Custom model development
Fine-tune a strong multimodal base on proprietary tasks, preferences, or domain data and export the resulting weights.
Independent tool overview
Inkling at a glance
Inkling is Thinking Machines Lab's open-weight multimodal model for text, images, and audio, built for adjustable reasoning, coding, tool use, and customization through Tinker.
Visit the official Inkling site ↗
- Developer
- Thinking Machines Lab
- Architecture
- 975B MoE; 41B active parameters
- Context
- Up to 1M tokens; 64K or 256K on Tinker
- Inputs
- Text, images, and audio
- License
- Apache 2.0
Overview
What Inkling is
Inkling is a general-purpose, open-weight model from Thinking Machines Lab. It is a 975-billion-parameter mixture-of-experts model with 41 billion active parameters and native reasoning across text, images, and audio.
Its defining feature is customization. The full Apache-2.0-licensed weights are available for self-hosting, while Thinking Machines' Tinker platform provides playground access, serverless inference, fine-tuning, and downloadable custom checkpoints. Reasoning effort can be adjusted to trade speed and cost against task performance.
Inkling-Small is now a full member of the family rather than a preview. It uses 276 billion total parameters and 12 billion active parameters, retains the same multimodal and controllable-effort approach, and is the more practical starting point when latency, hosted cost, or deployment scale matters.
Use cases
Who Inkling is best for
The strongest fit depends on the job you need the product to complete, not the size of its feature list.
Agentic coding and tools
Build coding assistants and agents that combine long reasoning, tool calls, and large working contexts.
Audio and image understanding
Create applications that reason jointly over text, images, and audio rather than relying on separate single-modality models.
Open-weight research
Inspect, host, customize, and evaluate a permissively licensed model without depending entirely on one hosted API.
Capabilities
Core Inkling features
Native multimodality
Inkling accepts text, image, and audio inputs in a shared model and generates text outputs.
Controllable thinking effort
Developers can vary the amount of reasoning to balance quality, latency, and generated-token cost.
Agentic capabilities
The model is trained for coding, tool use, instruction following, and long iterative workflows.
One-million-token context
The open model supports up to 1 million tokens, while Tinker currently offers 64K and 256K configurations.
Tinker customization
Tinker supports supervised and reinforcement fine-tuning, sampling, a playground, and downloadable checkpoints.
Inkling-Small
A 276B-total, 12B-active sibling offers similar modalities and reasoning controls with lower compute and hosted prices.
Process
How the Inkling workflow works
Step 1
Choose the model size
Benchmark Inkling-Small first, then move to full Inkling only when its quality gains justify higher inference, training, and deployment costs.
Step 2
Set context and effort
Use the shortest viable context and lowest reasoning effort that meets the task's quality target.
Step 3
Build an evaluation set
Measure factuality, task completion, tool use, latency, safety, and cost on representative examples before fine-tuning.
Step 4
Customize on Tinker
Train against domain data or a verifiable reward, compare the checkpoint with the base model, and retain a holdout set.
Step 5
Choose deployment
Use Tinker's beta serverless inference for testing, a supported partner for production, or self-host when control justifies the infrastructure.
Cost
Inkling pricing and free plan
Tinker prices sampling, fine-tuning, and beta serverless inference per million tokens. Inkling and Inkling-Small currently carry limited-time 50% fine-tuning discounts, so the listed promotional rates can change.
Inkling serverless beta
$1.00 input / $0.17 cached / $4.05 output per 1M
256K-context hosted inference for evaluation and lighter use.
- Serverless beta
- Not recommended for intensive production yet
- Usage based
Inkling-Small serverless beta
$0.30 input / $0.06 cached / $1.20 output per 1M
Lower-cost 256K-context inference for the smaller model.
- Text, image, and audio
- Serverless beta
- Usage based
Inkling 64K on Tinker
$1.87 input / $4.68 sample / $5.61 train per 1M
Limited-time discounted 64K pricing for sampling and fine-tuning full Inkling.
- Cached input: $0.374 per 1M
- 50% promotional discount
- Checkpoint storage extra
Inkling-Small 64K on Tinker
$0.58 input / $1.44 sample / $1.73 train per 1M
Limited-time discounted 64K pricing for the smaller model.
- Cached input: $0.116 per 1M
- 50% promotional discount
- Checkpoint storage: $0.10/GB-month
Pricing checked . Check current pricing at the source ↗
Assessment
Inkling strengths and limitations
Where it stands out
- Open weights and an Apache 2.0 license support customization and deployment flexibility
- Text, image, and audio inputs are handled natively in one model
- Adjustable reasoning effort creates a direct quality, latency, and cost control
- Tinker connects experimentation, fine-tuning, evaluation, and checkpoint export
- Inkling-Small offers a more deployable option without leaving the model family
What to consider
- The full Inkling checkpoint is roughly 1.9 TB and self-hosting requires substantial specialized infrastructure
- Tinker's hosted context options are 64K and 256K rather than the model's full 1M context
- Tinker serverless inference is beta and is not recommended by the vendor for intensive production use
- Current Tinker fine-tuning prices include a limited-time discount and may rise
- Thinking Machines says Inkling is not the strongest model overall, so its value depends on customization, modalities, and openness
- Published benchmark and capability results are primarily vendor-reported and should be reproduced on target workloads
Compare
Inkling alternatives
The right alternative depends on the specific output, workflow, controls and budget your project requires.
Business Operations
DeepSeek
A lower-cost open-weight model family with 1M context and mature hosted APIs.
Explore DeepSeek →Agents
Kimi K2.5
Another large open-weight multimodal agent model with strong coding and tool-use capabilities.
Explore Kimi K2.5 →Agents
Mistral AI
A broader open-weight platform with multiple model sizes and enterprise deployment options.
Explore Mistral AI →Project Management
Claude
A managed alternative for teams that value polished reasoning and coding tools over weight access.
Explore Claude →Questions
Inkling FAQs
Is Inkling open source?
Inkling's full model weights are published under the Apache 2.0 license. Open weights provide broad use and modification rights, though the complete Tinker hosted platform is a separate commercial service.
What can Inkling understand?
Inkling accepts text, image, and audio inputs and produces text. It is trained for general reasoning, coding, tool use, instruction following, and multimodal tasks.
What is the difference between Inkling and Inkling-Small?
Full Inkling has 975B total and 41B active parameters. Inkling-Small has 276B total and 12B active parameters, making it less expensive and easier to serve while retaining the family's main modalities and effort controls.
Can Inkling use a 1M-token context window?
The open model supports up to 1 million tokens. Tinker currently provides 64K and 256K configurations, so the usable limit depends on the deployment.
How much does Inkling cost?
Tinker's beta serverless Inkling pricing is $1.00 per million input tokens, $0.17 for cached input, and $4.05 for output. Fine-tuning and sampling have separate usage-based rates, with a limited-time 50% discount currently shown.
Can I fine-tune Inkling?
Yes. Tinker supports fine-tuning and lets users download trained checkpoints. The open weights also allow customization through supported third-party and self-hosted stacks.
Bottom line
Our Inkling verdict
Inkling is most compelling as a customizable multimodal foundation model, not simply another chat assistant. Its open weights, audio support, controllable reasoning, and Tinker workflow give developers unusual flexibility. Inkling-Small should be the default evaluation candidate for cost and deployment practicality, with full Inkling reserved for workloads where measured quality gains justify the much larger footprint.
Visit Inkling website ↗Similar Consumer Tools
View all Consumer AI tools →
Bonsai 27B
Bonsai 27B - PrismML's compressed 27B model small enough for iPhones
Hint
Hint - Martha Stewart's AI home app for maintenance, repairs, and fair quotes
ChatGPT Work
ChatGPT Work- OpenAI's Codex-powered agent for everyday non-coding work
.webp)
Muse Glimmer
Muse Glimmer - Meta's open-weights local model for on-device agents
AI training for the future of work.
Get access to all our AI courses, hundreds of real-world AI use cases, live expert-led workshops, an exclusive network of AI early adopters, and more.
AI Courses
Get unlimited access to all of our current & upcoming industry-specific AI courses for the duration of your subscription.
Daily Guides
To keep up with the rapid pace of AI, our team publishes AI implementation guides daily. Our library contains 300+ practical use cases to automate real-world work.
Workshops
Join weekly, live, interactive sessions with industry leaders who are at the forefront of AI for hands-on implementation guidance and exclusive insights.
Community
Network with an exclusive community of AI-first professionals who are working smarter with AI. Learn how early adopters are using AI in their work and businesses.