Inkling Review: Thinking Machines First Open-Weights Model (July 2026)

Thinking Machines Lab released Inkling on 15 July 2026: a 975B-total / 41B-active Mixture-of-Experts model under Apache 2.0, with up to 1M-token context and a continuous thinking-effort dial. Almost every article about it restates the same spec sheet. This one does something the coverage did not: it tags every headline number as vendor-run or externally sourced, flags a benchmark figure that two sources disagree on, and converts the effort dial into dollars.

What Thinking Machines actually shipped

Inkling landed on Wednesday 15 July 2026, with coverage running into the 16th. It is the first open-weights model from Thinking Machines Lab, the company founded by former OpenAI CTO Mira Murati, which TechCrunch reports has roughly 200 employees after a round of departures in early 2026 including two co-founders who left for OpenAI in January.

The architecture, as described on the model card and in the launch write-ups: a 66-layer decoder-only transformer with sparse Mixture-of-Experts feed-forward layers, 975B total parameters and 41B active per token. Each token is routed to 6 of 256 experts, plus 2 shared experts that fire on every token. Context runs up to 1M tokens for the downloadable weights, while the hosted Tinker offering exposes 64K and 256K options. Pretraining covered 45 trillion tokens of text, images, audio and video. It accepts text, image and audio input and emits UTF-8 text only — no image or audio generation.

The license is Apache 2.0, confirmed on both the model card and the Hugging Face repo. That is a real permissive grant: commercial use, modification, fine-tuning and redistribution, with none of the usage carve-outs that make some so-called open models awkward for commercial deployment. Apache 2.0 weights are still not open source in the OSI sense, because the training data and pipeline are not published — a distinction we work through model by model in the best open source LLMs of 2026, ranked and licensed.

The distinctive feature is the thinking effort control: a continuous parameter that scales how much the model reasons before answering. Thinking Machines says effort was trained in during reinforcement learning by varying the system message and the per-token cost, and the published sweep range runs 0.2 to 0.99. In the Hugging Face transformers integration it surfaces as named levels — none, minimal, low, medium, high, xhigh, max — with published numeric mappings of none at 0, low at 0.2, medium at 0.7, high at 0.9, and xhigh and max at 0.99. Note that none sits at 0, below the swept 0.2 floor, so the dial is not literally bounded at 0.2.

The most useful thing in the launch materials is a sentence from the vendor itself. Thinking Machines writes that Inkling is not the strongest overall model available today, open or closed, and positions it instead as a base for customization: multimodal, token-efficient, and fine-tunable on Tinker from day one with a limited-time 50% discount and a playground that is free for a limited time. TechCrunch frames Tinker as the company's revenue path. Hold that thought — it explains a lot about the release.

Vendor numbers vs independent numbers

Here is the problem with the Inkling coverage. Thinking Machines published a model-card benchmark table. Artificial Analysis published an independent evaluation. Press coverage blended the two, and several outlets — plus the research brief that started this article — attributed the 77.6% SWE-bench Verified score to Artificial Analysis. AA never reported a SWE-bench Verified figure for Inkling. That number is the vendor's own, measured at effort 0.99 and temperature 1.0.

It cuts the other way too, and almost nobody noticed: Thinking Machines states in its own launch post that it relies on externally reported evaluations — naming Artificial Analysis — for HLE, GPQA Diamond, GDPval, τ³ Banking, AA-Omniscience and MMMU Pro, for both its own model and its competitors. So parts of the "vendor table" are not vendor-run at all, and the clean two-column story most coverage told is wrong in both directions.

So the first original asset here is not another spec table. It is the same numbers with a provenance column attached.

Table: DataLLM Lab. Figures as published July 2026. Provenance column is ours; the underlying numbers are from the Thinking Machines model card and launch post and the Artificial Analysis launch evaluation.

Read the table top to bottom and a different story emerges from the launch headlines. The independently verified wins are about efficiency and agentic behaviour, not raw capability. The headline capability numbers that are genuinely vendor-run — SWE-bench Verified, AIME, Terminal Bench — were produced at maximum effort, with TM's own harness in the Terminal Bench case, against competitor figures TM took from those competitors' reports. That is a structurally flattering comparison, and Thinking Machines is more candid about it than most: it publishes effort-sweep curves for Inkling on Terminal Bench 2.1, HLE and IFBench, where competitors are represented by single externally reported results. A curve will always look good against a dot.

The Nemotron figure nobody checked

The most repeated line in the coverage is some version of "Inkling's 77.6% on SWE-bench Verified beats Nemotron 3 Ultra's 71.9%." Go to the source table and that comparison number is not there.

Table: DataLLM Lab. Cross-source reconciliation, July 2026.

A 1.2-point gap does not change who is ahead. It matters because of what it reveals: cross-vendor benchmark numbers are not apples-to-apples even on a benchmark as standardised as SWE-bench Verified. Harness differences, scaffold differences, checkpoint and quantisation differences all move the score by a point or two. Thinking Machines said as much about its own Terminal Bench runs. If a 1.2-point discrepancy can survive into near-universal repetition, so can a 5-point one, in either direction. Our take on Nemotron in isolation is in the Nemotron 3 Ultra review, and the broader field is ranked in the best open source LLMs of 2026.

The practical rule: never let a single cross-vendor benchmark decide a model choice. Run your own eval on your own harness, on your own tasks, and treat leaderboards as a shortlist generator only.

The effort dial is a cost dial

Every article about Inkling describes the effort parameter. None of them price it. That is the gap, because output tokens are what you actually pay for, and a knob that controls output token volume is a knob that controls your bill.

Hold price constant and the arithmetic is blunt: 25K output tokens per task at $4.68 per 1M is about $0.117 a task, or $117 per thousand tasks. At 38K tokens the same work costs about $178, and at 43K about $201. Inkling's independently measured verbosity advantage is worth roughly 40% of the output bill against the most verbose comparison point in AA's set — before you touch the effort dial at all.

Now add the dial. Because effort directly scales generated tokens, moving from a high setting to a medium one on tasks that do not need deep reasoning cuts the output bill proportionally. The buying question stops being "is Inkling ranked above model X" and becomes "what is the lowest effort setting that still passes my eval?" That is a far more tractable question, and it is the one nobody in the launch coverage asked.

A workable procedure: build a 50-task eval from your real traffic, run it at low, medium and high, and record pass rate and mean output tokens at each. You will usually find one or two task classes that need high effort and a long tail that does not. Route accordingly. We walk through the same measurement method for coding workloads in our LLM coding cost benchmark.

Open weights you probably cannot self-host

Apache 2.0 on a 975B-parameter model sounds like sovereignty. Check the hardware bill before you plan around it. Thinking Machines lists a minimum of roughly 2 TB of aggregated VRAM for the full BF16 checkpoint — 8x NVIDIA B300 or 16x H200 — and roughly 600 GB for the NVFP4-quantised checkpoint, or 4x B300 / 8x H200. Computerworld and MarkTechPost both repeat those configurations. Community GGUF quantisations exist for llama.cpp, but the vendor's own floor is the number to plan against.

For the overwhelming majority of teams, "open weights" here means: you can legally fine-tune it, redistribute your derivative, and audit the architecture — but you will rent the compute. Distribution reflects that. Weights sit on Hugging Face; hosted APIs come from Together AI, Fireworks, Modal, Databricks and Baseten among others; runtime support exists in SGLang, vLLM, llama.cpp and Hugging Face Transformers; and NVIDIA Build lists it, with BF16 requiring Hopper or later and NVFP4 targeting Blackwell. Inkling was itself trained on NVIDIA GB300 NVL72 systems.

This is also the quiet business logic. Tinker, the fine-tuning platform, is where TechCrunch says the revenue is meant to come from — training, fine-tuning, and a cut of the hosting ecosystem around the model. Giving away weights that almost nobody can serve at scale, while selling the fine-tuning and hosting layer, is a coherent strategy — and it is worth naming, because VentureBeat notes the company has not explained how it will cover its training costs, and TechCrunch reports a roughly $50B round said to be in progress as of November 2025 that had stalled by January, with the company declining to discuss funding since. Those are reported figures about a private company, not confirmed facts.

One more framing worth attributing rather than endorsing: VentureBeat built its coverage around low cost and resistance to censorship, reflecting Thinking Machines' emphasis on epistemics — calibration, instruction following, and answering directly on topics that may be subject to censorship. TM says the model showed strong patterns of censorship non-compliance when evaluated by Cognition on its Propaganda and Censorship Eval; that is a third-party eval, but it reaches you through the vendor. On safety, the model card's conclusion that Inkling did not present risk of material uplift beyond what is already available in the open-weight ecosystem is a vendor summary of its own testing, and TM acknowledges residual risks including occasional compliance with role-play and indirectly framed prompts on harmful topics. The published FORTRESS figures — 78.0% adversarial, 95.9% benign — are likewise from the model card. If licensing and distribution policy is your angle, see export controls and open weights.

Who should actually use Inkling

Take the vendor at its word. Inkling is not the strongest model available, and it is not trying to be. It is a customization base with three properties that are genuinely hard to find together: permissive licensing, native multimodal input across text, images and audio, and independently confirmed token efficiency.

A decision rule, in order:

1. You need the best raw capability and cost is secondary. Inkling is the wrong answer. Thinking Machines' own table puts several closed models well ahead on SWE-bench Verified, and three open-weights models — DeepSeek V4 Pro at 80.6%, Kimi K2.6 at 80.2% and GLM 5.2 at 80.0% — ahead of Inkling's 77.6%.
2. You need a fine-tunable base with a clean license and multimodal input. Strong fit. Apache 2.0 plus audio and image understanding plus day-one Tinker support is the actual pitch, and it is a good one.
3. Your workload is high-volume and verbosity-sensitive. Strong fit. The ~25K-token average is independently measured by Artificial Analysis, not a marketing claim, and the effort dial gives you a lever most models do not expose.
4. Your workload is knowledge-heavy and hallucination-sensitive. Be careful. AA's Omniscience result places Inkling at +2, ahead of Nemotron 3 Ultra at -1 but behind the leading open-weights field, with 40% accuracy and a 63% hallucination rate on that benchmark's construct. That metric measures behaviour on deliberately hard knowledge probes — it is not a claim that the model is wrong 63% of the time in general — but it is the honest counterweight to the launch narrative.
5. You want a drop-in for a Chinese open-weights model on leaderboard grounds alone. Re-read the U.S. qualifier. AA's framing was the new leading U.S. open weights model, and that scoping is doing real work.

Independent hands-on colour, for calibration: Simon Willison ran Inkling through the API to generate an SVG of a pelican on a bicycle, then fed the rendering back for description. The model described colours and composition accurately but called its own drawing a stylized white bird resembling a stork or seagull. He frames the release as a new viable contender in the U.S. open-weights ecosystem alongside NVIDIA Nemotron and Gemma 4 — which is, stripped of launch-day framing, exactly what it is.

As for Inkling-Small: 276B total / 12B active, announced as a preview, weights not released at launch while TM finishes testing. The claim that it matches or exceeds the larger model on many benchmarks is vendor-only and unverified. If it ships and holds up, it is the more interesting model for most readers, because 12B active is a serving proposition rather than a datacentre one. Until then it is an announcement. For where it would land against the current field, see our Kimi K3 review and the 2026 open-weights roundup.

FAQ