Large Language Models

Inkling

Mira Murati's first model — a 975B/41B-active open-weight MoE with Apache 2.0 license, native text/image/audio, 1M context, designed as a base for fine-tuning.

Inkling is the first model from Thinking Machines Lab — the company founded by Mira Murati (former OpenAI CTO) — released on July 15, 2026 with full weights under Apache 2.0. It’s a 975B-parameter Mixture-of-Experts transformer with 41B active parameters, native text/image/audio reasoning, and a 1M-token context window. And Thinking Machines is refreshingly honest about what it is: “not the strongest overall model available today, open or closed.”

The pitch is not raw capability — it’s the base. Inkling is positioned as a balanced, multimodal foundation model for downstream fine-tuning on the company’s Tinker platform. Apache 2.0 licensing means no revenue gates, no branding mandates, no separate commercial agreements. For teams that need an open, permissive starting point for customization, that combination is the selling point.

Architecture

SpecValue
Total parameters975B
Active parameters41B per token
ArchitectureDecoder-only MoE transformer (6 of 256 routed + 2 shared experts, 66 layers)
Context window1M tokens (weights) / 256K (Tinker API)
Input modalitiesText, image, audio
Output modalitiesText
Training data45T tokens (text, images, audio, video)
NumericsBF16 and NVFP4 (Blackwell-optimized)
LicenseApache 2.0

Pricing (Tinker platform)

Context lengthInput / 1MCached / 1MOutput / 1M
64K$1.87$0.374$4.68
256K$3.74$0.748$9.36

A 50% launch discount is available for a limited time. Inkling is also available via third-party inference providers.

Benchmarks

Thinking Machines published a comparison table benchmarking Inkling against both open and closed frontier models:

BenchmarkInklingKimi K2.6GLM-5.2DeepSeek V4-ProFable 5 (max)GPT-5.6 Sol
HLE (text only)29.7%35.9%40.1%35.9%53.3%47.2%
HLE (with tools)46.0%54.0%54.7%48.2%64.5%55.0%
AIME 202697.1%96.4%99.2%96.7%99.9%99.9%
GPQA87.2%91.1%89.5%88.8%92.6%94.1%

The pattern is clear: Inkling competes with the top open-weight models (Kimi K2.5/K2.6, GLM-5.2, DeepSeek V4-Pro) but sits meaningfully behind the closed frontier (Fable 5, GPT-5.6 Sol). On AIME 2026 (math) it scores 97.1% — strong, but not top. The AA Intelligence Index places Inkling at 41, well below frontier (Opus 5 at 61).

Why Apache 2.0 matters

Simon Willison’s characterization captures it well: “good to see the US open weights ecosystem gain a new viable contender to join NVIDIA Nemotron and Gemma 4.” Inkling is the first large-scale US-origin model at this capability level shipped under Apache 2.0 — making it directly competitive on licensing terms with the Chinese open-weight leaders.

ModelLicenseTotal params
InklingApache 2.0975B
Kimi K3Kimi K3 License (custom)2.8T
GLM-5.2MIT753B
Kimi K2.7 CodeModified MIT1T
DeepSeek V4-ProMIT

Inkling-Small

Alongside the main model, Thinking Machines previewed Inkling-Small: a 276B-parameter (12B active) variant trained with a similar recipe. It matches or exceeds the larger Inkling on many benchmarks — a result of improvements to the smaller model’s pre-training data and recipe. Full weights will be released after testing concludes.

Availability

SurfaceStatus
Hugging FaceFull weights (BF16 + NVFP4)
Tinker platformAPI access (64K and 256K context)
Third-party inferenceAvailable

Best for

  • Fine-tuning and distillation from a clean, permissive base model
  • Multimodal applications requiring text + image + audio reasoning
  • Building custom models on the Tinker fine-tuning platform
  • Research and experimentation with a frontier-class open MoE architecture

Pros & cons

Strengths

  • Apache 2.0 licensed — the most permissive frontier-class open license (no revenue gates, no branding mandates)
  • Natively multimodal (text, image, audio) with efficient and controllable thinking effort
  • First viable US-origin open-weight competitor to Chinese MoE models (Kimi, GLM, DeepSeek)
  • Full weights available on Hugging Face (BF16 + NVFP4 for Blackwell inference)

Limitations

  • Not the strongest model overall by Thinking Machines' own admission — positioned as a base for customization
  • AA Intelligence Index at 41 — well below frontier models (Opus 5 at 61, Fable 5 higher)
  • Tinker platform limited to 256K context (weights support 1M)
  • Inkling-Small (276B/12B) still in preview, weights not yet released

Frequently asked questions

When was Inkling released?

Thinking Machines Lab (founded by Mira Murati, former OpenAI CTO) released Inkling on July 15, 2026. Full weights went live on Hugging Face the same day, and API access launched on the Tinker platform.

Is Inkling the best open model?

No, and Thinking Machines says so explicitly. Per their launch post, "Inkling is not the strongest overall model available today, open or closed." Its value is as a balanced, multimodal, permissively licensed base for customization. On HLE-text it scores 29.7% (vs Kimi K2.6's 35.9% or GLM-5.2's 40.1%).

What is Inkling-Small?

Inkling-Small is a 276B-parameter (12B active) variant trained with a similar recipe. It matches or exceeds the larger Inkling on many benchmarks. Weights will be released after testing concludes — it is currently in preview only.

How is Inkling licensed?

Apache 2.0 — the most permissive option among frontier-class models. No revenue gates, no branding mandates, no separate commercial agreements needed. Compare to Kimi K3's custom license or Meta's Llama community license.

What hardware runs Inkling?

Hugging Face provides weights in BF16 and NVFP4 (optimized for NVIDIA Blackwell). At 975B total parameters, self-hosting requires significant GPU infrastructure. Alternatively, use the Tinker API or third-party inference providers.

Sources

Last updated: 2026-07-29 · Specs and pricing change fast — verify on the vendor's site before relying on them.