Inkling: Thinking Machines' First Open-Weights Model

Thinking Machines, the lab founded by former OpenAI CTO Mira Murati, has released its first model: Inkling. Unlike the flagship models from OpenAI, Anthropic, and Google, Inkling ships with full weights available for download, so developers can modify it directly rather than only calling it through a metered API.

Inkling is a Mixture-of-Experts (MoE) transformer with 975B total parameters and 41B active, a 1M-token context window, and pretraining on 45 trillion tokens of text, images, audio, and video.

Thinking Machines is upfront that it isn't the strongest model available: the company states plainly that Inkling is 'not the strongest overall model available today, open or closed.' The pitch is well-rounded performance plus a design meant to be customized through Tinker, the company's fine-tuning platform.

In this article, I'll cover everything new with Inkling, looking at its features, exploring the benchmarks, and proposing hands-on tests you can run yourself. 

What Is Inkling?

Inkling is an open-weights MoE transformer with 975B total parameters and 41B active per token, positioned as a broad base model for fine-tuning rather than a top-of-leaderboard chatbot. It's the first release in a planned family, released July 15, 2026, alongside a preview of a smaller sibling, Inkling-Small.

The design goal is breadth. Thinking Machines trained Inkling across agentic, reasoning, coding, instruction-following, factuality, vision, and audio tasks instead of optimizing for a single domain.

That matters because the company is betting organizations will fine-tune the model on their own data, and a broad base adapts to more workflows than a narrow one.

The headline capability claim is efficiency rather than a peak benchmark. On Terminal Bench 2.1, Thinking Machines reports Inkling spends one-third as many tokens as NVIDIA's Nemotron 3 Ultra to hit the same coding performance.

For a model you might run millions of times inside longer workflows, that token cost curve matters more than a single max-effort score.

How Inkling was built

Inkling's MoE design largely follows DeepSeek-V3, with 256 routed experts, 2 shared experts, and 6 routed experts active per token, using a sigmoid router with auxiliary-loss-free load balancing.

Thinking Machines interleaves sliding-window and global attention at a 5:1 ratio and uses relative positional embeddings instead of RoPE, reporting better extrapolation to longer sequences.

Pretraining ran on 45 trillion multimodal tokens using a hybrid optimizer: Muon for large matrix weights and Adam for other parameters.

Post-training was bootstrapped with a small amount of SFT on synthetic data from open-weights models, including Kimi K2.5, followed by large-scale asynchronous RL scaled to over 30M rollouts across two continuous runs.

Thinking Machines Inkling Key Features

Inkling's differentiators sit less in raw capability and more in how it's built to be customized and controlled. Here are the features that stand out.

Dial thinking effort up or down

Inkling lets you set a "thinking effort" level to trade thoroughness against token cost and latency.

You can push effort to 0.99 for maximum performance on a hard problem, or drop it low for a high-volume task where speed and cost dominate.

This is trained behavior, not a wrapper.

During large-scale reinforcement learning, Thinking Machines varied the effort level per sample by changing the system message and adjusting the per-token cost, so the model learned to spend different token budgets on different rollouts.

The effort setting can be passed from inside a coding or agent harness.

For practitioners running production workloads, this is the feature I'd care about most. Latency is often the binding constraint for interactive tools, and being able to control it per request without swapping models is genuinely useful.

Native multimodal input across text, images, and audio

Inkling reasons natively over text, images, and audio, with the multimodal components trained from scratch on general-domain data.

It can transcribe speech, follow spoken instructions, answer questions about recordings, describe images, and reason over charts and diagrams.

The architecture is encoder-free.

Audio is fed in as dMel spectrograms, images are encoded as 40x40 pixel patches through a four-layer hMLP, and both pass through a lightweight embedding layer to be processed jointly with text tokens.

On the vision side, Inkling can call a Python tool to zoom and crop while reasoning, combining visual and code-based reasoning in one pass.

One caveat worth flagging: input is multimodal, but output in the current tooling is text-only. If your use case needs the model to generate images or audio directly, Inkling won't do that.

Calibrated answers that admit uncertainty

Inkling is trained to express the right amount of confidence, including saying "I don't know" when it's likely to be wrong, which Thinking Machines groups under the model's "epistemics."

The company trained for calibration using reinforcement learning against proper scoring rules on a large corpus of resolved real-world questions.

They used two automated graders in post-training: a rubric grader that checks a response against a checklist of what a good answer should contain, and a claims grader that verifies each factual claim through agentic web search.

The claims grader penalizes claims that don't check out, which reduces hallucination instead of just rewarding recall.

Why this matters is that a model confident in every answer forces you to double-check everything. On the ForecastBench Brier Index without search, Inkling scores 61.1, matching Grok 4.3 and edging out GPT-5.5 at 59.1 and Gemini 3.1 Pro at 54.6.

Strong built-in safety for an open-weights model

Inkling posts the strongest FORTRESS adversarial safety score among the open-weights models Thinking Machines compared, at 78.0%. FORTRESS tests refusal of requests related to weapons and violence alongside benign look-alike queries, so a high score means refusing harmful requests without over-refusing safe ones.

For comparison, Nemotron 3 Ultra scored 77.6%, Kimi K2.5 scored 54.1%, GLM 5.2 scored 65.6%, and DeepSeek V4 Pro scored 36.0% on the same adversarial set.

Inkling also clears 98.6% on StrongREJECT, a refusal test of unambiguous harmful requests, in line with other open and closed models.

The open question is what happens after fine-tuning.

Thinking Machines says it's still studying how safety behavior shifts when customers fine-tune on Tinker, which is a real concern for any downloadable model.

Inkling Benchmarks

Inkling lands mid-pack among open-weights models: ahead of Nemotron 3 Ultra and Kimi K2.5 on most reasoning and coding tests, but trailing GLM 5.2 and Kimi K2.6 on the hardest ones. Against closed frontier models like GPT-5.6 Sol, Gemini 3.1 Pro, and Claude Fable 5, it consistently trails. All scores below are reported at effort 0.99.

Source

SWE-Bench Verified and Terminal Bench 2.1

Terminal Bench 2.1 tests agentic coding in a terminal environment.

Inkling scores 63.8% on its best harness, ahead of Nemotron 3 Ultra at 56.4% and Kimi K2.5 at 51.3%, but well behind GLM 5.2 at 82.7% and GPT-5.6 Sol at 89.5%.

This is where the token-efficiency claim comes in: Inkling reaches Nemotron 3 Ultra's coding performance using roughly a third of the tokens.

Humanity's Last Exam and GPQA Diamond

Humanity's Last Exam (HLE) is a hard, broad reasoning benchmark.

Inkling scores 29.7% text-only and 46.0% with tools, ahead of Nemotron 3 Ultra (26.6% / 37.4%) but behind GLM 5.2 (40.1% / 54.7%) and the closed models, with GPT-5.6 Sol at 47.2% text-only.

GPQA Diamond tests graduate-level science questions.

Factuality: SimpleQA Verified and AA Omniscience

Inkling scores 43.9% on SimpleQA Verified, ahead of Nemotron 3 Ultra at 32.4% and Kimi K2.5 at 36.9%, but far behind Gemini 3.1 Pro at 77.3% and GPT-5.6 Sol at 71.6%. Raw factual recall isn't its strength.

Where Inkling stands out is AA Omniscience, which rewards knowing what you don't know.

Inkling scores a positive 2.1, while several open models go negative: Kimi K2.5 at -8.0, DeepSeek V4 Pro at -10.0, and Nemotron 3 Ultra at -1.0. That gap reflects the calibration training paying off.

Multimodal: MMMU Pro, Charxiv RQ, and audio

On MMMU Pro (Standard 10), a vision reasoning benchmark, Inkling scores 73.5%, behind Kimi K2.6 at 79.0% and Gemini 3.1 Pro at 82.0%.

On Charxiv RQ for chart reasoning, it hits 78.1%, rising to 82.0% when allowed a Python tool.

For audio, Inkling scores 91.4% on VoiceBench, 77.2% on MMAU, and 56.6% on Audio MC, placing it among the strongest open-weights audio models.

Gemini 3.1 Pro still leads on all three, at 94.3%, 82.5%, and 66.8% respectively.

Inkling Pricing and Availability

Inkling is available for fine-tuning on Tinker today, with context length options of 64K and 256K tokens, and full weights are on Hugging Face as both the original checkpoint and an NVFP4 checkpoint for NVIDIA Blackwell inference.

Thinking Machines is offering a 50% discount for a limited time, with full pricing in its documentation.

With this discount applied, pricing per 1 million tokens breaks down as follows:

• Standard (64K context): $1.87 for input/prefill ($0.374 if cached), $4.68 for output generation, and $5.61 for training.
• Extended (256K context): $3.74 for input/prefill ($0.748 if cached), $9.36 for output generation, and $11.23 for training.

(Note: While Tinker is increasing prices for most models on July 17, Inkling's rates are shielded from this hike. Standard platform storage fees of $0.10 per GB/month still apply for fine-tuning).

Deployment support is broad. Fine-tuned checkpoints run via APIs on TogetherAI, Fireworks, Modal, Databricks, and Baseten, with open-source inference support in SGLang, vLLM, and llama.cpp, plus a chat interface with integrated web search in the Inkling Playground, free for a limited time.

Inkling vs. Inkling-Small

Alongside Inkling, Thinking Machines previewed Inkling-Small, a 276B-parameter MoE with 12B active versus 41B for the full model. It matches or exceeds its larger sibling on several benchmarks thanks to improvements to the smaller model's pre-training data and recipe.

• HLE with tools: Inkling-Small 46.6% vs Inkling 46.0%
• GPQA Diamond: Inkling-Small 88.3% vs Inkling 87.2%
• IFBench: Inkling-Small 83.4% vs Inkling 79.8%
• SimpleQA Verified: Inkling-Small 20.9% vs Inkling 43.9%
• Terminal Bench 2.1: Inkling-Small 52.7% vs Inkling 63.8%

Inkling-Small trades factual recall and hard agentic coding for lower cost and latency, which suits workloads like grading, synthetic data generation, and high-volume coding. Full weights ship once testing is complete.

Final Thoughts

Inkling is a statement about strategy more than a leaderboard win. Thinking Machines is betting that organizations willing to fine-tune and own a model will get more value than they would from a metered call to a closed frontier model, and Inkling is the base they're offering for that.

I find the honesty refreshing. The company says outright that Inkling isn't the strongest model, which is unusual for a launch, and the benchmarks back that up: it trails GLM 5.2 and Kimi K2.6 on the hardest coding and reasoning tests, and it sits well behind closed models like GPT-5.6 Sol and Gemini 3.1 Pro. What it offers instead is breadth, native multimodal input, calibrated answers, controllable thinking effort, and the best FORTRESS adversarial safety score among the open-weights models compared.

If you want a top score on a single benchmark, Inkling isn't your pick. If you want an open-weights base to fine-tune on your own data, with strong calibration and safety out of the box, it's a reasonable choice, though the per-token pricing and default verbosity are worth watching for high-volume deployments.

If you're keen to build the fine-tuning skills that make a model like Inkling useful, I recommend starting with our AI Fundamentals skill track.