Inkling Guide: Thinking Machines' 975B Open Multimodal Model

Thinking Machines Lab is best known for its research posts and its Tinker fine-tuning service. Inkling, released on 15 July 2026, is an open-weight model of its own, and the launch post is unusually blunt about where it stands: Inkling "is not the strongest overall model available today, open or closed."

That sentence is the right frame for evaluating it. Inkling is not pitched as a leaderboard winner. It is pitched as a broad, balanced open-weight foundation that reasons over text, images and audio, ships under Apache 2.0, and is meant to be customised. Whether that is useful depends on whether you need those three things together.

What Thinking Machines shipped

Inkling is a mixture-of-experts transformer with 975B total parameters and 41B active per token. Each MoE layer has 256 routed experts plus 2 shared, with 6 routed experts active per token, behind a sigmoid router with auxiliary-loss-free load balancing. Attention interleaves sliding-window and global layers at 5:1, with 8 KV heads, and it uses a relative positional embedding instead of RoPE, which the authors say extrapolates better to long sequences.

Pretraining covered 45 trillion tokens of text, images, audio and video. Post-training started with supervised fine-tuning on synthetic data from open-weight models, including Kimi K2.5, followed by reinforcement learning over more than 30 million rollouts. Thinking Machines states a context window of up to 1M tokens.

The weights are on Hugging Face under Apache 2.0, in BF16 and in an NVFP4 checkpoint for Blackwell-class hardware. If total versus active parameters is unfamiliar, mixture of experts explained covers why 41B active is what sets your per-token compute.

Encoder-free multimodality, audio included

Most open multimodal models bolt a separate vision encoder onto a language model. Inkling does not. Audio enters as dMel spectrograms and images as 40-by-40-pixel patches through a small four-layer MLP, then pass through a light embedding layer into the same transformer as the text.

It can also call a Python tool during inference to zoom and crop images, which helps with dense charts and documents. The Hugging Face card asks for images between 40 and 4,096 pixels per side and audio as 16kHz WAV, ideally under 20 minutes.

The vendor's multimodal scores are respectable rather than leading: 73.5% on MMMU Pro, 77.2% on MMAU and 91.4% on VoiceBench. The launch post calls the multimodal capability a first release it expects to improve. Multimodal models explained covers how image and audio inputs become tokens and why that affects cost.

The effort dial runs from 0.2 to 0.99

Reasoning depth is a continuous setting rather than a choice of named levels. Thinking Machines sweeps effort from 0.2 to 0.99 and reports its headline benchmarks at 0.99 with temperature 1.0.

At that setting the vendor reports 77.6% on SWE-bench Verified, 54.3% on SWE-bench Pro, 63.8% on Terminal Bench 2.1, 87.2% on GPQA Diamond, 97.1% on AIME 2026, and 29.7% on Humanity's Last Exam (46.0% with tools). It also claims Inkling matches Nemotron 3 Ultra on Terminal Bench 2.1 at roughly a third of the tokens.

Two cautions. Every one of those numbers is at maximum effort, so your default setting will score lower and cost less. And they are vendor-run, with some contaminated Terminal Bench solutions zeroed out by the vendor. Vendor-claimed vs independent benchmarks covers how to read figures like these.

The awkward fact: the smaller sibling scores higher

Two weeks later Thinking Machines released Inkling Small, at 276B total and 12B active. On the vendor's own comparison table it beats Inkling on most reasoning and agentic coding benchmarks: 80.2% against 77.6% on SWE-bench Verified, 64.7% against 63.8% on Terminal Bench 2.1, and 89.5% against 87.2% on GPQA Diamond.

So why run the big one? The same table answers it. Inkling keeps a clear lead on knowledge and factuality: 43.9% on SimpleQA Verified against 20.6% for Small, and 2.1 against -9.0 on the AA Omniscience index. It also leads on Tau 3 Banking (23.7% against 15.5%), on adversarial refusal in FORTRESS, and slightly on the AA Index v4.1, 41.0 against 40.0.

Thinking Machines credits Small's gains to a revised pre-training data mix and recipe, on-policy distillation from Inkling, and two further weeks of agentic coding RL. The knowledge gap fits the usual expectation that more total parameters hold more facts. If your workload leans on recall, such as answering questions without retrieval, multilingual knowledge or domain trivia, Inkling is the safer pick. If it leans on reasoning over material you supply, Small is likely the better buy.

What it costs to run

On OpenRouter at the time of writing, Inkling lists at $1.00 per million input tokens and $4.05 per million output, with cached input at $0.17. Thinking Machines uses the same $4.05 output figure in the cost comparisons of its Inkling Small announcement. The paid OpenRouter listing exposes 524,288 tokens of context, below the vendor's 1M maximum, so check what your provider actually serves.

There is also an inkling:free variant on OpenRouter. It is the same model on a rate-limited free endpoint: fine for trying it, not for production traffic. For fine-tuning, Inkling is available on Thinking Machines' Tinker service with 64K and 256K context options.

Self-hosting is a serious undertaking. 975B total parameters is a multi-GPU deployment even at NVFP4, and the card names SGLang, vLLM, TokenSpeed, Unsloth and Transformers as supported stacks without stating minimum hardware.

How to decide

1. Do you need audio or images in the same model? If not, compare Inkling against text-only open models on price, and it may lose.
2. Is your task recall-heavy or reasoning-heavy? Run twenty real prompts through both Inkling and Inkling Small at the same effort. If Small matches it, take the cost saving.
3. Sweep effort. Try 0.5 and 0.99 on the same tasks and record output tokens. The best value point is rarely the maximum.
4. Check the licence fit. Apache 2.0 lets you fine-tune and resell, which is the reason to choose an open model in the first place. The model card also links an acceptable use policy, so read that too.