Inkling: Thinking Machines Releases 975B Open-Weights Multimodal Model

Thinking Machines Lab has entered the open-weights arena with a serious contender. Inkling, announced today, is a Mixture-of-Experts transformer with 975 billion total parameters and 41 billion active per forward pass. It supports a context window of up to 1 million tokens and was pretrained on 45 trillion tokens spanning text, images, audio, and video. Perhaps most notably, the full model weights are available — and the model is designed from the ground up to be customized.

Not the Strongest Model — The Most Customizable One

Thinking Machines is refreshingly honest about where Inkling stands. By their own admission, Inkling is not the strongest overall model available today, whether open or closed. Instead, the team has optimized for a combination of qualities that make it an excellent base for fine-tuning: multimodal capabilities, efficient controllable thinking, and availability on their Tinker platform for customization.

This is a fundamentally different value proposition than what most labs are offering. Rather than chasing the top of a leaderboard, Thinking Machines is betting that developers want a model they can adapt to their specific workflows — and that the best model for your use case is one you have shaped yourself.

Architecture and Training

Inkling is a Mixture-of-Experts (MoE) model, a design choice that allows massive total parameter counts while keeping active compute per token manageable. At 975B total with only 41B active, the efficiency ratio is striking — roughly 4.2% of the model lights up for any given token. This means inference costs are closer to what you would expect from a 41B dense model, while the knowledge capacity benefits from the full 975B.

The training corpus is equally ambitious: 45 trillion tokens across text, images, audio, and video. Inkling processes audio as dMel spectrograms and images as 40x40 pixel patches through a lightweight embedding layer, using an encoder-free architecture that keeps the pipeline simple and efficient. Both modalities are processed jointly with text tokens, enabling true native multimodal reasoning rather than bolted-on vision or speech modules.

Alongside the main release, Thinking Machines also previewed Inkling-Small, a lighter variant with 12B active parameters trained with a similar recipe, targeting even lower cost and latency.

Controllable Thinking: The Efficiency Story

One of Inkling's most interesting features is controllable thinking effort. Developers can dial the model's reasoning intensity from 0.2 to 0.99, trading compute for accuracy based on the task at hand. This is not just a gimmick — the efficiency gains are real. According to Thinking Machines' own benchmarks, Inkling matches Nemotron 3 Ultra on Terminal Bench 2.1 (agentic coding) at roughly one-third the tokens. In a world where inference costs dominate AI budgets, that kind of efficiency curve is a compelling story.

The ability to sweep effort dynamically means developers can use Inkling at low effort for quick iterations and crank it up for final answers — a workflow that mirrors how humans actually solve problems. Low latency enables faster feedback loops, which means more iterations, which means better final results.

Agentic Coding and Real-World Demos

Thinking Machines showcased several impressive demos of Inkling's agentic capabilities:

• A one-shot web app build where Inkling created a functional job-application web app in a single pass, then powered an embedded AI assistant that could operate the app through natural language.
• A polished nine-page PDF food and travel journal generated from a single prompt, demonstrating precise instruction following and cohesive multi-page styling.
• A multiplayer snake game refined through 40 iterations of feedback from GPT Codex acting as a reviewer — demonstrating sustained collaborative refinement.

On Design Arena's Agentic Web Dev leaderboard, where blinded human evaluators compare generated web apps head-to-head, Inkling scored 1257 — ranking it among the strongest open-weights models, just behind GLM 5.2 (1275) and ahead of Grok 4.5 (1271).

Epistemics: Calibration Over Confidence

Thinking Machines paid unusual attention to what they call "epistemics" — the model's calibration, instruction following, and resistance to censorship. Rather than training Inkling to always sound confident, they trained it to express appropriate uncertainty, including the ability to say "I don't know" when it should.

This involved RL against proper scoring rules on resolved real-world questions, plus two automated graders during instruction-following training: a rubric grader that scores responses against checklists, and a claims grader that verifies each factual claim through agentic web search. The result is a model that hallucinates less without sacrificing helpfulness — a tradeoff most labs struggle with.

On ForecastBench, Inkling scored 61.1 Brier Index without search and 63.7 with search — competitive with GPT-5.5 (59.1/64.7) and Grok 4.3 (61.7/63.2). For a model that's supposed to be a customizable base rather than a frontier flagship, that's strong.

Safety: Best-in-Class for Open Weights

For an open-weights model, safety is the perennial question. Thinking Machines commissioned external safety testers and published results on FORTRESS (adversarial refusal) and StrongREJECT (harmful request refusal). Inkling scored 78.0% on FORTRESS adversarial and 95.9% on FORTRESS benign — meaning it refuses harmful requests at a higher rate than Nemotron 3 Ultra, Kimi K2.5/K2.6, GLM 5.2, and DeepSeek V4 Pro, while over-refusing fewer benign look-alike queries. On StrongREJECT, it scored 98.6%, in line with other open and closed-weights models.

The model also exhibited strong censorship non-compliance patterns when evaluated by Cognition, meaning it doesn't reflexively refuse politically sensitive topics — a design choice aligned with Thinking Machines' mission of building AI that extends human judgment.

The Tinker Platform and Self-Fine-Tuning

Inkling is available for fine-tuning on Thinking Machines' Tinker platform starting today. In a demonstration that borders on showing off, the team had Inkling fine-tune itself — the model wrote its own fine-tuning job, ran it, and evaluated the results. This self-improvement loop is both a marketing flex and a genuine illustration of what becomes possible when a model is designed for customization from day one.

The Tinker console also includes the Inkling Playground, a developer-facing chat interface for interacting with the model before committing to a fine-tuning pipeline.

What This Means for the Open-Weights Ecosystem

The open-weights landscape in mid-2026 is crowded: DeepSeek V4 Pro, Nemotron 3 Ultra, Kimi K2.6, GLM 5.2, and now Inkling. What sets Inkling apart is not raw benchmark dominance but the combination of:

• Native multimodal training (text, vision, audio) from scratch — not bolted on
• Controllable thinking effort with a smooth efficiency curve
• Best-in-class safety among open-weights models on adversarial benchmarks
• Strong epistemics — calibration and honest uncertainty built into training
• First-class fine-tuning support via Tinker from day one

Thinking Machines describes Inkling as "just the start" — the first in a family of models they will continue to build on. If the next releases improve on the raw benchmark front while maintaining this design philosophy, the open-weights ecosystem is in for an interesting trajectory.

The Bigger Picture

Inkling's release reinforces a trend that's been building throughout 2026: the gap between closed and open-weights models continues to narrow. While frontier models from Anthropic, OpenAI, and Google still lead on raw benchmarks, the combination of multimodal capability, safety, calibration, and customizability in a single open package is something no single closed model fully offers either.

For organizations that need to own their inference pipeline, customize model behavior, or operate without dependency on a single provider's API, Inkling represents a serious option. The fact that it's paired with a fine-tuning platform from day one — rather than being tossed over the wall for the community to figure out — suggests Thinking Machines understands that open weights without tooling is only half the promise.

The real test will be what the community builds with it. Open-weights models live or die by their ecosystem. If Inkling's combination of multimodal reasoning, efficient thinking, and fine-tuning accessibility attracts enough developers, it could carve out a meaningful niche — not as the strongest model, but as the most useful one to make your own.