, and reasoning in general.Full definition → ran on 45 trillion tokens of text, images, audio and video. The smaller sibling, Inkling-Small at 276B total and 12B active parameters, came in unexpectedly close to the big model on several community evaluations.
The attention stack is hybrid: five local sliding-window layers (window size 512) per global layer, plus 8 multi-token-prediction heads for speculative decodingA speed trick where a small model drafts several tokens ahead and the big model verifies them in one pass, often doubling generation speed.Full definition →. That is why vLLM reported up to 380 tokens/sec per user on four GB200s on day one.
Watch the context windowThe maximum amount of text a model can consider at once — its working memory for the current conversation or task.Full definition → gotcha: the open checkpoints handle up to 1M tokens, but TML's own Tinker API serves 256K, with separate pricing tiers at 64K and 256K. The headline context number is not what you get through the hosted path.
Terms in this piece · Glossary
mixture-of-experts — A model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.
context window — The maximum amount of text a model can consider at once — its working memory for the current conversation or task.
pretraining — The first, biggest phase of building a model: training it on enormous amounts of text so it learns language, facts, and reasoning in general.
speculative decoding — A speed trick where a small model drafts several tokens ahead and the big model verifies them in one pass, often doubling generation speed.
Why it matters
Inkling is a 975B-param (41B active) multimodal mixture-of-expertsA model built from many specialist sub-networks where only a few activate per token, giving big-model capability at small-model running cost.Full definition → released under Apache 2.0 with 1M-token context and same-day fine-tuning on Tinker, giving builders a permissively licensed U.S. open base model to customize rather than a locked-down API. Its novel architecture (relative positional bias, scaled short-conv layers, dual shared-expert MoE) makes it a distinct alternative to Llama, DeepSeek, and GLM/Kimi for teams needing open weights they can adapt.
Key quotes
“Thinky only seems to come up for air once every few months; most recently with Interaction models - but each time they do they impress, showing both taste and depth.”
“Our model, called Inkling, is a Mixture-of-Experts transformer with 975B total parameters, 41B active. It supports a context window of up to 1M tokens. It was pretrained on 45 trillion tokens of text, images, audio and video.”
“Thinking Machines Lab launched Inkling, its first fully released open-weights foundation model family entry, positioning it as a customizable multimodal base model rather than a benchmark-maxed flagship.”
