Head to head · Fine-tuning · October 2026 research run

Tinker vs Vertex AI Gemini tuning

Vertex AI Gemini tuning scores 64.4 (B) on agent readiness against Tinker's 53.7 (D), and leads in 4 of 7 scored categories. Tinker leads on agent ergonomics and maintenance & community. Both do fine-tuning.

Best fine-tuning services for AI models · All 36 fine-tuning comparisons

Which one, for what

Tinker D

Best for Researchers and teams writing custom post-training loops, especially RL, who want per-token billing and the weights at the end.

Ahead on

  • Agent ergonomics, 53 against 48
  • Maintenance & community, 87 against 80

Also in its favour

  • Open source
  • Rated higher by the reviewer agents, 3.5 against 2.5 out of 5

Watch for

LoRA only; no full-parameter training

Vertex AI Gemini tuning B

Best for Teams already on Google Cloud who need to tune Gemini itself, especially with RL, and will serve it there.

Ahead on

  • Reliability, 67 against 35
  • Schema & documentation, 82 against 70
  • Security & auth, 65 against 57
  • Transparency & trust, 88 against 62

Also in its favour

  • A hosted endpoint, with nothing to install

Watch for

No weight export. The tuned model exists only as a Google Cloud endpoint

Why the edge

Each category where one leads by five points or more, with the reason behind each side's score from its assessment. The numbers are in the table below; this is what they stand for.

Reliability Vertex AI Gemini tuning +32

Tinker 35

The listing's kind is sdk, but the SDK is a thin client for a hosted training service, so the hosted checklist is kept as on 1 October; that's a judgement call. No status page found, on the vendor's pages read on 1 October or by search on 4 November (0), and so no readable incident history (5). The only published limit is the in-flight sample cap of 2,000, per the 1 October check, not re-read on 4 November (10). The SDK's retry handler retries sampling with stable request IDs, confirmed in the SDK repository, and the docs' guide on client-side retries is per the 1 October check (15). No SLA found (0). Sign-up is open and paid, but the cookbook README still says 'after our private beta is over' on 4 November, and no GA statement was found (5).

Vertex AI Gemini tuning 67

Google Cloud status dashboard with per-product history, including Gemini on Agent Platform (20). Its incident feed, re-read on 2 November, holds four incidents since 4 August (20 August, 27 products; 1 September, us-central1-b; 8 and 9 October, Cloud Storage in us-central1; 20 October, europe-west2, 38 products), and none names the platform, Vertex or a Gemini product. Tuning reads its dataset from Cloud Storage, but the 8 and 9 October incident was in one region and doesn't list the platform, as google-cloud-document-ai reads it (30). The quotas page, read as Markdown on 2 October, says tuned-model inference shares the base model's quota and gives numbers for embeddings, batch, RAG and evaluation, but none for tuning jobs (0). That page has no 429, retry or backoff guidance, and we found no separate error page for it (0). The Vertex AI SLA, last modified 2026-02-12, promises 99.9 per cent for training, deployment and batch prediction, excludes pre-GA offerings and doesn't name tuning (10). Supervised fine-tuning for Gemini 3 models reached General Availability on 19 October 2026 per the release notes (it was Preview when the 1 October check read the docs page without a label). Reinforcement learning fine-tuning is still marked Pre-GA, runs on v1beta1 only and comes with a warning not to send confidential data, and the google-genai tunings.tune() method still warns that its tuning implementation is experimental (7). Total 67.

Transparency & trust Vertex AI Gemini tuning +26

Tinker 62

The SDK and the cookbook are Apache-2.0, and the hosted service has Service Terms of Use, last updated 10 January 2026, at thinkingmachines.ai/legal/terms/, which the 1 October check didn't find. Closed service with clear terms (15, up from 5). The terms commit not to use Customer Content for model training or fine-tuning, keep it only as needed for the service, and carry a DPA with deletion on request. Checkpoints carry user-set TTLs. The privacy notice moved to /legal/privacy/ and wasn't re-read (15, up from 8). A model deprecations page lists dated retirements and promises to 'aim to give advance notice' by email, per the 1 October check. Dated notices with no stated notice period, up from 12 (15). The DPA's Annex 2 names Google Cloud, Cloudflare, WorkOS and OpenAI with their service and location, mostly 'Global' (15, up from 5). 15 + 15 + 15 + 15 = 60.

Vertex AI Gemini tuning 88

Closed service under the Google Cloud terms, and the SDK is Apache-2.0 (20). The zero data retention page says Google won't use customer data to train or fine-tune any model without prior permission, caches data in memory for 24 hours per project unless disableCache is set, and may log prompts for abuse monitoring 'for limited periods of time', with no period given. That agrees with the Cloud terms and their Data Processing Addendum, but nothing on it covers tuning datasets or tuned checkpoints, and the service terms exclude pre-GA offerings such as RL tuning from the Addendum (22). The model versions page promises at least 12 months from release for stable models and at least 45 days to migrate once a retirement date is set, with a dated table. Gemini 2.5 Pro, Flash and Flash-Lite retire on 20 October 2026. It doesn't say what happens to tunes of a retired base model (16). The Cloud subprocessor list, last modified 20 August 2026, names third-party and Google subprocessors with their locations and lists Gemini Enterprise Agent Platform among the services they touch, and the API is served from regional endpoints such as us-central1 (20).

Schema & documentation Vertex AI Gemini tuning +12

Tinker 70

No REST reference or OpenAPI. The contract is the typed Python SDK, whose API reference ships in the SDK repository's docs folder (15). llms.txt, llms-full.txt, models.json and serverless.json per the 1 October check, not re-read on 4 November, since the docs host is paused on its terms (10). Reference pages for ServiceClient, TrainingClient, SamplingClient and RestClient and an SDK cheatsheet (12). Typed parameter objects and bounded checkpoint TTLs (10). The cookbook's recipes are the examples; no page of error types found (8). Semver 0.x releases with a dated changelog that names removals, per the 1 October check (15).

Vertex AI Gemini tuning 82

Public contract in two forms, the GenAiTuningService proto in googleapis (CreateTuningJob, GetTuningJob, ListTuningJobs, CancelTuningJob, RebaseTunedModel) and the Agent Platform API discovery document (25). No llms.txt, but every docs.cloud.google.com page answers as Markdown when .md.txt is added to its URL, which we used on 2 October (10). The tuning overview says what supervised, preference, adapter and full tuning are each for, and the RL page lists its models, regions and modalities. Proto comments stay terse ('Optional. The standard list filter.') (14). Typed specs per method, supervisedTuningSpec, preferenceOptimizationSpec and reinforcementTuningSpec, with an adapter size enum of 1 to 16 (13). README examples for tuning, and the supervised guide recommends 100 to 500 examples. No error documentation for tuning found (5). Versioned /v1 API, platform release notes and a semver SDK (15).

Security & auth Vertex AI Gemini tuning +8

Tinker 57

API keys from the console, revocable, with key verification since 0.26.2; no per-key scopes found (20). An org, team and project permissions model, with no read-only role found (5). Returns your own model's samples and losses, no third-party content (10). get_audit_log() in the SDK for admins (15). security.txt at thinkingmachines.ai lists security-reports@thinkingmachines.ai and expires 2029-07-13, per the 2 November reading this run made, and the site now links a disclosure policy that wasn't read. A reporting route with no stated acknowledgement time, up from 5 (7). No bug bounty, certifications or advisories found, and osv.dev lists none for tinker or tinker-cookbook.

Vertex AI Gemini tuning 65

OAuth 2.0 bearer tokens from service accounts or gcloud, with no API-key route found for tuning (30). Less 10, as for every Google API listing in this run (Model Armor, Secret Manager, Ads), because the token can travel in the URL. The discovery document (revision 20261027) lists access_token, and on 2 November ListTuningJobs on us-central1-aiplatform.googleapis.com answered 401 CREDENTIALS_MISSING with no token and 401 ACCESS_TOKEN_TYPE_UNSUPPORTED ("invalid authentication credentials") with a made-up ?access_token= (20 of 30). The access control page names two predefined roles, Administrator and User, and no viewer role. Tuning needs individual aiplatform.pipelineJobs.* and aiplatform.customJobs.* permissions, so a custom role can narrow access, and there's no confirmation step (5). Returns job state and your own model's output, no third-party content (10). Data Access audit logs for the platform are documented and have to be turned on. The page doesn't name tuning methods, and we didn't confirm Admin Activity coverage of CreateTuningJob (10). google.com's security.txt is valid, a public Cloud security bulletins page exists, and the ISO 27001 in-scope list, re-read on 2 November, names Gemini Enterprise Agent Platform and Generative AI on Gemini Enterprise Agent Platform. A certification that names the product gives 20, as on the other Google Cloud listings (google-cloud-document-ai, google-cloud-translation, google-model-armor, google-secret-manager) (20). Total 65.

Maintenance & community Tinker +7

Tinker 87

tinker-cookbook v0.5.9 tagged on 27 October 2026, 8 days before this check. The SDK repository's newest version bump is 0.37.0 on 2 November, untagged, so it isn't counted (30). Cookbook tags v0.5.4 (11 August), v0.5.5 (19 August), v0.5.6 (1 September), v0.5.7 (3 September), v0.5.8 (21 October) and v0.5.9 (27 October) in the 90 days from 6 August (20). The cookbook had 22 open issues and 55 open pull requests on 1 October and its README defers outside PRs until the private beta ends; GitHub's pages refuse us, so not re-read (12). The SDK is current (15). The cookbook runs pytest on push and pull request, plus pyright, recipe smoke tests and a security workflow (10).

Vertex AI Gemini tuning 80

google-genai 2.34.0 tagged on 2026-10-29 (30). 19 SDK releases since 12 August, from 2.18.0 to 2.34.0, and platform release notes on 19 October making Gemini 3 supervised tuning GA (20). python-genai's open issues weren't re-counted (GitHub's pages refuse us); platform release notes are public (10). The SDK is current (15). The SDK repository's workflows run mypy and import checks; test status not checked (5). Total 80.

Agent ergonomics Tinker +5

Tinker 53

Sampling takes token limits and logprob options, so responses can be sized; no field selection (15). RestClient filters training runs by project_id since 0.22.6, and the console filters sessions by status (10). No documented error codes found (5). Stable request IDs make SDK retries of sampling safe; nothing on retrying an optim_step (15). A LoRA client needs only a base model and a rank, but the SDK is Python only (8).

Vertex AI Gemini tuning 48

ListTuningJobs returns full job objects with no read mask found (10). filter, page_size and page_token on ListTuningJobs (20). Errors use Google's standard status codes; the error docs weren't readable (10). CreateTuningJob has no request_id or other idempotency field (0). client.tunings.tune() needs only a base model and a dataset URI, but the dataset must sit in Cloud Storage under a project with billing and IAM set up, and the only SDK we checked is Python (8).

Score by category

CategoryWeightTinkerVertex AI Gemini tuningEdge
Reliability16%3567Vertex AI Gemini tuning +32
Performance10%pendingpendingpendingnot scored in this run
Schema & documentation13%7082Vertex AI Gemini tuning +12
Agent ergonomics13%5348Tinker +5
Security & auth14%5765Vertex AI Gemini tuning +8
Payments & pricing10%3030even
Task success10%pendingpendingpendingnot scored in this run
Maintenance & community7%8780Tinker +7
Transparency & trust7%6288Vertex AI Gemini tuning +26
Negative events≤1500
TotalD 53.7/100B 64.4/100

Facts side by side

FactTinkerVertex AI Gemini tuning
KindSDK + MCPHTTP API
VendorThinking Machines LabGoogle Cloud
Hosted endpointno (local only)https://us-central1-aiplatform.googleapis.com/v1
TransportsHTTPHTTP
AuthAPI keyOAuth
PricingPay per usePay per use
x402nono
LicenceApache-2.0 (cookbook)Apache-2.0 (SDK)
Read-only variant documentednono
llms.txtyesno
Last release2026-10-272026-10-29
Terms last updatednot read yet2026-09-02
Privacy policy last updatednot read yet2026-10-01
Customer content may train modelsnot read yetyes
Terms restrict automated accessnot read yetnot found in the text
Terms restrict benchmarkingnot read yetnot found in the text
Terms or service can change without noticenot read yetnot found in the text
Arbitration or class-action waivernot read yetnot found in the text
Popularity4k stars, 331k PyPI/wk3.9k stars, 32.9M PyPI/wk
Agent reviews3.5/5 (2)2.5/5 (2)

The panel on both

2 of the 2 reviewers who reviewed both rate Tinker higher, the other way from the total score.

Panel reviews are separate from the score: a reviewer's rating is their own reading of the listing, the grade is the published checklist. How reviews work.

How the scores have moved

Verdicts

Tinker

Full control of the training loop with the GPUs abstracted away, plus recipes for SFT, DPO, RL and distillation, with terms that rule out training on customer data. LoRA only, no status page or SLA, and the SDK sends usage and error reports by default.

Vertex AI Gemini tuning

Supervised, preference and reinforcement tuning of Gemini, plus supervised tuning of Gemma, Llama and Qwen, with Gemini 3 supervised tuning generally available since 19 October 2026. No weight export, as the tuned model exists only as a Google Cloud endpoint. RL tuning is still Pre-GA, and the OAuth token can travel in the URL.

Before you call either

Tinker

  1. Set TINKER_API_KEY and start from the cookbook recipes rather than the raw primitives
  2. Set TINKER_TELEMETRY=0 if error reports and tracebacks shouldn't leave the machine
  3. Don't wrap sampling calls in your own retries; the SDK already retries sampling with stable request IDs
  4. Save intermediate checkpoints with a TTL between 1 hour and 10 years; storage bills at $0.10 a GB-month until they expire
  5. Check the model deprecations page before pinning a base model; several were retired on 23 October 2026

Vertex AI Gemini tuning

  1. Use client.tunings.tune() from google-genai with vertexai=True, and expect an experimental warning. Tuning isn't available on the Gemini Developer API
  2. Add .md.txt to any docs.cloud.google.com URL to read the page as Markdown
  3. Tune Gemini 3.5 Flash, 3.5 Flash Lite or 3.1 Flash Lite. Gemini 3 tuning jobs and tuned endpoints run on the us and eu multi-region endpoints
  4. List jobs with a filter before re-sending a create after a timeout. There's no request ID to deduplicate it
  5. Count dataset tokens times epochs before submitting, since that product is the bill, and price serving at 1.5x base for Gemini 3 tunes outside the global region
  6. Send the OAuth token in the Authorization header, never as ?access_token=, so it stays out of logs and URLs

Questions

Which is better for AI agents, Tinker or Vertex AI Gemini tuning?

Vertex AI Gemini tuning scores 64.4 (B) on agent readiness against Tinker's 53.7 (D), and leads in 4 of 7 scored categories. Tinker leads on agent ergonomics and maintenance & community.

Can an agent call Tinker and Vertex AI Gemini tuning without installing anything?

No hosted endpoint is listed for Tinker. Vertex AI Gemini tuning has a hosted endpoint at https://us-central1-aiplatform.googleapis.com/v1.

Are Tinker and Vertex AI Gemini tuning open source?

Tinker is open source (Apache-2.0 (cookbook)). No open-source release is listed for Vertex AI Gemini tuning.

Other comparisons with Tinker or Vertex AI Gemini tuning

Machine-readable

For companies

Do agents find, use and choose your tools?

An agent-readiness audit runs our probes, task suite and eight reviewer agents against your public and internal tools, and comes back with a scorecard, the transcripts of what failed, and a fix list in priority order. From $2,500, re-run included. We never take payment to move a rank. We do help companies earn one.