Thinking Machines Built a Dial, Not a Toggle. That Breaks the Way We Usually Test Effort.

Thinking Machines shipped its first open-weight model on July 15, and the headline spec isn't the parameter count. It's a thinking-effort control that runs on a continuous scale from 0.2 to 0.99, not a three-step low/medium/high menu like almost everything else we test. That single design choice undercuts the assumption behind our own does-thinking-effort-matter piece: that effort is a small number of discrete settings you can just enumerate and vote on.

What Inkling actually is

Inkling is a mixture-of-experts model with 975 billion total parameters and 41 billion active per token, pretrained on 45 trillion tokens of text, image, audio, and video. It supports a 1 million token context window and ships in two checkpoint formats: BF16, which needs roughly 2TB of aggregate GPU memory to run, and NVFP4 for Nvidia Blackwell hardware, which drops that to about 600GB. A smaller sibling, Inkling-Small (276B total, 12B active), is in preview with weights to follow after more testing.

• Pricing: $1.87 per million input tokens, $4.68 per million output, roughly $1.10 blended at a 7:2:1 cache-hit:input:output ratio
• Reported scores: 41 on the Artificial Analysis Intelligence Index (open-weight median is ~25), 77.6% on SWE-bench Verified, 91.4% on VoiceBench
• Distribution: full weights on Hugging Face, inference live on Together AI, Fireworks, Modal, Databricks, and Baseten
• Fine-tuning access through Tinker, Thinking Machines' own customization platform, from day one

Those are the vendor's numbers, and they're the kind of numbers we're generally wary of printing without a live run behind them. What's new here isn't the score, it's the shape of the effort control underneath it.

A dial changes what 'testing effort' means

Every effort test we've published so far, including our own, treats thinking effort as a small, enumerable set: a model ships two or three named tiers, you run the same prompt at each, and you compare. That works because the vendor already collapsed a continuous underlying behavior into a handful of marketed checkpoints. Inkling doesn't do that collapsing for you. The knob goes from 0.2 to 0.99, which is not three settings, it's dozens of usable ones, and Thinking Machines is explicit that the model is meant to be tuned per deployment through Tinker rather than shipped as one fixed personality.

Inkling is 'not the strongest overall model available today, open or closed.' ... [it's] less as a finished product than as a starting point, something for organizations to fine-tune themselves through Tinker — TechCrunch, July 15, 2026

That's an unusually honest thing for a model launch to say out loud. Most releases lead with a leaderboard number. This one leads with 'we're not trying to win the leaderboard, we're trying to be a starting point you reshape.' It's a coherent strategy, but it's also a strategy that a blind one-shot arena vote is a poor instrument for.

What a blind vote can and can't tell you here

Our arena is built to answer one question cleanly: given the same prompt, run at the same moment, which output do people actually prefer, with no logo attached to bias the vote. That's a fair test of a fixed model at a fixed setting. It says almost nothing about a model whose entire pitch is that the setting shouldn't be fixed, and that the real product is what you get after fine-tuning it on your own data through Tinker. Voting on Inkling straight out of the box measures the starting point, not the thing Thinking Machines is actually selling.

This matters beyond one model. Open-weights releases have been arriving in clusters this year, GLM-5.2, DeepSeek V4, Qwen 3.6, and now Inkling, and the pattern in each writeup we've done is the same: the vendor's benchmark table is a snapshot of one configuration, not a description of the model. A continuous effort dial just makes that fact impossible to paper over with a tidy three-row table.

It also raises a fair question about SWE-bench Verified specifically. 77.6% is a strong number, but SWE-bench measures patch correctness against a fixed test suite, not code a reviewer would actually want to merge, and it says nothing about which effort setting produced it. Until we run it ourselves, treat it as a ceiling under ideal conditions, not an expectation for a default deployment.

For readers following this beat from Europe, nowosci.ai is covering the same open-weights wave in Polish, worth a look if Inkling's Tinker-first pitch is relevant to your stack.

The short version: a model that ships with a 0.2-to-0.99 dial instead of a three-tier menu isn't just a UX difference. It's an admission that 'thinking effort' was always a continuum vendors were rounding down to a marketing-friendly toggle. Inkling stopped rounding. Our testing needs to stop treating effort as three checkboxes too, at least for models built this way.