All models
Text & reasoningStable
Thinking Machines LabInkling
Thinking Machines Lab's large open-weights model for customizable reasoning, coding, tools, vision, and audio workflows.
Plain-English overview
What Inkling actually is
Inkling is a 975-billion-parameter mixture-of-experts model, but it activates 41 billion parameters for each token. It can reason over text, images, and audio, use tools, and vary how much thinking effort it spends on a task.
Its main appeal is control. The Apache-2.0 weights can run on your infrastructure, while Tinker provides a playground and fine-tuning workflow. Hosted partners make it easier to start, but each route has different pricing, privacy, hardware, and context limits.
Good fit for
- Teams that want an adaptable open-weight foundation model
- Agentic coding and tool-use experiments
- Private or specialized deployments with multimodal inputs
Category comparison
The facts that matter for text models
These are provider-published specifications, not Cody benchmark scores. Follow the linked sources for current limits and endpoint-specific exceptions.
- Context window
- Up to 1M tokens; Tinker offers 64K and 256KMaximum combined prompt and working context documented by the provider.
- Maximum output
- Not separately publishedProvider-published response limit, where available.
- Knowledge cutoff
- Not publishedLatest reliable knowledge date explicitly published by the model provider. Search and connected tools can retrieve newer information but do not change the model's built-in cutoff.
- Input price
- Hosting dependentCurrent standard list price per million input tokens unless noted.
- Output price
- Hosting dependentCurrent standard list price per million output tokens unless noted.
- Inputs
- Text, image, audioMedia types accepted by the listed model endpoint.
- Tools & agents
- Agentic coding, tools, Python-assisted vision, controllable thinkingSelected native tools and agent-building capabilities, not an exhaustive list.
Pricing & comparisons
Estimate your cost
Set your usage. Your estimate updates as you type.
Example: 6,000 input tokens for 10 pages, plus a 500-token summary. Page lengths vary; adjust the numbers below.
One run sends your input to the model once and receives an answer. Tokens are pieces of text: input is what you send, output is the answer you receive.
Advanced options
Inkling
Thinking Machines Lab
Estimated total (USD)
For the usage above · USD · API pricing, not a subscription
How this estimate works
Estimates exclude taxes, tools, cache storage/writes, free allowances and custom discounts. Image estimates cover output only, not prompt or reference-image charges. Quality modes differ by model. Unlisted settings are not treated as free.
API and provider access
Where to get Inkling
Thinking Machines LabInkling release and availabilityOfficial architecture, capabilities, Tinker, and partner-access overview.Hugging FaceInkling model card and weightsApache-2.0 weights, model properties, and self-hosting examples.TinkerInkling PlaygroundPlayground and fine-tuning access; hosted terms and pricing apply.
Availability
Regions and access stage
Open weights on Hugging Face, fine-tuning and a playground through Tinker, plus hosted partner and self-hosted inference routes.
Self-hosting determines the processing region. Tinker and partner availability follow each provider's account and regional terms.
Check live availabilityData and training
The route matters.
The open weights can be self-hosted so request data stays in infrastructure you control. Tinker and partner-hosted routes have separate retention and training-use terms that must be checked with the selected provider.
This is a concise reading of the cited provider material, not legal advice. A third-party gateway can have different storage, routing, training, and residency terms from the model maker's direct API.
Read the provider policyFrequently asked questions
Inkling FAQ
What is Inkling?
Thinking Machines Lab's large open-weights model for customizable reasoning, coding, tools, vision, and audio workflows. Inkling is a 975-billion-parameter mixture-of-experts model, but it activates 41 billion parameters for each token. It can reason over text, images, and audio, use tools, and vary how much thinking effort it spends on a task.
When was Inkling released?
Inkling was released on July 15, 2026 according to the cited provider materials.
Where can I access Inkling?
Open weights on Hugging Face, fine-tuning and a playground through Tinker, plus hosted partner and self-hosted inference routes. The access routes listed in this guide are Thinking Machines Lab, Hugging Face, and Tinker.
How much does Inkling cost?
Open weights; hosted pricing varies. The model is Apache 2.0. Tinker, partner inference, and self-hosted deployments each have their own compute and usage costs.
Where is Inkling available?
Open weights on Hugging Face, fine-tuning and a playground through Tinker, plus hosted partner and self-hosted inference routes. Self-hosting determines the processing region. Tinker and partner availability follow each provider's account and regional terms.
Is my Inkling API data used for training?
The open weights can be self-hosted so request data stays in infrastructure you control. Tinker and partner-hosted routes have separate retention and training-use terms that must be checked with the selected provider. The policy belongs to the provider route and account terms, so verify it again before production use.
Price comparison
USD per million tokens for the shown API routes at standard context length. Cache and long-context rates may differ. Models without matching reviewed prices are omitted; lower cost does not mean better quality.
- Mistral Small 4$0.15
Mistral AI
- Gemini 3.8 Flash$0.75
Google
- GPT-5.6 Sol$4.00
OpenAI
- GPT-6 Astra$10.00
OpenAI
- Claude Fable 5.1$10.00
Anthropic
- Mistral Small 4$0.60
Mistral AI
- Gemini 3.8 Flash$3.75
Google
- GPT-5.6 Sol$20.00
OpenAI
- GPT-6 Astra$50.00
OpenAI
- Claude Fable 5.1$50.00
Anthropic
Related comparisons
Head-to-head comparisons
Inkling vs GPT-6 Astra
OpenAI's most capable model for difficult end-to-end work, combining frontier reasoning with a million-token context window and a broad set of agent tools.
Open full comparisonInkling vs GPT-5.6 Sol
OpenAI's flagship general model for difficult coding, analysis, and professional work, with a very large context window and a broad native tool set.
Open full comparisonInkling vs Claude Fable 5.1
Anthropic's frontier model for ambitious, long-running agentic and coding work, with strong vision and enterprise marketplace availability.
Open full comparisonInkling vs Gemini 3.8 Flash
Google's stable, high-efficiency multimodal model for agents, software work, and large mixed-media inputs at an introductory Flash-tier price.
Open full comparisonInkling vs Grok 4.6
SpaceXAI's frontier text-and-image model for coding, agentic tasks, and knowledge work, with direct web, X, and code tools.
Open full comparisonInkling vs Grok 4.20 Multi-Agent Beta
A specialist Grok model that sends several AI agents to investigate a difficult question in parallel, then combines their work into one researched answer.
Open full comparisonInkling vs Mistral Medium 3.5
Mistral's open-weight frontier model for demanding multimodal, coding, reasoning, and agent workflows, with a 256K context window.
Open full comparisonInkling vs Mistral Small 4
Mistral's lower-cost open model that combines normal instruction following, reasoning, coding, vision, and agent tools in one endpoint.
Open full comparisonInkling vs Qwen3.8 Max
Alibaba's frontier Qwen model for long-context reasoning, coding, tools, and understanding text, images, and video.
Open full comparisonInkling vs MiniMax M3
MiniMax's million-token multimodal model for coding, agents, computer use, and long projects at a low direct API price.
Open full comparisonInkling vs GLM-5.3
Z.ai's open-weight long-context reasoning model for coding, tools, and agentic work across self-hosted and managed routes.
Open full comparisonInkling vs Kimi K3
Moonshot AI's frontier multimodal model for million-token coding, reasoning, knowledge work, and tool-driven agents.
Open full comparisonInkling vs Falcon-H1R-7B
TII's compact open reasoning model for text tasks, long contexts, and self-hosted function-calling workflows.
Open full comparisonInkling vs Falcon-H1-34B-Instruct
TII's 34B open instruction model for multilingual text, coding, and controlled self-hosted applications.
Open full comparisonInkling vs Muse Spark 1.3
Meta's latest multimodal reasoning model for long-running agents, coding, tools, and complex user collaboration.
Open full comparisonInkling vs Celeris-1 Magnus
Celeris' agent-focused diffusion language model for fast reasoning, tool loops, structured actions, and OpenAI-compatible integration.
Open full comparisonInkling vs NVIDIA Nemotron 3 Ultra
NVIDIA's largest Nemotron 3 reasoning model for complex agents, coding, planning, tools, RAG, and million-token analysis.
Open full comparisonInkling vs NVIDIA Nemotron 3.5 Lightning
NVIDIA's compact 30B mixture-of-experts model for efficient specialist agents and high-volume text workflows.
Open full comparisonInkling vs Qwen3.8-Flash
Alibaba's efficient million-context multimodal model for fast agents, coding, document work, vision, video understanding, and tool use.
Open full comparisonInkling vs DeepSeek-V4-Pro
DeepSeek's flagship million-context text model for difficult reasoning, coding, long-running agents, tools, and very large outputs.
Open full comparisonInkling vs DeepSeek-V4-Flash
DeepSeek's lower-cost V4 model for high-volume reasoning, coding, agents, and million-token text workloads.
Open full comparison
