Thinking Machines Inkling: 975B, Open, and Hard to Run

Thinking Machines Inkling is a 975-billion-parameter open-weight model released on July 15, 2026, built as a sparse mixture-of-experts that activates only 41B parameters per token and takes text, image, and audio input. It ships under Apache-2.0, so you can download the full weights and fine-tune them commercially. The catch: even a 4-bit quant needs roughly 600GB of VRAM, so this is a model teams evaluate on a cluster, not something you run on a gaming GPU.

The specs, and what the headlines got wrong

Here is what the Hugging Face model card and the official Inkling page state (accessed July 16, 2026).

Three things floating around the launch coverage need correcting:

• Context window. Several write-ups quoted a 1M-token window. The official Inkling page lists 64K standard, and 256K when run through the Tinker platform; the Hugging Face card doesn't state a number at all. Treat 64K/256K as the figures you can verify and 1M as unconfirmed.
• License. It's confirmed Apache-2.0, one of the most permissive licenses available, and it clears commercial use. Thinking Machines does put the burden of fine-tuning safety on you.
• "Inkling-Small." A smaller variant with ~12B active parameters appeared in some reports, but it is not listed on the Hugging Face model card. Until it shows up there, don't plan around it.

What "975B parameters" really means here

Inkling routes each token to 6 of 256 experts plus 2 shared experts, so only about 41B parameters fire at a time even though the full pool is 975B. That's why Thinking Machines claims it matches Nvidia's Nemotron 3 Ultra on coding while spending roughly a third fewer tokens to get there: you get a large model's breadth at a mid-size model's inference cost.

Inkling also exposes a "thinking effort" dial. Turn it up for harder problems and accept slower, more deliberate answers, or turn it down for speed. It's trained to flag uncertainty rather than guess, which matters more for a fine-tuning base than for a consumer chatbot.

Can you actually run Inkling?

This is where the "open weights" label gets complicated, because open doesn't mean light. Here's roughly what it takes to serve the full 975B model, based on the model card and early community estimates (these are ballparks, not guarantees):

• BF16 (full precision): ~2TB of VRAM for the weights alone, so 16+ H200-class GPUs, or an 8-GPU B300 node, before you add serving overhead.
• NVFP4 / 4-bit: ~600GB, still multi-GPU server territory (roughly 4× H200 or 8× 80GB cards).
• 1-2 bit GGUF (llama.cpp / Unsloth quants): ~280-350GB of combined RAM+VRAM, reachable on a high-end workstation or a maxed-out Mac Studio Ultra, and slower once you push toward long context because the KV cache grows.

The honest read: a well-funded team can evaluate and fine-tune Inkling, but there is no consumer-GPU path to running it locally today. If you're a solo developer hoping to load it on a single 24GB card, this isn't your model. You download it from Hugging Face at thinkingmachines/inkling, or fine-tune it through Thinking Machines' Tinker platform, which also unlocks the 256K context.

Where Inkling lands against other models

Thinking Machines says outright that Inkling is not the strongest model available, and the model card benchmarks back that up: Inkling scores well but trails the closed frontier on reasoning and coding.

*Source: Inkling model card and official benchmark page, accessed July 16, 2026.*

The official radar chart tells the same story visually. Inkling holds its own on reasoning and multimodal tasks but sits inside GPT 5.6 Sol and Claude Fable 5 on agentic coding (SWE-bench Pro, Terminal Bench).

Where it's strong is breadth: native audio and image understanding in one open model, with a coding gap you can narrow through fine-tuning. If you need best-in-class agentic coding out of the box, a closed frontier model still wins. If you need an open, multimodal base you own, Inkling is the more interesting option.

Who Inkling is really for

Inkling is a base to build on, not a finished assistant. Thinking Machines makes money from Tinker, its fine-tuning platform, not from metered API calls, so the model itself is free and the pitch is "take this and specialize it."

The clearest proof point is Bridgewater Associates: per Thinking Machines, a finance-tuned version of the model scored 84.7% on the firm's reasoning test at roughly one-fourteenth the cost of a proprietary model. That's the intended use, a capable generalist that a team reshapes for a specific domain.

So it's a fit if you have the infrastructure and the fine-tuning expertise to specialize an open model and own the result, and you can accept responsibility for safety tuning. Skip it if you want an out-of-the-box chatbot or you're working from consumer hardware, because the general-purpose closed models are cheaper to reach and stronger on day one. One caveat worth knowing: Inkling's post-training used data generated by other open models, including Moonshot's Kimi K2.5, and Thinking Machines says its next generation will be fully self-trained.

Inkling FAQ

Is Inkling free to use commercially?

Yes. It's released under Apache-2.0, which permits commercial use and modification. You are responsible for any safety fine-tuning before you deploy it.

Where can I download Inkling?

The full weights are on Hugging Face at thinkingmachines/inkling. Quantized GGUF builds from the community appear there too.

Does Inkling really have a 1M-token context window?

The official page lists 64K by default and 256K through the Tinker platform. The 1M figure in some coverage isn't confirmed on the model card, so plan around 64K/256K.

Inkling vs DeepSeek or Qwen, which is better?

It depends on the job, not a single score. Inkling's edge is native multimodal input (text, image, audio) in one Apache-2.0 model plus the Tinker fine-tuning path; the leading Chinese open models remain strong general and coding choices. Benchmark them on your own task before committing.

What is Tinker, and do I need it?

Tinker is Thinking Machines' hosted fine-tuning platform. You don't need it to run the open weights, but it's the official route for customizing Inkling and it raises the context window to 256K.

---

Related reading

• Kimi K3: what's leaked and what to use right now
• Grok Build goes open source: what the code lets you check