Tinker

LLM Fine-Tuning / Training API

Tinker is Thinking Machines Lab’s training API for researchers and developers. You write the data, loss, and loop on a CPU machine; Tinker runs distributed LoRA fine-tuning and sampling on large open-weight models (dense and MoE, including multimodal), with downloadable weights.

Alternatives & similar tools

Related picks from editorial notes.

Also mentioned

Together Fine-Tuning / Fireworks (hosted fine-tunes)
OpenAI fine-tuning API (closed models)
Self-managed Axolotl/Unsloth on rented GPUs

Features & details

Full overview from our catalog (read-only reference).

Category (short)

LLM Fine-Tuning / Training API

Sponsored note

Target audience

Technical users who want algorithmic control over training. Not ideal for: no-code “upload CSV and chat with a fine-tune” buyers, or teams needing full-weight fine-tuning rather than LoRA.

Best for

Technical users who want algorithmic control over training. Not ideal for: no-code “upload CSV and chat with a fine-tune” buyers, or teams needing full-weight fine-tuning rather than LoRA.

Pricing notes

Verified August 2026 at https://tinker-docs.thinkingmachines.ai/tinker/models/. Usage-based only; no published seat subscription. Examples: GPT-OSS-20B prefill $0.18 / sample $0.45 / train $0.396 per 1M tokens; Qwen3-8B $0.195 / $0.60 / $0.44; Inkling currently 50% off list (e.g. train $5.61 vs $11.22). Checkpoint storage $0.10/GB/month. Cached prefill 80% off. Machine-readable models.json provided. Sign-up: https://auth.thinkingmachines.ai/sign-up ; console https://tinker.thinkingmachines.ai/. Model retirements occur (many Llama/Qwen IDs retired mid-2026)—check live table before quoting a specific model.

Pros & cons

Editorial notes to help compare fit before opening the vendor site.

Pros

Clean training primitives without cluster ops
Published per-token tables and JSON feed
Wide open-weight model catalog including MoE efficiency

Cons

No no-code path
LoRA-only (not full FT)
Prices and model SKUs change
retirements happen
Can exceed DIY GPU cost for small dense runs
--
*Ambiguity / QA flags for the parent batch file**
**Caret**: list as **Careti**
`caret.team` is expired.
**BidHelm**: Google Ads agent, not gov contracting.
**Spira**: `spira.ai` social agents ≠ Inflectra SpiraTest
ignore `tryspira.ai` ladder unless confirmed same product.
**Scrollsequence**: Classic vs v2 AI beta dual surface.
**Ideamap / ChatPlayground / Averi Agency / Atoms Max disk**: soft spots called out in Pricing/Review Notes.
--
## Wave 1C (tools 28–42)

Review notes

Context from the listing review and editorial research.

NEW tool. DISAMBIGUATION: Thinking Machines Lab Tinker (training API) ≠ random “Tinker” toys/apps. Product URL path thinkingmachines.ai/tinker; docs are the pricing source of truth. Mira Murati’s lab; expect researcher audience.

Extended features

In-depth description and capability notes.

LoRA fine-tuning API (`forward_backward`, `optim_step`, `sample`, checkpoint save/load) Models from small dense up to frontier MoE (Qwen, GPT-OSS, DeepSeek, Kimi, Inkling, Nemotron, etc.) Vision/audio-capable models (e.g. Inkling family) Per-model prefill / sample / train meters; 80% off cached prefill Checkpoint storage billing Python SDK / cookbook examples; console at tinker.thinkingmachines.ai Serverless inference beta for select Inkling models

Use cases

ML engineers and researchers run custom SFT/RL/post-training loops without operating their own multi-GPU clusters, then export LoRA weights to their inference stack.