Inkling: How Thinking Machines Lab Built a 975B Open-Weight Model Around Controllable Thinking

Thinking Machines Lab — the startup founded by former OpenAI CTO Mira Murati — released its first in-house model on July 15, 2026. Called Inkling, it is a 975-billion-parameter Mixture-of-Experts (MoE) transformer with 41 billion active parameters, a 1-million-token context window, and native support for text, image, and audio inputs. The weights are available on Hugging Face under the Apache 2.0 license.

What makes Inkling worth examining is not its raw benchmark position — the company is explicit that it is "not the strongest overall model available today, open or closed." Instead, the design choices around efficiency, multimodality, and a feature called controllable thinking effort reflect a specific set of tradeoffs that are worth understanding on their own terms.

Architecture: A Familiar Spine with Deliberate Departures

Inkling is a 66-layer decoder-only transformer. Its MoE feed-forward layers follow the DeepSeek-V3 design closely: each layer contains 256 routed experts and 2 shared experts, with 6 routed experts activated per token. This keeps the active parameter count at 41B while the full parameter count sits at 975B — a ratio that makes inference substantially cheaper than a dense model of equivalent capacity.

A few choices diverge from the common recipe:

• Attention: Inkling uses an interleaved mix of sliding-window and global attention layers at a 5:1 ratio, with 8 KV heads. Rather than RoPE (the positional encoding used by most recent models), it uses learned relative positional embeddings.
• Multimodal encoding: The model uses an encoder-free early fusion approach. Audio is ingested as dMel spectrograms; images are encoded as 40×40 pixel patches via a four-layer hierarchical MLP (hMLP). Both modalities are projected into the same hidden space as text tokens and processed jointly by the decoder.
• Optimization: Training used a hybrid strategy — Muon for large matrix weights, Adam for everything else. The model was trained entirely on NVIDIA GB300 NVL72 systems.

The encoder-free design for vision and audio is notable. Most multimodal models use a separate vision encoder (like a ViT) and project its outputs into the language model's space. Inkling skips that step, processing raw patches directly. This simplifies the architecture and avoids the mismatch between a separately trained encoder and the main model, but it also means the model has to learn visual representations from scratch rather than inheriting them from a pretrained encoder.

Controllable Thinking Effort: What It Is and How It Works

The most distinctive feature of Inkling is its controllable thinking effort — a parameter that lets developers set the model's reasoning budget on a scale from 0.2 to 0.99 at inference time.

This is not a simple temperature knob. It was trained into the model through large-scale asynchronous reinforcement learning over more than 30 million rollouts. During RL training, Thinking Machines varied the system message and applied a per-token cost to the model's chain-of-thought reasoning. By adjusting that cost across different training samples, the model learned to compress or expand its reasoning depending on the effort level specified.

The practical result: at lower effort settings, the model produces shorter reasoning chains and responds faster. At higher settings, it reasons more extensively before generating a final answer. According to Thinking Machines' benchmarking, Inkling at effort=0.99 uses roughly one-third as many tokens as Nemotron 3 Ultra to reach comparable coding performance on certain tasks.

This matters for real-world deployment. Most production applications do not need maximum reasoning on every query. A customer support system, a document classifier, or a retrieval-augmented generation pipeline can often get adequate results with a fraction of the token budget that a hard reasoning task requires. Having a single model that can operate across that range — rather than maintaining separate fast and slow models — simplifies infrastructure.

Multimodal Benchmarks: Where It Stands

Inkling's multimodal performance is competitive among open-weight models, though closed models still lead on most audio and vision tasks. On MMMU Pro (Standard 10), it scores 73.5% — above Qwen3-Omni (60.0%) but below Gemini 3.1 Pro (82.0%). On VoiceBench it reaches 91.4%, and on MMAU (audio understanding) 77.2%, roughly on par with Qwen3-Omni.

On agentic coding, Inkling scores 77.6% on SWE-Bench Verified and 63.8% on Terminal Bench 2.1. On instruction following (IFBench), it scores 79.8% — above Claude Fable 5 (63.5%) and GPT-5.6 Sol (72.7%), which is one of its stronger relative results.

Inkling-Small: A Preview of the Smaller Sibling

Alongside the main release, Thinking Machines previewed Inkling-Small: a 276B-parameter MoE with 12B active parameters. The performance numbers are striking. On HLE (with tools), Inkling-Small scores 46.6% versus Inkling's 46.0%. On GPQA Diamond, it scores 88.3% versus 87.2%. On MCP-Atlas, it scores 74.9% versus 74.1%.

The smaller model underperforms on factuality (SimpleQA Verified: 20.9% vs. 43.9%) and on Terminal Bench 2.1 (52.7% vs. 63.8%), but for many agentic and reasoning tasks, the gap is narrow enough that the lower cost and latency of Inkling-Small may be the better tradeoff.

The Customization Argument

Thinking Machines is positioning Inkling primarily as a base for fine-tuning rather than a finished product. The company's platform, Tinker, provides fine-tuning infrastructure and a developer playground. The Apache 2.0 license allows commercial use and modification without restriction.

The business logic differs from OpenAI or Anthropic, which sell metered API access to locked models. As TechCrunch noted, once the weights are public, nothing obligates anyone to pay Thinking Machines to run them — so Tinker, not Inkling itself, is where the company's revenue has to come from.

One example: Thinking Machines worked with Bridgewater Associates to fine-tune a model on the hedge fund's internal financial expertise. The result reportedly scored 84.7% on financial reasoning tests at roughly one-fourteenth the running cost of comparable proprietary models — though those numbers come from the two companies' own evaluation, not an independent one.

Hardware Requirements and Deployment

Running the BF16 checkpoint requires at least 2 TB of aggregated VRAM — either 8× NVIDIA B300 GPUs or 16× H200 GPUs. A quantized NVFP4 checkpoint reduces this to 600 GB (4× B300 in W4A4 mode, or 8× H200 in W4A16 mode). Supported inference frameworks include SGLang, vLLM, TokenSpeed, Unsloth, and Hugging Face Transformers.

For most organizations, self-hosting a model at this scale is not practical. The more realistic path is using Tinker's hosted fine-tuning and inference, or one of the third-party inference providers that Thinking Machines has partnered with.

What to Make of It

Inkling is a technically coherent model with a clear design philosophy: broad multimodal capability, efficient inference through sparse MoE, and a training-time mechanism for controlling reasoning cost at inference. It is not trying to top every benchmark — it is trying to be a good starting point for organizations that want to adapt a model to their own data and workflows.

The controllable thinking effort mechanism is the most interesting technical contribution. Training a model to genuinely compress or expand its reasoning chain — rather than just truncating output — requires careful RL design, and the token efficiency numbers suggest it works. Whether that holds up after fine-tuning on domain-specific data remains to be seen.

Weights, model card, and fine-tuning documentation are at thinkingmachines.ai.