Inkling

The first open-weight model from Thinking Machines Lab, released July 15, 2026. Inkling is a decoder-only multimodal mixture-of-experts — 975B total parameters with 41B active (256 experts, 6 routed + 2 shared) — that accepts text, image and audio input across a 1M-token context window. It was trained on 45T tokens of text, image, audio and video, and the weights ship on Hugging Face under Apache 2.0 with day-0 support in transformers, SGLang, vLLM, TokenSpeed and llama.cpp.

Inkling strengths

• First open-weight release from Thinking Machines Lab
• Multimodal input: text, image and audio
• 1M-token context window
• Apache 2.0 weights in three formats (BF16, NVFP4, 1-bit GGUF)
• Day-0 support in transformers, SGLang, vLLM, TokenSpeed and llama.cpp

Pricing & context

Cost guide: there is no first-party API list price — Inkling is open weights, so you run it yourself or through a hosting provider. Published formats: BF16 (~2TB VRAM), NVFP4 (~600GB VRAM, for NVIDIA Blackwell) and a 1-bit GGUF via Unsloth.

Published benchmarks

Scores as reported by Thinking Machines Lab at effort=0.99.

When to choose Inkling

Inkling is best for self-hosted deployments that need multimodal input — text, image and audio — with a 1M-token context window under a permissive Apache 2.0 license. If you want a smaller open-weight model instead, see gpt-oss-120b.

Inkling FAQ

How much does Inkling cost?

Inkling is an open-weight release, so there is no first-party API price. The weights are published on Hugging Face under Apache 2.0 in three formats: BF16 (~2TB VRAM), NVFP4 (~600GB VRAM, for NVIDIA Blackwell) and a 1-bit GGUF via Unsloth. Hosted pricing is not yet published.

What is Inkling best for?

Inkling by Thinking Machines Lab is best for self-hosted deployments that need multimodal input — text, image and audio — with a 1M-token context window under a permissive Apache 2.0 license.

Is Inkling multimodal?

Yes. Inkling accepts text, image and audio input. It was trained on 45T tokens of text, image, audio and video data.

How does Inkling score on published benchmarks?

At effort=0.99, Thinking Machines Lab reports AIME 2026 97.1%, SWE-Bench Verified 77.6%, MMMU Pro 73.3% and VoiceBench 91.4% for Inkling.