Inkling by Thinking Machines: 975B Open-Weight Multimodal Model With Controllable Reasoning

On July 15, 2026, Thinking Machines Lab announced Inkling — its first in-house model, and a statement of intent. Founded by former OpenAI CTO Mira Murati, the company has spent 18 months building AI infrastructure largely out of public view. Inkling is that first proof point: a 975-billion-parameter mixture-of-experts model, trained on 45 trillion tokens of text, image, audio, and video, released under Apache 2.0 with full open weights on Hugging Face.

What makes Inkling worth paying attention to is not that it's the strongest model available — Thinking Machines explicitly says it isn't — but that it represents a fundamentally different approach to how enterprises should adopt AI. Rather than selling metered API access to a black-box model, Thinking Machines is giving you the weights and saying: "Make it yours."

The model is designed to be a starting point for organizations to fine-tune through Tinker, Thinking Machines' model-customization platform. It's the company's bet that AI shaped by organizations for their own use cases will outperform the one-size-fits-all models currently dominating the market.

Architecture: Sparse MoE at Scale

Inkling uses a sparse Mixture-of-Experts architecture — the same family of designs powering DeepSeek V3 and Mixtral, but pushed to a much larger scale. The key numbers:

• 66-layer decoder-only transformer with hybrid local and global attention layers
• 256 experts per feed-forward layer, with 6 routed per token plus 2 shared experts active on every token
• 975B total parameters, but only 41B active per forward pass
• 1M token context window
• Support for BF16, MXFP8, and NVFP4 numerics

The sparse routing is what makes this architecture economically viable. Despite having nearly a trillion parameters, Inkling only activates about 4% of them for any given token — similar to running a 41B dense model, but with access to a much larger parameter pool for specialized knowledge.

Natively Multimodal

Unlike models that bolt on vision or audio capabilities as afterthoughts, Inkling was designed as multimodal from the ground up. All modalities are projected into a shared hidden space and processed jointly by the decoder:

• Text: UTF-8 encoded text input
• Image: Any pixel-based input via a hierarchical patch encoder (optimal: 40px–4096px per dimension)
• Audio: WAV format at 16kHz via discrete token encoding (optimal: up to 20 minutes)

Training data spans text, images, audio, and video — 45 trillion tokens total. However, outputs are currently limited to text, including code, styled artifacts, and structured data.

Controllable Thinking Effort

One of Inkling's standout features is the "thinking effort" dial — a mechanism that lets users control how much computational effort the model invests in reasoning before producing an answer. The effort parameter ranges from 0 to 1, with higher values meaning more reasoning tokens and potentially better answers.

This is a practical feature for production deployments. Not every query needs maximum reasoning effort — a simple factual question doesn't require the same depth as a complex coding task. By dialing effort up or down, organizations can optimize the speed-cost-quality tradeoff for each use case.

Thinking Machines claims that on one benchmark, Inkling uses a third as many tokens as Nvidia's Nemotron 3 Ultra to achieve the same coding performance — a significant efficiency advantage.

The model is also designed for calibrated uncertainty — it's trained to flag when it's uncertain rather than guessing, which is critical for enterprise deployments where hallucinated answers can have real consequences.

Benchmark Performance

Thinking Machines has published comprehensive benchmarks comparing Inkling (at effort=0.99) against both open-weight and closed-weight competitors. Here's the breakdown:

Reasoning

On mathematical and reasoning benchmarks, Inkling holds its own against open-weight peers but trails the strongest closed models:

• AIME 2026: 97.1% (vs. 99.9% for Claude Fable 5 and GPT 5.6 Sol)
• GPQA Diamond: 87.2% (vs. 94.1% for Gemini 3.1 Pro and GPT 5.6 Sol)
• HLE text only: 29.7% (vs. 47.2% for GPT 5.6 Sol)
• HLE with tools: 46.0% (vs. 64.5% for Claude Fable 5)

Agentic Coding

Strong performance in coding, particularly notable for an open-weight model:

• SWEBench Verified: 77.6% (competitive with GLM 5.2 at 80.0% and DeepSeek V4 Pro at 80.6%)
• SWEBench Pro Public: 54.3% (vs. 80.0% for Claude Fable 5)
• Terminal Bench 2.1: 63.8% (vs. 89.5% for GPT 5.6 Sol)

Agentic General

On general agentic tasks, Inkling shows solid but not leading performance:

• MCP Atlas: 74.1% (vs. 83.3% for Claude Fable 5)
• BrowseComp: 77.1% (vs. 90.8% for GPT 5.6 Sol)
• Tau 3 Banking: 23.7% (vs. 33.0% for GPT 5.6 Sol)

Vision

Competitive multimodal understanding:

• MMMU Pro Standard: 73.5% (vs. 84.2% for Claude Fable 5)
• Charxiv RQ: 78.1% (vs. 86.5% for Claude Fable 5)

Audio

Here's where Inkling stands out — as one of the few open-weight models with native audio capabilities:

• MMAU: 77.2%
• VoiceBench: 91.4%
• Audio MC: 56.6%

Hardware Requirements

Inkling is distributed in two checkpoint formats, each with different hardware requirements:

BF16 Checkpoint (Full Precision)

• Requires 2 TB aggregated VRAM
• 8x NVIDIA B300 GPUs, or
• 16x NVIDIA H200 GPUs

NVFP4 Checkpoint (Quantized)

• Requires 600 GB aggregated VRAM
• 4x NVIDIA B300 GPUs (W4A4 mode, requires SM100+ architecture)
• 8x NVIDIA H200 GPUs (W4A16 mode)

For inference, the model supports deployment through SGLang, vLLM, TokenSpeed, Unsloth, or Hugging Face — the major inference frameworks in the ecosystem.

Thinking Machines: The Company Behind the Model

Thinking Machines Lab was founded by Mira Murati, former CTO of OpenAI, along with a team of AI researchers and engineers. The company has roughly 200 employees as of July 2026, though it experienced departures earlier this year, including two co-founders who left for OpenAI in January.

The company's central thesis: AI that organizations can adapt for themselves will outperform the one-size-fits-all models the biggest labs currently sell. This is a direct challenge to the API-as-a-service model that OpenAI, Anthropic, and Google have built their businesses around.

Tinker: The Fine-Tuning Platform

Inkling isn't sold as a standalone product — it's the foundation for Tinker, Thinking Machines' platform for model customization. The revenue model is built around training, fine-tuning, and hosting services rather than metered API calls to a base model.

This means:

• Organizations download Inkling's weights (Apache 2.0 — free for commercial use)
• They fine-tune on their own data through Tinker
• Thinking Machines earns revenue from the fine-tuning infrastructure, not from the model itself

Training Infrastructure

Inkling was trained entirely on Nvidia's GB300 NVL72 systems. Thinking Machines struck a partnership with Nvidia in March 2026 to deploy a gigawatt of Vera Rubin computing capacity. The company hasn't disclosed training costs, and a reported $50 billion fundraising round that was coming together in November 2025 had stalled by January 2026.

Speed to Market

Thinking Machines emphasizes how quickly it got here. OpenAI took roughly five years to bring its technology to market and show revenue. Anthropic took roughly three. Thinking Machines claims it achieved comparable results in about nine months.

The Distillation Question

Thinking Machines has been transparent about its training methodology: Inkling was pre-trained from scratch, but the company used other open-weight models — including Moonshot AI's Kimi K2.5 — to help generate early post-training data before large-scale reinforcement learning took over. The company states its next model will use fully self-contained post-training instead.

The Open vs. Closed Debate: Why This Matters

Inkling's release comes at a moment when the industry is re-examining the economics of closed AI models. Several voices have recently argued against the proprietary model approach:

Microsoft CEO Satya Nadella — whose company has invested billions in both OpenAI and Anthropic — warned in a blog post that enterprises using proprietary AI models "effectively pay twice": once in subscription costs, and again by handing over business knowledge embedded in their prompts and corrections, which can be absorbed into future model versions. [TechCrunch Coverage]

Hugging Face CEO Clem Delangue predicted that frontier models will increasingly be reserved for experimentation and high-value tasks, while most production AI work shifts to private or open-source alternatives — the exact split Thinking Machines is building around.

The clearest real-world example came from a recent project with Bridgewater Associates, the world's largest hedge fund. Researchers from both Bridgewater and Thinking Machines took an existing open-source model and trained it further on Bridgewater's own financial expertise. The result scored 84.7% on financial reasoning tests, beating top proprietary AI models, while costing roughly one-fourteenth as much to run.

Safety and Responsible Use

Thinking Machines conducted safety evaluations ahead of release, covering:

• Everyday interaction: Sycophancy, harmful manipulation, parasocial dependency, and validation of delusional reasoning — including multi-turn external red-teaming
• CBRN and cyber: Knowledge and procedural uplift through internal evaluations, external testing, and refusal-suppressed variants
• Loss of control: Agentic capability, strategic deception, and sabotage potential — benchmarked against public frontier models

The company concluded that Inkling "did not present risk of material uplift beyond what's already available in the open-weight ecosystem." Residual risks include occasional compliance with role-play and indirectly framed prompts concerning harmful topics — consistent with any open-weight model.

Thinking Machines recommends defense-in-depth rather than relying on the model's built-in refusals alone. They suggest layering downstream moderation tools like Llama Guard around Inkling for production deployments.

Getting Started with Inkling

The model weights are available on Hugging Face under Apache 2.0. For fine-tuning, Thinking Machines provides the Tinker platform with a cookbook and the tml-renderers Python package.

Key resources:

• Model Card: thinkingmachines.ai/model-card/inkling/ — full technical specifications
• Announcement: thinkingmachines.ai/news/introducing-inkling/ — official blog post with benchmarks
• Tinker Cookbook: Step-by-step guide for fine-tuning on your own data
• Inkling Playground: Interactive demo for testing the model

The original announcement tweet by @thinkymachines garnered 5.3M views, signaling strong community interest in the release.

What to Watch

Inkling is Thinking Machines' first model — and the company has already indicated its next release will use fully self-contained post-training. Key things to monitor:

• Tinker adoption: How quickly do enterprises adopt the fine-tuning platform? Is the "make it yours" model compelling enough to drive revenue?
• Next model iteration: Thinking Machines has signaled improvements in post-training methodology. Expect a stronger successor.
• Funding picture: The stalled $50B raise is a wild card. Can the company sustain its compute commitments?
• Community fine-tunes: With Apache 2.0 weights, expect the open-source community to produce specialized variants rapidly.
• Audio capabilities: Native audio reasoning is rare in open-weight models. Watch for innovative applications in voice AI, meeting transcription, and audio analysis.

Inkling may not be the strongest model available today. But it represents a meaningful vote of confidence in the open-weight approach — from a company founded by one of the most prominent figures in the AI industry. If Thinking Machines' thesis is correct, the models that matter most in 2027 won't be the ones you can't touch — they'll be the ones you've shaped for your own needs.