Inkling

Inkling

Quick answer: Inkling is Thinking Machines Lab's first open-weights model — a 975B total / 41B active Mixture-of-Experts transformer natively trained for text, image, and audio reasoning, with a 1M token context window, controllable thinking effort, and Apache 2.0 licensing. It scores 77.6% on SWE-bench Verified and 87.2% on GPQA Diamond.

At a Glance

Where Inkling leads

• Strong open-weights multimodal capability: native text, image, and audio reasoning in a single model
• Controllable thinking effort — matches frontier open-weights performance at a fraction of the token cost
• SWE-bench Verified 77.6% and GPQA Diamond 87.2% are competitive with much larger closed-weight models
• Apache 2.0 license with weights on Hugging Face; fine-tunable via Thinking Machines' Tinker platform

Where it lags

• Terminal Bench 2.1 score of 63.8 trails closed-weight frontier models (GPT-5.6 Sol: 89.5, Claude Fable 5: 84.6)
• HLE (with tools) 46.0% is mid-tier among frontier models; strong closed models exceed 64%
• Self-hosted BF16 checkpoint requires at least 2 TB of aggregated VRAM

Best for: Developers who need a customizable, multimodal open-weights base that covers agentic coding, vision, and audio — especially when fine-tuning on domain-specific data via Tinker.

What Inkling Is

Inkling is Thinking Machines Lab's first from-scratch model release, designed explicitly as a broad foundation for fine-tuning rather than a narrowly optimised benchmark maximiser. It is a decoder-only Mixture-of-Experts transformer with 975B total parameters and 41B active per token. The architecture broadly follows DeepSeek-V3's MoE design (256 routed experts, 2 shared experts, 6 active per token), with departures including a sigmoid-based router, a hybrid of sliding-window and global attention layers, and a relative positional embedding instead of RoPE for better long-context extrapolation.

The model was pretrained on 45 trillion tokens of text, images, audio, and video, then post-trained with large-scale RL over 30M+ rollouts. An unusual design choice is controllable thinking effort: by adjusting a scalar effort parameter at inference time, users can trade off token cost against performance — Inkling reaches the same Terminal Bench score as Nemotron 3 Ultra at roughly one-third the tokens. Both image and audio inputs are processed natively (encoder-free architecture) and projected into the same token space as text.

Inkling is released under Apache 2.0 and available for download on Hugging Face in both BF16 and NVFP4 (quantised for NVIDIA Blackwell) formats. API access is available through Thinking Machines' Tinker platform and third-party providers including TogetherAI, Fireworks, Modal, Databricks, and Baseten.

Specifications

Pricing

Inkling is open weights (Apache 2.0) — free to download and self-host.

Self-hosted BF16: requires ≥2 TB aggregated VRAM (e.g. 16× H200 or 8× B300). NVFP4 quantised checkpoint reduces the requirement to ≥600 GB (e.g. 8× H200 W4A16).

Related Models

• Inkling-Small (276B total, 12B active) — preview released alongside Inkling; full weights pending.
• Compare open-weights frontier MoE: Kimi K2.6 · Nemotron 3 Ultra · DeepSeek V4 Pro

Last updated 2026-07-16.