On July 15, 2026, Thinking Machines Lab — the AI startup founded by ex-OpenAI CTO Mira Murati alongside chief scientist and OpenAI co-founder John Schulman — released Inkling, its first open-weight model. Full weights landed on Hugging Face under Apache 2.0, alongside same-day fine-tuning access on Tinker.
This is a research-based guide. We reviewed Thinking Machines’ official model card, primary coverage from TechCrunch, Axios, Bloomberg, TechTimes, The Next Web, HPCwire, Simon Willison’s notes, and the OpenRouter model page. We did not fine-tune or run the model ourselves.
What Inkling is (and what it explicitly is not)
Thinking Machines published an unusual positioning statement alongside the release: Inkling is not the strongest model available today, open or closed. The company frames that not as a disclaimer but as the design intent. Inkling is a starting point — a large, permissively licensed, multimodal base model meant to be narrowed and sharpened for specific workloads through Tinker, the company’s fine-tuning platform.
The comparison class is Moonshot AI’s Kimi K2.5 and NVIDIA’s Nemotron 3 Ultra — Chinese and NVIDIA open-weight entries that compete for enterprise fine-tuning workloads. TechStrong AI framed Inkling specifically as a US challenge to Chinese open-source dominance in the customizable model tier.
Architecture
According to the official model card, Inkling is a 66-layer decoder-only transformer with sparse Mixture-of-Experts feed-forward layers. Each token is routed to 6 of 256 experts plus 2 shared experts. Total parameters: 975 billion. Active parameters per forward pass: approximately 41 billion — a 4.2% activation ratio, which keeps inference costs roughly in line with a 41B dense model despite the massive knowledge base encoded in the full weight set.
Training data: 45 trillion tokens spanning text, images, audio, and video. Per the model card, the model is natively multimodal — images are encoded via a hierarchical patch encoder and audio via discrete token encoding, with all modalities projected into a shared hidden space and processed jointly by the decoder rather than through separate adapter modules.
Context window: 1 million tokens.
Benchmark numbers
From Thinking Machines’ own release materials, corroborated by HPCwire:
| Benchmark | Score |
|---|---|
| HLE (text-only) | 29.7% |
| HLE (with tools) | 46.0% |
| AIME 2026 | 97.1% |
| GPQA Diamond | 87.2% |
The HLE scores put it behind the current frontier — Claude Fable 5 scores 53% on HLE, more than 20 points above Inkling’s text-only score — which is consistent with Thinking Machines’ own framing of Inkling as not the strongest model available today. The math benchmarks are strong: AIME 2026 at 97.1% is a high score by current frontier standards.
VRAM requirements for self-hosting
Full weights are on Hugging Face in two checkpoints. Per the official model card:
- BF16 checkpoint: Minimum 2 TB aggregate VRAM (e.g., 16× NVIDIA H200 or 8× NVIDIA B300)
- NVFP4 checkpoint: Minimum 600 GB aggregate VRAM (e.g., 8× NVIDIA H200 in W4A16, or 4× NVIDIA B300 in W4A4), optimized for NVIDIA Blackwell systems
Self-hosting frameworks supported: SGLang, vLLM, TokenSpeed, and Hugging Face Transformers, per the Hugging Face model page; quantized variants are also compatible with llama.cpp, Ollama, LM Studio, and Jan. The Apache 2.0 license allows commercial use and derivative works without royalty obligations.
For most builders, 600 GB of VRAM is not a practical starting point. The managed inference and fine-tuning paths below are the realistic entry points.
Managed inference: access without owning a cluster
Inkling is available through several third-party inference providers as of launch day: Together AI announced day-zero support for the full 1M-token context, and TechStartups reported the full launch-day roster:
- TogetherAI
- Fireworks AI
- Modal
- Databricks
- Baseten
Pricing varies by provider. On Thinking Machines’ own Tinker platform, Inkling is priced at:
- Input: $1.87 per 1M tokens (64K context tier)
- Output: $4.68 per 1M tokens
- Cached input: $0.374 per 1M tokens
- 256K context tier: approximately double the above rates
Thinking Machines is offering a 50% discount on Tinker pricing at launch. The revenue model is fine-tuning, not inference — Inkling at list price may not be the most competitive inference option compared to similarly-sized models, but the discount makes it accessible for experimentation.
Fine-tuning with Tinker: the Bridgewater case study
The commercial rationale for Inkling is Tinker, the fine-tuning platform Thinking Machines launched in October 2025. Inkling gives Tinker customers a frontier-adjacent, permissively licensed base model to customize.
Thinking Machines disclosed one customer example around launch. In a joint post with Bridgewater Associates, the two companies describe using Tinker to fine-tune a model on specialized financial-reasoning data. In their own first-party evaluation — not independently replicated — the fine-tuned model averaged 84.7% accuracy versus 78.2% for the best frontier model tested with expert prompting, at roughly 13.8x cheaper inference (well under 10% of the cost), per Thinking Machines’ and Bridgewater’s own figures.
The Bridgewater case establishes the builder template:
- Start with Inkling as the base (Apache 2.0, no royalties, weights you can inspect)
- Fine-tune on your domain data via Tinker
- The fine-tuned checkpoint becomes proprietary to you — Thinking Machines does not retain rights to derivative models under Apache 2.0
What builders should evaluate
Why Inkling over a proprietary API:
- Fine-tuned weights are yours — no vendor lock-in, no per-token royalty on the derivative model
- Apache 2.0 is the most permissive major open license for commercial use
- 1M context window is competitive with any frontier model at launch
- Native multimodal (text + image + audio) without adapter layers
Why Inkling might not be the right choice:
- Raw benchmark performance is below Fable 5 and GPT-5.6 Sol for general tasks — Gemini 3.5 Pro remains unreleased as of this writing, so the current Google flagship for comparison is Gemini 3.1 Pro
- 600 GB VRAM minimum for self-hosting puts it out of reach for most teams without a cluster partnership
- Tinker fine-tuning is Thinking Machines’ proprietary platform — you’d be trading one vendor lock-in for another unless you self-host the fine-tuning process on the open weights
When to consider Inkling:
- You have a narrow, well-defined domain (financial analysis, legal review, scientific literature) where a fine-tuned 41B-active model can outperform a general frontier model at lower cost
- Your legal or compliance team requires full weight transparency and the ability to audit the model on-premises
- You want to build a product on customized weights you own, not on an API you rent
Company context
Thinking Machines Lab is valued at $12 billion, with revenue currently driven by Tinker. The existing Thinking Machines article on chatforest.com covers the company’s founding philosophy — the core argument that centralized alignment concentrates power and that builders should own weights rather than rent API access. Inkling is the first product that operationalizes that argument.
The short version
Inkling is a 975B MoE, 41B active, Apache 2.0, multimodal, 1M context, weights-on-Hugging-Face. It is not the strongest model available. It is designed to be fine-tuned into something narrow, purpose-built, and proprietary to you. Access it via Tinker at $1.87/$4.68 per MTok (50% launch discount), or via TogetherAI, Fireworks, Modal, Databricks, or Baseten. Self-hosting requires 600 GB VRAM minimum (NVFP4 checkpoint) or 2 TB (BF16). The Bridgewater financial fine-tune hit 84.7% on domain benchmarks at under 10% of proprietary model cost. If you need a domain-specific model you own outright, this is the most serious US open-weight option available today.