⚡Stack Radar

Model Fine-Tuning · Reviewed 22 Oct 2026

Tinker

API for fine-tuning open-weight models with managed infrastructure.

free tier · thinkingmachines.ai ↗

Paid against free › Model Fine-Tuning › Tinker

Tinker at a glance
330.3KEstimated total trafficModeled monthly visits582Ranking keywordsPublic search sample96/100Technical health15 pages checked7.2KReferring domainsLive backlink summary

Overview

01

Summary

The useful answer before the supporting detail.

Tinker is a managed API for fine-tuning open-weight models, letting researchers and developers train custom models without managing GPU infrastructure S20 S26. It solves the problem of complex post-training by providing low-level primitives and an open-source cookbook for common methods S26. Its practical value is enabling users to control their data and training loop while accessing scalable infrastructure and pay-per-token pricing S32 S1.

02

Product identity and positioning

Name, category, positioning, audience, and first-party claims.

  • Identity: Tinker is Thinking Machines Lab's API for fine-tuning open-weight models S20.
  • Core purpose: It simplifies post-training by handling distributed training complexity while giving users control over algorithms and data S26.
  • Intended users: Researchers and developers who want to customize models for their own applications S26.
  • Product scope: A managed service running on internal clusters, offering training, checkpoint storage, and sampling S26 S32.
  • Positioning: Positioned as frontier-grade infrastructure for training your own models S20.
  • Claimed distinction: Provides low-level primitives like forward_backward and sample to express most post-training methods S26.
03

Use cases and audiences

Who the product serves and the jobs it is designed to handle.

AudiencePractical use case
ResearchersConduct studies on post-training methods and customize models for new applications S40.
DevelopersIntegrate fine-tuned models into custom workflows and applications via API S32.
Security-sensitive industriesUse enterprise controls like SSO and private VPC to fine-tune models with strict data governance S32.
Financial analystsAutomate information triage tasks by training models to replicate expert judgment S29.

Features

09

Features and capabilities

Core functions and evidence-backed product capabilities.

CapabilityWhat it does
Fine-tuning APIProvides low-level primitives like forward_backward and sample to express custom post-training methods S26.
Managed infrastructureRuns on Thinking Machines' clusters, managing GPUs for sampling and training, and storing checkpoints S26 S32.
OpenAI-compatible samplingOffers an inference interface compatible with the OpenAI API, enabling plug-and-play with OpenAI-compatible platforms S39.
Vision inputSupports interleaving image chunks with text for multimodal fine-tuning and inference S39.
Enterprise controlsIncludes single sign-on, role-based access control, audit-log API, and private VPC connectivity for security-sensitive industries S32.
AutoscalingShifts capacity between models as demand changes, handling tens of thousands of concurrent requests S32.
Tinker CookbookOpen-source library with modern implementations of post-training methods that run on top of the Tinker API S26.
10

Integrations and ecosystem

Connected products, exports, APIs, and workflow compatibility.

Integration or interfaceTypeWhat it enables
OpenAI API-compatible samplingAPIAllows sampling from models using OpenAI-compatible scaffolding, enabling integration with any OpenAI-compatible platform S39.
Tinker CookbookOpen-source libraryProvides modern implementations of post-training methods that run on top of the Tinker API S26.
Tinker PlaygroundWeb interfaceOffers a quick chat interface for sampling from checkpoints S32.
Audit-log APIAPIEnables programmatic access to audit logs for enterprise compliance S32.

Pricing

11

Pricing and plans

Published plans, free access, trials, credits, and pricing limitations.

Plan or offerVerified price and billingIncludedLimits
Nemotron-3.5-Lightning-30B-A3B$0.80 / 1M train tokens, pay per tokenSample $0.495 / 1M; prefill $0.184 / 1MCheckpoint storage $0.10 per GB per month S1
Qwen3.5-4B$1.34 / 1M train tokens, pay per tokenSample $1.005 / 1M; prefill $0.33 / 1MCheckpoint storage $0.10 per GB per month S1
AI modelOSS-120B$2.33 / 1M train tokens, pay per tokenSample $0.84 / 1M; prefill $0.33 / 1MCheckpoint storage $0.10 per GB per month S1
Qwen3.5-397B-A17B$12.00 / 1M train tokens, pay per tokenSample $7.50 / 1M; prefill $1.65 / 1MCheckpoint storage $0.10 per GB per month S1
Kimi-K2.6$15.40 / 1M train tokens, pay per tokenSample $5.49 / 1M; prefill $2.205 / 1MCheckpoint storage $0.10 per GB per month S1

Competition

12

Competitive landscape

Domains competing for the same organic search demand.

Sampled competitors

CompetitorFirst-page appearancesSampled queriesAverage positionQueries observed
youtube.com226inkling, interaction model
inkling.com111inkling
linkedin.com111interaction model
huggingface.co112inkling
alphaxiv.org112interaction model
merriam-webster.com113inkling
slideshare.net113interaction model
semaphore.substack.com115inkling
bookdown.org115interaction model
splatoon.fandom.com116inkling

Website intelligence — Collected evidence

Customers, testimonials, and first-party review claims

No supporting public-page evidence was collected for this category.

Independent review sources discovered

No links to supported independent review platforms were discovered on the collected pages.

14

AI visibility

Measured ChatGPT and Google AI mentions with cited sources.

AI-assistant referral traffic

  • Estimated current generative-AI referral share: 1.4%
MonthSource
—All AI assistants
—All AI assistants

Website intelligence — Collected evidence

Site-stated AI capabilities

Page typeEvidenceSource
otherThinking Machines Lab is an artificial intelligence research and product company. We’re building a future where everyone has access to the knowledge and tools to make AI work for their unique needs and goals.https://thinkingmachines.ai/
otherWhile AI capabilities have advanced dramatically, key gaps remain. The scientific community’s understanding of frontier AI systems lags behind rapidly advancing capabilities. Knowledge of how these systems are trained is concentrated within the top research labs, limiting both the public discourse on AI and people’s abilities to use AI effectively. And, despite their potential, these systems remain difficult for people to customize to their specific needs and values. To bridge the gaps, we’re building Thinking Machines Lab to make AI systems more widely understood, customizable and generally capable.https://thinkingmachines.ai/
otherAI that works for everyonehttps://thinkingmachines.ai/
otherMore flexible, adaptable, and personalized AI systems. We see enormous potential for AI to help in every field of work. While current systems excel at programming and mathematics, we’re building AI that can adapt to the full spectrum of human expertise and enable a broader spectrum of applications.https://thinkingmachines.ai/
otherEmpirical and iterative approach to AI safety. The most effective safety measures come from a combination of proactive research and careful real-world testing. We plan to contribute to AI safety by (1) maintaining a high safety bar—preventing misuse of our released models while maximizing users’ freedom, (2) sharing best practices and recipes for how to build safe AI systems with the industry, and (3) accelerating external research on alignment by sharing code, datasets, and model specs. We believe that methods developed for present day systems, such as effective red-teaming and post-deployment monitoring, provide valuable insights that will extend to future, more capable systems.https://thinkingmachines.ai/
otherWe’re building AI systems that push technical boundaries while delivering real value to as many people as possible. Our team combines rigorous engineering with creative exploration, and we’re looking for collaborators to help shape this vision.https://thinkingmachines.ai/
pricingPricing documentationhttps://thinkingmachines.ai/tinker/pricing/
integrationsThis document provides information regarding the datasets used by Thinking Machines Lab to develop its generative artificial intelligence systems and services, including its models (collectively “AI services”). This information reflects general practices, not specific to any single model.https://thinkingmachines.ai/training-data-documentation/
integrationsThinking Machines Lab’s AI services are developed using a variety of data sources, including publicly available data, data provided by partners via access agreements, and internally generated data, including synthetic data.https://thinkingmachines.ai/training-data-documentation/
integrationsThinking Machines Lab uses the collected data to develop its AI services, which are designed to understand and generate content, which may include text, images, audio, and video. The datasets are selected and curated to enable the AI services to develop broad capabilities, which may include language understanding, reasoning, visual comprehension, and audio and video processing.https://thinkingmachines.ai/training-data-documentation/
integrationsThinking Machines Lab’s AI services are developed using datasets containing trillions of tokens. The training data consists of a broad variety of content types, including text, images, audio, and video, drawn from publicly available sources or acquired from third-parties. Thinking Machines Lab also uses synthetically generated data to supplement its training datasets.https://thinkingmachines.ai/training-data-documentation/
integrationsThe datasets Thinking Machines Lab uses to develop its AI services includes content that is in the public domain as well as content that may be subject to intellectual property protection.https://thinkingmachines.ai/training-data-documentation/
integrationsThinking Machines Lab takes steps to clean, process, and modify datasets used for its AI services. These processing steps, which vary by data type, may include deduplication and filtering to remove junk or other low-quality data, and are intended to improve the data’s usefulness for model training and support the responsible development of the AI services.https://thinkingmachines.ai/training-data-documentation/
integrationsThinking Machines Lab collects and uses datasets to train and improve its AI services on an ongoing basis. Thinking Machines Lab began collecting and using datasets for model development in 2025.https://thinkingmachines.ai/training-data-documentation/
integrationsThinking Machines Lab generates and uses synthetic data in the development of its AI services.https://thinkingmachines.ai/training-data-documentation/
changesToday, we are launching Tinker, a flexible API for fine-tuning language models. It empowers researchers and hackers to experiment with models by giving them control over the algorithms and data while we handle the complexity of distributed training. Tinker advances our mission of enabling more people to do research on cutting-edge models and customize them to their needs.https://thinkingmachines.ai/news/announcing-tinker/
changesRedwood Research used Tinker to RL Qwen3-32B on difficult AI control taskshttps://thinkingmachines.ai/news/announcing-tinker/
changesTinker is now in private beta for researchers and developers. You can sign up for the Tinker waitlist here. We will be onboarding users to the platform starting today.https://thinkingmachines.ai/news/announcing-tinker/
changesWe launched Tinker to enable builders and researchers to train models their own way, whether they’re conducting studies or customizing models for new applications. We plan to publish regular roundups of the coolest projects from the Tinker community, and we invite you to submit what you’ve been Tinkering on to be featured on our blog.https://thinkingmachines.ai/news/call-for-community-projects/
changesWe’re interested in featuring ML research projects, AI-enabled research in other domains, custom models, and other contributions. Some examples:https://thinkingmachines.ai/news/call-for-community-projects/

These are first-party product claims, not independent measurements of AI-search visibility.

15

Marketing opportunities

Actions calculated directly from the collected evidence.

OpportunityRecommended actionEvidence
Competitor gapReview the sampled queries where youtube.com appears.2 first-page appearances across inkling, interaction model

Website intelligence — Collected evidence

Website evidence gaps

  • features: no usable public page was collected; verify whether the site publishes this evidence.
  • customers: no usable public page was collected; verify whether the site publishes this evidence.
  • reviews: no usable public page was collected; verify whether the site publishes this evidence.
  • identity: no usable public page was collected; verify whether the site publishes this evidence.
  • technical: no usable public page was collected; verify whether the site publishes this evidence.

Trust and proof

16

Customers, proof, and reviews

Published proof, customer signals, and review evidence.

Independent reviews

PlatformPageHow it was foundNote
eesel AI BlogInkling review: is Thinking Machines' open model worth it? / eesel AIGrounded web searchInkling review: is Thinking Machines' open model worth it? / eesel AI ... The Hacker News launch thread (1,214 points, ~292 comments) is the clearest read on reception. The mood was enthusiastic but measured, and the excitement was less about scores and more about who shipped it: ... That framing got pushback (people named Mistral, Cohere's North, Gemma), but the underlying "a credible Western open-weights option was overdue" sentiment ran through the whole thread. There was also a healthy "how

Customer proof

No named customers, testimonials, or customer case studies were found.

Independent mentions

External links were grounded in one live web search. Search snippets are leads, not independently scraped page content; rejected links remain available in Review evidence.

17

Company, security, and trust

Ownership, policies, security signals, and public trust evidence.

SignalFindingSource
Company and ownership“AI is the most powerful knowledge discovery instrument in human history,” said Jensen Huang, founder and CEO of NVIDIA. “Thinking Machines has brought together a world-class team to advance the frontier of AI. We are thsource
Contact and supportIf you’re an organization interested in using Tinker, please contact us here.source
Legal and privacySee how to apply for proposal requirements, selection criteria, timeline, and complete terms of service for the grant program.source
Security and complianceCustom models draw strong interest from security-sensitive industries, and their requirements set a high bar. We met it with enterprise-grade controls. Organizations divided into projects with data and training access gasource
Uptime and status pageNot published on the collected pages.—
Domain registered2024-07-09—
Domain expires2030-07-09—

Not found on the collected pages: uptime and status page. This only means Scout did not find it on the pages it read; it says nothing about the company itself.

18

Documentation and product updates

Help resources, product guidance, and visible changes over time.

Resource or updateTypeWhat is availableDate or freshness
Training Data Documentation - Thinking Machines LabDocumentationTraining Data Documentation This document provides information regarding the datasets used by Thinking Machines Lab to develop its generative artificial intelligence systems and services, including its models (collectively “AI services”). This information reflects general practices, not specific to any single model. 1. Dataset Sources Thinking Machines Lab’s AI services are developed using a variety of data sources, including publicly available data, data provided by partners via access agreements, and internally generated data, including synthetic data. (source)Not dated
Inkling: Our Open-Weights Model - Thinking Machines LabProduct updateProduct update is publicly available. (source)Not dated

Not found in the collected public sources: public roadmap, changelog or release notes, status page, separate help center or knowledge base.

Technical

19

Technical and delivery signals

Delivery model, platforms, API, data ownership, reliability claims and release activity.

SignalFindingSource
Delivery model (cloud, self-hosted, desktop, browser extension)Managed service running on Thinking Machines' internal clusters and training infrastructure S26.S26
Platforms and appsAccessible via web API and Tinker Playground chat interface S32.S32
API and extensibility (API, SDK, webhooks, MCP, plugins)Provides low-level API primitives like forward_backward and sample, plus an OpenAI-compatible sampling interface S26 S39.S26 S39
Data ownership and exportCustomer data and training loop stay under customer control; checkpoints can be sampled from S32.S32
Performance and reliability claimsAutoscaling handles tens of thousands of concurrent requests without queueing S32.S32
Release activityTinker launched in private beta, then general availability with vision input; ongoing improvements shipped over a year S26 S39 S32.S26 S39 S32
20

Measured infrastructure

Domain, certificate, and hosting signals, measured directly.

SignalFinding
Domain registered2024-07-09 — MarkMonitor Inc.
Domain expires2030-07-09
TLS certificateGoogle Trust Services, valid, expires 2026-11-29
Security headersX-Content-Type-Options, Referrer-Policy
Headers not servedHSTS, CSP, X-Frame-Options
Server headercloudflare
Mail handled byaspmx.l.google.com, alt1.aspmx.l.google.com
Nameserverskellen.ns.cloudflare.com, alexa.ns.cloudflare.com

Not measurable: wayback: rate limited by archive.org, ladder stopped after 1 probe(s). A gap here means the lookup was unavailable, not that the finding is negative.

21

Technical site audit

A bounded multi-page crawl with page-level issues and URLs.

Scout — Bounded technical crawl

  • Pages crawled: 15 of 15 maximum
  • Technical health: 96/100
  • Critical issues: 0
  • Warnings: 10
  • Notices: 6

Issues

SeverityIssueValueAffected URL
warningPage should have one primary H10https://thinkingmachines.ai/
noticeImages are missing alt text90https://thinkingmachines.ai/tinker/
noticeLow visible word count139https://thinkingmachines.ai/blog/
noticeMeta description length should be reviewed31https://thinkingmachines.ai/news/
warningTitle length should be reviewed85https://thinkingmachines.ai/news/one-year-of-tinker/
warningTitle length should be reviewed89https://thinkingmachines.ai/blog/interaction-models/
noticeMeta description length should be reviewed749https://thinkingmachines.ai/legal/terms/
warningPage should have one primary H120https://thinkingmachines.ai/legal/terms/
noticeMeta description length should be reviewed684https://thinkingmachines.ai/legal/privacy/
warningPage should have one primary H113https://thinkingmachines.ai/legal/privacy/
warningTitle length should be reviewed67https://thinkingmachines.ai/security/disclosure-policy/
noticeMeta description length should be reviewed452https://thinkingmachines.ai/security/disclosure-policy/
warningPage should have one primary H19https://thinkingmachines.ai/model-card/inkling/
warningPage should have one primary H19https://thinkingmachines.ai/model-card/inkling-small/
warningDuplicate meta description in audit sample2https://thinkingmachines.ai/
warningDuplicate meta description in audit sample2https://thinkingmachines.ai/blog/

Crawled pages

StatusTitleH1sWordsResponseURL
200Thinking Machines Lab072270 mshttps://thinkingmachines.ai/
200Inkling - Thinking Machines Lab1404100 mshttps://thinkingmachines.ai/inkling/
200Tinker - Thinking Machines Lab14833110 mshttps://thinkingmachines.ai/tinker/
200Connectionism - Thinking Machines Lab113996 mshttps://thinkingmachines.ai/blog/
200News - Thinking Machines Lab117624 mshttps://thinkingmachines.ai/news/
200One Year of Tinker: Turning Your Expertise into Custom Models - Thinking Machines Lab1196059 mshttps://thinkingmachines.ai/news/one-year-of-tinker/
200Inkling: Our Open-Weights Model - Thinking Machines Lab1567174 mshttps://thinkingmachines.ai/news/introducing-inkling/
200Introducing Inkling-Small - Thinking Machines Lab1296083 mshttps://thinkingmachines.ai/news/inkling-small/
200Interaction Models: A Scalable Approach to Human-AI Collaboration - Thinking Machines Lab1403546 mshttps://thinkingmachines.ai/blog/interaction-models/
200The Future Worth Building Is Human - Thinking Machines Lab1216062 mshttps://thinkingmachines.ai/blog/the-future-worth-building-is-human/
200Service Terms of Use - Thinking Machines Lab20924598 mshttps://thinkingmachines.ai/legal/terms/
200Privacy notice - Thinking Machines Lab131899108 mshttps://thinkingmachines.ai/legal/privacy/
200Coordinated Vulnerability Disclosure Policy - Thinking Machines Lab1748102 mshttps://thinkingmachines.ai/security/disclosure-policy/
200Inkling Model Card - Thinking Machines Lab9151155 mshttps://thinkingmachines.ai/model-card/inkling/
200Inkling-Small Model Card - Thinking Machines Lab9161393 mshttps://thinkingmachines.ai/model-card/inkling-small/

This is a bounded Scout crawl. Core Web Vitals and Lighthouse scores are not claimed unless a dedicated performance audit is connected.

22

Search Console and Analytics

First-party search and behavioral data connection state.

Modeled acquisition channels

ChannelEstimated share
direct43.5%
search38.1%
search organic37.9%
search paid0.1%
social6.6%
social organic6.6%
social paid0.0%
referrals9.2%
paid referrals0.1%
mail1.1%
gen ai1.4%
affiliate0.0%
other0.0%

First-party connections

IntegrationStatusData available after connection
Google Search ConsoleNot connectedClicks, impressions, CTR, average position, queries, pages, countries, and devices
Google AnalyticsNot connectedOrganic sessions, engagement, landing pages, conversions, geography, and devices

No first-party Search Console or Analytics numbers are shown until the site owner authorizes the corresponding Google account. Scout traffic model values above are modeled estimates, not private analytics totals.

Limits and method

23

Limitations and open questions

Known constraints and missing evidence from the collected pages.

The collected pages do not mention specific competitors. Evidence gaps include lack of independent reviews, detailed customer case studies, and explicit uptime or status page information. Open questions remain about the full feature set of the serverless inference beta and the exact scope of the enterprise plan.

24

Sources, method, and limitations

Evidence inventory, collection method, provider use, and constraints.

This report is based on 15 collected pages from thinkingmachines.ai, including the Tinker product page, pricing page, training data documentation, and news announcements. Static and browser rendering rendering were used; external proof search was enabled but only verified the existence of linked results. Pricing data was owner-reviewed and read on 2026-10-21. Limitations: no third-party reviews or detailed case studies were collected, and some claims rely on company-authored announcements.

Paid against free, in your inbox

One job at a time: the paid tools worth paying for, the free ones you can run yourself, and who should pick which.

No tracking pixels. One click to leave. We use your address to send Stack Radar and nothing else. Privacy.