Inkling Small AI model arrives with near-Inkling benchmark scores

Thinking Machines has released the Inkling Small AI model, a 276-billion-parameter multimodal reasoning system under an Apache 2.0 license, two weeks after launching its first Inkling model. The startup led by former OpenAI CTO Mira Murati is pitching the model as a lower-cost, easier-to-run alternative to the 975-billion-parameter Inkling, while claiming only a limited performance tradeoff.

According to Thinking Machines, Inkling Small accepts text, image and audio inputs, returns text, and supports context windows of up to one million tokens. The company has published full weights on Hugging Face and supports fine-tuning through its Tinker training API.

The model is smaller by total parameter count and by active compute. Thinking Machines says Inkling Small uses 12 billion active parameters per token, compared with 41 billion for Inkling. That difference is the practical point for enterprise buyers: lower inference cost and a smaller deployment footprint, if the benchmark results hold for their own workloads.

What is Inkling Small?

Inkling Small is a sparse mixture-of-experts model, meaning it has a large pool of parameters but activates only part of that pool for each token. Thinking Machines says the 42-layer decoder sends each token to six of 256 specialized experts, plus two shared experts that are active for every token.

The architecture lets the company describe the model as 276 billion total parameters while using 12 billion active parameters during inference. That does not make it a local desktop model. It means the model is lighter than Inkling while remaining in the enterprise GPU class.

How does Inkling Small compare with Inkling?

Artificial Analysis gave Inkling Small a 40 on its Intelligence Index, compared with 41 for Inkling. The firm also reported that no open-weight model at Inkling Small's size or below had a higher score on that index.

Thinking Machines says the smaller model beats Inkling on several tests, including 80.2% versus 77.6% on SWE-bench Verified and 64.7% versus 63.8% on Terminal Bench 2.1. The company also reports higher results on SciCode, Humanity's Last Exam, GPQA Diamond and CritPt.

The comparison is not one-way. Inkling still leads on factual knowledge and some agentic tasks, according to the reported results. Inkling Small scored 15.5% on tau3-Banking, versus 23.7% for Inkling, and its Artificial Analysis Omniscience score was negative, which points to weaker factual coverage despite a slightly lower reported hallucination rate.

For operators, that suggests a specific fit rather than a general replacement. Coding assistants, retrieval-augmented generation systems, document workflows and multimodal applications may be plausible targets. Uses that depend on high-confidence factual recall still need retrieval, verification and human review.

What hardware does Inkling Small need?

Thinking Machines says the BF16 checkpoint requires at least 600 GB of aggregate GPU memory, with supported configurations including four Nvidia B300 GPUs or eight Nvidia H200 GPUs. A quantized NVFP4 checkpoint cuts the requirement to about 180 GB of aggregate VRAM.

The company says the quantized version can run in W4A4 mode on one Nvidia B300, or in W4A16 mode on two H200 GPUs. That rules out ordinary laptops, most desktop workstations and consumer gaming PCs. The likely users are enterprises with GPU servers, cloud infrastructure or specialized inference providers.

What does the license and pricing change?

Inkling Small is released under Apache 2.0, a permissive license that generally gives companies room to use, modify, fine-tune and commercialize software while meeting notice and attribution obligations. That is cleaner for legal and procurement teams than many custom AI licenses with added commercial conditions.

At launch, Thinking Machines is offering a limited-time 50% discount on API access. For the standard 64K-context model, the company lists pricing at $0.58 per million input tokens, $1.44 per million output tokens, $1.73 per million training tokens and $0.116 per million cached input tokens. A 256K-context version is available at higher prices.

The release also signals that Thinking Machines is turning Inkling into a model family rather than a single launch. The company says Inkling Small used an improved pre-training data mix, recipe changes, on-policy distillation from Inkling and two additional weeks of agentic coding reinforcement learning.

This story draws on original reporting from VentureBeat.