Foreword¶

On July 15, 2026, Thinking Machines Lab, founded by former OpenAI CTO Mira Murati, released its first open-weight model trained from scratch, Inkling. This is the first official model product from the new lab established after Murati briefly served as interim CEO of OpenAI in 2023 and left the company in 2024.

Unlike the approach of chasing top leaderboard rankings taken by models like DeepSeek and Kimi, Thinking Machines stated bluntly in its official blog that Inkling is not the most powerful model currently available, whether closed-source or open-source. The team positions it as a customizable enterprise-grade base model: a MoE architecture with 975B total parameters and 41B activated parameters, released under the Apache 2.0 license, natively supporting text, image, and audio inputs, and paired with the Tinker fine-tuning platform. For enterprise developers needing private deployment and domain customization, this “open-source MoE + controllable thinking effort” approach deserves serious evaluation.

What is Inkling¶

Inkling is the first Mixture-of-Experts (MoE) multimodal large model trained by Thinking Machines Lab. Its core specifications are as follows:

Murati mentioned on X that Inkling is a model trained from scratch, and its weights have been publicly released on Hugging Face. A preview version Inkling-Small (276B total parameters, 12B activated parameters) was also released simultaneously, with a volume about one-quarter of the full-size model and performance close to the large model on multiple benchmarks.

MoE Architecture and Controllable Thinking Effort¶

Sparse MoE Design¶

Inkling’s MoE design generally follows the DeepSeek-V3 roadmap, but with several modifications for long context and efficiency:
- Each MoE layer contains 256 routed experts + 2 shared experts, with each token activating 6 routed experts and all shared experts
- The Router uses sigmoid + auxiliary-loss-free load balancing to avoid training interference caused by traditional auxiliary losses
- The Attention layer alternates between sliding-window and global attention at a 5:1 ratio, with 8 KV heads
- Relative positional embeddings (not RoPE) are used, which the official claims perform better on long sequence extrapolation

For multimodality, images are encoded via a hierarchical patch encoder (40×40 pixel patches), and audio is discretized via dMel spectrogram before being projected into the same hidden space as text tokens, jointly processed by the decoder. This is an encoder-free multimodal solution, consistent with the design philosophy of the interaction models previously released by the team.

Controllable Thinking Effort¶

A key differentiating capability of Inkling is controllable thinking effort — developers can adjust the model’s “thinking depth” during inference to balance performance and token costs.

Official benchmark data shows: On Terminal Bench 2.1, Inkling can achieve scores equivalent to Nemotron 3 Ultra with about one-third of the tokens. When the effort parameter is swept from 0.2 to 0.99, performance and the number of generated tokens follow a tunable curve. For agent scenarios requiring large-scale calls or integration into long workflows, this means costs can be dynamically controlled based on task difficulty, rather than maxing out the inference budget uniformly.

Capability Overview¶

Thinking Machines trained Inkling as a breadth-first generalist model, covering areas such as agentic coding, reasoning, instruction following, factuality, vision, audio, and safety, rather than chasing leaderboard rankings on single tasks. Several official reference data points are provided (effort=0.99):
- SWE-bench Verified: 77.6%
- Terminal Bench 2.1: 63.8%
- AIME 2026: 97.1%
- VoiceBench (audio): 91.4%
- FORTRESS Adversarial (safety): 78.0%, the highest among the open-source models compared

On the Design Arena’s Agentic Web Dev blind review leaderboard, Inkling scored 1257, ranking among the top open-source models. The team emphasized that when selecting a base model for fine-tuning, benchmark numbers are only part of the equation, and “feel” is equally important — to this end, an Inkling Playground has been added to the Tinker console for developers to directly chat and experience the model.

Tinker Fine-tuning Platform¶

Inkling’s commercial ecosystem is deeply bound to Tinker. Tinker is a training and inference platform for developers from Thinking Machines, and fine-tuning capabilities went live on the same day Inkling was released.

Basic Usage¶

1. Select Inkling on the Tinker console, with context window options of 64K or 256K
2. Refer to the official Tinker Cookbook to write fine-tuning tasks; multimodal post-training requires installing the tml-renderers package
3. Sampling and validation can be performed at any time from sampler checkpoints during training, without waiting for the entire task to complete
4. Fine-tuned checkpoints can be deployed via the Tinker API or partner inference providers

Tinker also provides an Anthropic-compatible API: Point ANTHROPIC_BASE_URL to the Tinker endpoint and replace the API Key, and existing tools such as Claude Code and the Anthropic SDK can directly connect to Inkling or self-trained checkpoints, reducing migration costs.

Self-fine-tuning Demo¶

There is a symbolic demo in the official blog: Inkling writes its own fine-tuning tasks, runs training, and evaluates results on Tinker — customizing the model with the model itself, demonstrating the potential of Tinker in agentic workflows.

Private Deployment and Inference¶

Inkling’s weights can be downloaded on Hugging Face, with two checkpoint formats available: BF16 and NVFP4.

Hardware Requirements¶

BF16 full precision (minimum 2 TB aggregated VRAM):
- 8× NVIDIA B300, or
- 16× NVIDIA H200

NVFP4 quantized version (minimum 600 GB aggregated VRAM):
- W4A4 mode: 4× B300 (requires SM100+ architecture)
- W4A16 mode: 8× H200

For enterprise private deployment, the threshold is not low, but the NVFP4 version has reduced hardware requirements to a manageable cluster scale.

Supported Inference Frameworks¶

Day-0 ecosystem support is relatively complete, with options including:
- SGLang (cooperated with RadixArk, with RL support)
- vLLM (Inferact)
- TokenSpeed (Lightseek)
- Unsloth (including llama.cpp path)
- Hugging Face transformers

If you do not want to build your own cluster, you can call the model via third-party APIs: Together AI, Fireworks, Modal, Databricks, and Baseten have all integrated Inkling.

Position in the Open-source MoE Track¶

The competition in the 2026 open-source large model market is fierce: Kimi K3, DeepSeek-V4-Flash, Tencent Hy3, Poolside Laguna, and others have been released one after another. Interconnects.ai pointed out in its open artifacts review that when Thinking Machines was founded in February 2025, few outsiders classified it as an “open-source model company”; today, Tinker’s fine-tuning service has reached an annual revenue of hundreds of millions of dollars, and Inkling is also regarded as one of the strongest open-weight models trained in the United States.

Inkling’s differentiation does not lie in single-point SOTA, but in a three-point combination:
1. Full-weight open-source under Apache 2.0 with no commercial restrictions
2. Native multimodality + controllable effort, suitable for agents and enterprise customization
3. Tinker one-stop fine-tuning + inference + Anthropic-compatible API, reducing the friction from trial to production

For domestic developers, attention should be paid to the fact that the model is primarily English-focused, with multilingual as a supplement; the computing power cost of long context and multimodal inference; and the safety compliance and content review responsibilities of open-weight models — the official recommends adding moderation tools such as Llama Guard on the deployment layer, rather than relying solely on the model’s built-in refusal responses.

Summary¶

The first model submission led by Mira Murati and Thinking Machines Lab has a clear strategy: Do not be the strongest, be the best base model for fine-tuning. The 975B-A41B MoE, 1M context window, Apache 2.0 license, and Tinker ecosystem — Inkling targets the enterprise AI implementation path of “obtain weights, customize by domain, deploy privately”.

If you are evaluating the next fine-tune base model, you might as well first try it in the Tinker Playground or the Hugging Face weight page, and scan the effort curve combined with your own task’s latency and cost constraints before deciding whether to invest in fine-tuning — this is closer to real selection than just looking at a single benchmark score.

References:
- Inkling: Our Open-Weights Model (Official Blog)
- Inkling Model Card
- The Verge Report
- Interconnects.ai Open Artifacts #23