What Is Inkling? The 975B Open-Weights MoE From Mira Murati's Thinking Machines

Inkling is an open-weights Mixture-of-Experts foundation model released by Thinking Machines Lab on July 15, 2026, with 975B total parameters, 41B active per token, a 1M-token context window, and native reasoning across text, images, and audio. It is the lab's first production model, shipped under the Apache 2.0 license with full weights on Hugging Face and same-day fine-tuning available on the company's Tinker platform. Thinking Machines explicitly does not position Inkling as the strongest overall model — its stated role is a balanced, multimodal, customizable base for downstream fine-tuning.

Key Takeaways

• Inkling is a 975B-parameter Mixture-of-Experts transformer with 41B active parameters per token, released on July 15, 2026 by Thinking Machines Lab.
• Weights are Apache 2.0 licensed and available on Hugging Face in both BF16 and NVFP4 quantizations, with day-zero support in SGLang, vLLM, llama.cpp, and Unsloth.
• The model natively ingests text, image, and audio, and supports up to 1M-token context (256K on the Tinker API).
• It debuts at 41 on the Artificial Analysis Intelligence Index, the top score among U.S. open-weights models, ahead of Nemotron 3 Ultra at 38.
• A controllable Thinking Effort dial (0.00–0.99) trades compute for accuracy at inference time.
• A smaller preview variant, Inkling-Small, has 276B total and 12B active parameters; full weights are not yet released.

What Is Inkling?

Inkling is the first production language model from Thinking Machines Lab, the AI research company founded by former OpenAI CTO Mira Murati. Announced on July 15, 2026, roughly eighteen months after the lab was founded, it is a general-purpose multimodal MoE trained from scratch on 45 trillion tokens spanning text, images, audio, and video.

The model is deliberately released as a base for customization, not a leaderboard-topping frontier system. In the launch blog, Thinking Machines writes that "Inkling is not the strongest overall model available today, open or closed," and frames the release around three qualities: multimodal capabilities, efficient thinking, and availability on Tinker for fine-tuning. Mira Murati's launch post captured the pitch in one line: "Our first model, Inkling. Trained from scratch, weights are open, fine-tunable on Tinker today."

Two model variants exist. The main Inkling model (975B/41B) is fully released. A smaller preview called Inkling-Small (276B total, 12B active) uses a similar training recipe; per the official model card it matches or slightly beats the larger model on several benchmarks including HLE-with-tools (46.6% vs 46.0%) and GPQA Diamond (88.3% vs 87.2%). Full weights for Inkling-Small are not yet published.

Inkling at a Glance

Field	Value
Developer	Thinking Machines Lab
Release date	July 15, 2026
Type	Decoder-only Mixture-of-Experts transformer
Total parameters	975B
Active parameters	41B per token
Experts	6 of 256 routed + 2 shared
Layers	66
Input modalities	Text, image, audio
Output modalities	Text
Context window	1M tokens (weights) / 256K tokens (Tinker API)
Training data	45T tokens (text, images, audio, video)
Numerics	BF16 and NVFP4
License	Apache 2.0
Weights	Hugging Face (thinkingmachines/Inkling)
Tinker pricing (64K ctx)	$1.87 / 1M input · $0.374 / 1M cached · $4.68 / 1M output
Tinker pricing (256K ctx)	$3.74 / 1M input · $0.748 / 1M cached · $9.36 / 1M output
AA Intelligence Index	41
Small variant	Inkling-Small — 276B total / 12B active (preview)

How Inkling Works

Inkling's architecture makes several unusual choices for a large open-weights release. Per the Hugging Face model card and the Hugging Face welcome blog, the model is a 66-layer decoder-only transformer with the following distinguishing features:

Sparse MoE routing with shared experts. Each token is routed to 6 of 256 routed experts, plus 2 shared experts that fire on every token. Only 41B parameters activate per forward pass despite the 975B total.

Relative attention instead of RoPE. Rather than rotary position embeddings, Inkling uses a learned relative-position projection injected directly into attention logits, with a fourth projection tensor (beyond Q/K/V) producing per-token, per-head relative features.

Hybrid attention pattern. Layers alternate between global attention and sliding-window attention in a 5:1 ratio, with the final layer using global attention. This is how Inkling supports a 1M-token context window while keeping inference tractable.

Short convolution (SConv). A 1D convolution over hidden states reads the current token plus the previous W-1 states, freeing the attention and MoE modules from having to encode local patterns.

Encoder-free multimodality. Images use a hierarchical MLP patchifier that progressively merges pixels into one embedding per patch. Audio is encoded as discretized mel spectrograms in 100 ms chunks. All modalities project into a shared hidden space and are processed jointly by the decoder — there is no separate vision or audio tower.

Thinking Effort dial. A number from 0.00 to 0.99 in the system prompt scales how much compute the model spends before answering. Reported benchmarks are at effort=0.99. According to Unsloth's documentation, this replaces the usual "reasoning: high/medium/low" convention with a continuous curve.

MTP layers for speculative decoding. The release includes speculative multi-token-prediction layers for faster inference.

What You Can Do With Inkling

Thinking Machines positions Inkling as a base model rather than an application. Real-world uses surfacing in the first day of the release cluster around a few patterns.

Domain fine-tuning on Tinker. The launch demo showed Inkling writing its own fine-tuning job, running it on Tinker, and evaluating the resulting model — a self-fine-tune loop captured in a post from Rohan Paul.

Source: @rohanpaul_ai

Multimodal agents. Native audio and vision inputs make Inkling a candidate for voice assistants, document analysis, and video-adjacent workflows without bolting on separate encoders. Max Weinbach called it "the seemingly best base for a full scale post train/RL run for specific workloads."

Enterprise coding workflows. Databricks announced day-zero availability through its Unity AI Gateway, targeting enterprises that want to fine-tune Inkling on proprietary codebases and connect it to coding agents like Cursor and OpenCode.

Local inference via quantization. Unsloth shipped Dynamic 1-bit GGUFs the same day. Per Unsloth's post, the 1-bit quant is 86% smaller than BF16 and retains 74.2% of top-1 accuracy, runnable on ~280 GB of combined RAM and VRAM.

Source: @UnslothAI

How Inkling Compares

Inkling's benchmarks are competitive within the open-weights bracket but trail closed-frontier models on text reasoning. Selected scores from the official model card, all at effort=0.99:

Benchmark	Inkling	GLM 5.2	Nemotron 3 Ultra	Claude Fable 5 (max)
HLE (text only)	29.7%	40.1%	26.6%	53.3%
HLE (with tools)	46.0%	54.7%	37.4%	64.5%
AIME 2026	97.1%	—	—	—
GPQA Diamond	87.2%	—	—	—
SWE-bench Verified	77.6%	—	—	—
SWE-bench Pro	54.3%	—	—	—
Terminal-Bench 2.1	63.8%	—	—	—
MMMU Pro	73.5%	—	—	—
VoiceBench	91.4%	—	—	—

The Artificial Analysis writeup ranks Inkling at 41 on its v4.1 Intelligence Index, three points above Nemotron 3 Ultra and ahead of Gemma 4 31B (29) and gpt-oss-120b (24). It also flags token efficiency: Inkling averages 25K output tokens per Intelligence Index task, versus 37–43K for GLM 5.2, Kimi K2.6, and DeepSeek v4 Pro.

Community reaction reflects the framing tension. Bindu Reddy welcomed it as "a bit worse than GLM 5.2" but the "first US open-source model worth benchmarking." Teortaxes offered a contrarian read: "I like that Inkling is mediocre on benchmarks — this suggests they haven't been cutting corners too much with distillation." For comparisons against closed peers, see our earlier GPT-5.6 deep dive and Muse Spark 1.1 analysis.

Availability: How to Access Inkling

Four access paths exist on day one:

1. Weights on Hugging Face — thinkingmachines/Inkling under Apache 2.0, in BF16 and NVFP4. Model card notes the 1M-token context ceiling.
2. Tinker API — Thinking Machines' hosted fine-tuning and inference platform, at $1.87 / $4.68 per 1M input/output tokens for the 64K tier, with a launch discount reported in early coverage. The Tinker API context ceiling is 256K.
3. Inkling Playground — a chat interface inside the Tinker console for developers to try the model before committing to fine-tuning, per TestingCatalog.
4. Third-party inference and enterprise platforms — day-zero support in SGLang, vLLM, llama.cpp, and Unsloth, with hosted access on Databricks' Unity AI Gateway and other partners.

For local deployment, Unsloth documents six memory tiers, from 280 GB (1-bit) to 1,900 GB (BF16) of combined RAM and VRAM. A Mac Studio Ultra or a multi-GPU rig with roughly 290 GB of unified memory can host the smallest quant.

What We Don't Know Yet

Several details remain unpublished or unverified as of launch day:

• Inkling-Small release date. The 276B/12B preview is described but weights are not yet on Hugging Face.
• Training data provenance. The model card says data was "drawn from publicly available sources, acquired from third-parties, or synthetically generated or augmented," with no dataset composition breakdown. One community post from op7418 claims post-training data was distilled from open models including Kimi 2.5 — this is unconfirmed by Thinking Machines.
• Training compute and cost. Not disclosed.
• Independent benchmark replication. Most reported scores come from the official model card or Artificial Analysis; broad third-party evaluation is still in progress.
• Post-fine-tune safety behavior. How much fine-tuning erodes the base model's safety training is an open question, especially given the Apache 2.0 license.
• Long-form multilingual and video capabilities. The model card highlights English; video was in the pretraining mix but video-input support at inference is not documented in the released tooling.

Frequently Asked Questions

What is Inkling?

Inkling is an open-weights Mixture-of-Experts foundation model released by Thinking Machines Lab on July 15, 2026. It has 975B total parameters with 41B active per token, a 1M-token context window, and natively reasons over text, images, and audio.

Who made Inkling?

Inkling was built by Thinking Machines Lab, the AI research company founded by former OpenAI CTO Mira Murati. It is the lab's first production model release, roughly a year and a half after the company was founded.

Is Inkling open source?

Inkling's weights are released under the Apache 2.0 license, which permits commercial use, fine-tuning, and redistribution. Thinking Machines calls it "open-weights" rather than fully open source because the training data and training pipeline are not released.

How much does Inkling cost?

On Thinking Machines' Tinker platform, Inkling is priced at $1.87 per 1M input tokens and $4.68 per 1M output tokens at a 64K context window, with cached input at $0.374 per 1M tokens. The 256K context tier costs roughly double, and self-hosting the weights is free apart from compute.

What can Inkling do?

Inkling accepts text, image, and audio inputs and produces text outputs, with a controllable Thinking Effort dial from 0.00 to 0.99. It is designed as a customizable base for coding agents, tool-use systems, multimodal reasoning apps, and domain fine-tuning through the Tinker platform.

Inkling vs GLM 5.2 — which is better?

GLM 5.2 scores higher on several text reasoning benchmarks such as HLE, where GLM 5.2 reaches 40.1% versus Inkling's 29.7%. Inkling's stated advantages are native multimodal input, token efficiency on agentic tasks, and being the leading US open-weights model on the Artificial Analysis Intelligence Index at a score of 41.

How do I run Inkling locally?

Inkling weights are on Hugging Face under thinkingmachines/Inkling, with day-zero support for SGLang, vLLM, llama.cpp, and Unsloth. Full BF16 requires roughly 1.9 TB of storage, while Unsloth's Dynamic 1-bit GGUF runs on about 280 GB of combined RAM plus VRAM and retains around 74% of top-1 accuracy.

What to watch next

Three signals will determine whether Inkling becomes a durable base model or a first-attempt curiosity. First, the full release of Inkling-Small (276B/12B) — its smaller footprint could drive more fine-tuning experiments than the 975B flagship. Second, independent evals from third parties beyond Artificial Analysis, especially on long-horizon agentic coding where early hands-on tests have been mixed. Third, the first wave of published fine-tunes on Tinker: Thinking Machines' entire pitch rests on customization being where the real value lives, so the quality of downstream models is the real benchmark.

Building similar multimodal or agentic systems? On kie.ai you can try Claude Fable 5, GPT-5.6, and Gemini 3.1 Pro.