AI agents · OpenClaw · self-hosting · automation

Quick Answer

Tinker vs OpenAI Fine-Tune vs Together AI (Jul 2026)

Published:

Tinker vs OpenAI Fine-Tune vs Together AI (Jul 2026)

With Inkling’s release on July 15, 2026, Thinking Machines Lab’s Tinker customization platform became newly relevant. Tinker is the commercial hook — Inkling is the base model, Tinker is where the money is made.

But Tinker doesn’t operate in a vacuum. Together AI has been the default open-weight fine-tuning platform since 2024. OpenAI’s Model Customization ships closed-weight fine-tunes at scale. AWS Bedrock, Google Vertex AI, and Azure ML all offer fine-tuning services. And Modal, Anyscale, and Runpod let you run raw fine-tune jobs yourself.

Here’s how the three most likely candidates for your enterprise fine-tune project actually compare.

Last verified: July 19, 2026

Head-to-Head Overview

SpecTinker (Thinking Machines)OpenAI Model CustomizationTogether AI
VendorThinking Machines LabOpenAITogether AI
Public sinceLate 2025 (as customization platform); expanded July 2026 with InklingPublic since 2023, expanded through 2026Public since 2023, actively expanded through 2026
Model coverageInkling, Hugging Face open-weight modelsGPT-4o mini, GPT-5.5 and 5.6 variants (per OpenAI’s list)Wide open-weight catalog (Inkling, Kimi K3, DeepSeek V4 Pro, GLM-5.2, Muse Spark, Qwen 3.5, Llama family)
Portable weightsYes (you get open-weight fine-tune)No (runs on OpenAI infra only)Yes (you get open-weight fine-tune)
Fine-tune methodsFull FT, LoRA, QLoRA, PEFTVendor-managed (limited method choice)Full FT, LoRA, QLoRA, DPO, RLHF variants
Managed hosting after FTYes (Tinker-hosted inference)Yes (OpenAI-hosted inference)Yes (Together-hosted inference)
Bring-your-own inferenceYes (portable weights)NoYes (portable weights)
Pricing modelNot fully public at launch; expected premium-managed pricingHigher per-training-token; includes managed inferenceCheaper per-training-token; separate inference pricing
Best forEnterprise open-weight fine-tune with integrated workflowFine-tuning closed OpenAI modelsCost-optimized open-weight fine-tuning at scale
Data privacyEnterprise contract; standard SLAsOpenAI privacy defaultsEnterprise contract; SOC 2
Deployment paths after fine-tuneTinker-hosted, self-host, third-party providersOpenAI-hosted onlyTogether-hosted, self-host, third-party providers

What Each Platform Actually Does

Tinker — the integrated open-weight customization stack

What it is: Thinking Machines Lab’s fine-tuning platform, designed around the Inkling base model but also supporting other open-weight bases from Hugging Face.

Key features:

  • Integrated with Inkling — one-click fine-tune workflows against the Thinking Machines model.
  • Managed data upload + validation — same UX pattern as OpenAI’s fine-tune console.
  • PEFT / LoRA / QLoRA support — standard parameter-efficient fine-tune methods.
  • Managed inference post-training — spin up an endpoint immediately.
  • Portable weights — you can download the fine-tuned weights and run them elsewhere.
  • Tinker Cookbook — Thinking Machines has published Cookbook-style recipes for common enterprise use cases.

What differentiates Tinker:

  • Curated model catalog. Not “any Hugging Face model” — Thinking Machines vets bases they’ll fine-tune. Inkling is the default.
  • Integrated with Thinking Machines’ research thesis. Publications on determinism, on-policy distillation, RLHF variants inform the Tinker feature set.
  • Enterprise-first positioning. Not competing on raw fine-tune price; competing on managed workflow quality for enterprise buyers.

OpenAI Model Customization — closed-weight fine-tune with OpenAI’s managed hosting

What it is: OpenAI’s official fine-tuning offering for GPT-4o mini and GPT-5.x variants. Runs entirely inside OpenAI’s platform.

Key features:

  • Fine-tune GPT-4o mini, GPT-5.5 and 5.6 variants (per OpenAI’s supported model list).
  • Managed workflow — data upload, validation, training-run monitoring in the OpenAI console.
  • Reinforcement-learning fine-tune (RFT) for supported models.
  • Direct integration with OpenAI’s Response API and Assistants API — your fine-tune is a first-class model ID.

What differentiates OpenAI Model Customization:

  • Closed-weight base. You’re fine-tuning a proprietary model; weights never leave OpenAI’s platform.
  • Best base capability. GPT-5.6 Sol at the base is stronger than any open-weight base for many benchmarks.
  • Integrated with the rest of OpenAI’s stack — Response API, tool use, structured output, etc., all work with fine-tuned models automatically.

Limitations:

  • Portability = zero. You cannot self-host, export, or use elsewhere.
  • Per-token training cost is high.
  • Fine-tune method choice is limited — OpenAI abstracts most of it away.

Together AI — the open-weight fine-tune commodity provider

What it is: Together AI has been the default open-weight fine-tune platform since ~2023. Broad model catalog, competitive pricing, mature enterprise features.

Key features:

  • Wide model catalog — Inkling, Kimi K3, DeepSeek V4 Pro, GLM-5.2, Muse Spark, Qwen 3.5, Llama descendants, and effectively everything on Hugging Face’s popular list.
  • Multiple fine-tune methods — full FT, LoRA, QLoRA, DPO, and RLHF variants.
  • Cheap per-training-token pricing — historically $2-5 per million training tokens depending on model size (verify current pricing on their site).
  • Managed inference endpoints post-training.
  • Portable weights — download and run anywhere.
  • Serverless inference tier — pay-per-token on their infrastructure without managing endpoints.

What differentiates Together AI:

  • Model-agnostic. No preferred base; they’ll fine-tune anything commercially viable.
  • Pricing pressure — historically the cheapest managed option in the space.
  • Infrastructure focus. Not competing on curated workflow; competing on running inference and training at scale efficiently.

Sub-Questions People Are Asking

Which one should I default to for a first enterprise fine-tune project?

If you’re open to open-weight base models and want integrated managed workflow: Tinker (especially if fine-tuning Inkling).

If you must fine-tune GPT-5.6 or want to stay inside OpenAI’s ecosystem: OpenAI Model Customization.

If you want cheapest managed fine-tune with the widest model catalog: Together AI.

How does Tinker’s pricing compare?

Not fully public at the July 2026 Inkling launch. Thinking Machines positions Tinker as an enterprise premium managed offering — expect pricing above Together AI’s commodity rates and below OpenAI Model Customization’s per-token training rates. The value prop is workflow quality and Inkling integration, not raw price.

Can I fine-tune Kimi K3 or DeepSeek V4 Pro?

Yes, via Together AI, Modal, or Anyscale (any open-weight-friendly infrastructure). Tinker’s Inkling-first positioning may or may not include Kimi K3 in the curated catalog — check their docs. Kimi K3’s 2.8T total parameters make fine-tuning genuinely expensive; most teams will use LoRA / QLoRA rather than full fine-tunes.

What about Anthropic Claude fine-tuning?

As of July 2026, Anthropic does not offer a general public fine-tuning API for Claude. They have limited enterprise fine-tuning programs — talk to Anthropic sales if you have a specific use case. For most teams, “fine-tune Claude” isn’t currently an option; use RAG or prompt engineering instead.

When does fine-tuning beat RAG?

Fine-tuning wins for:

  • Behavior / voice / brand consistency — teach the model your tone.
  • Structured output formats — enforce JSON, XML, or custom schemas reliably.
  • Narrow-domain reasoning — improve on tasks where the base model is close but not quite right.
  • Latency-sensitive workloads — a smaller fine-tuned model can beat a larger base model + long-context RAG.

RAG wins for:

  • Domain knowledge that changes (docs, product info, news).
  • Retrieval over large corpora — where the answer is “which of these 10,000 documents is relevant.”
  • Auditability — the retrieved document is your citation.

Most enterprise deployments use both: fine-tune for behavior, RAG for facts.

How do these compare to AWS Bedrock, Azure ML, Google Vertex fine-tuning?

Cloud provider fine-tuning (Bedrock, Vertex, Azure ML) tends to be model-catalog-limited (each provider curates a smaller catalog than Together AI) and priced with cloud-provider premiums. They’re strongest when you’re already deeply committed to that cloud’s stack (IAM, S3, VPCs). Tinker and Together AI compete more on model coverage and workflow.

For most enterprises, the decision tree is:

  • Already deep on AWS/Azure/GCP: use their fine-tune (Bedrock / Vertex / Azure ML).
  • Multi-cloud or cloud-agnostic: Together AI for cost or Tinker for workflow.
  • Locked into OpenAI: OpenAI Model Customization.

Is Modal / Anyscale / RunPod a better DIY option?

Yes, if you have the ML engineering to run fine-tune jobs directly on GPU rentals. You give up managed workflow but get full control and often cheaper aggregate cost at scale. Larger AI teams (10+ ML engineers) may prefer this. Smaller teams should use managed platforms.

Which one is best for LoRA / PEFT vs full fine-tune?

For LoRA / QLoRA (the common case), all three platforms handle it. Together AI is cheapest; Tinker is smoothest for Inkling; OpenAI is closed-only.

For full fine-tune on very large models (Kimi K3, Fable 5-scale), you generally need managed infrastructure — Together AI or Tinker are appropriate; Modal / Anyscale if you’re doing DIY.

How do I evaluate fine-tune quality?

Standard eval loop:

  1. Hold out a validation set from your training data.
  2. Compare pre-fine-tune and post-fine-tune accuracy on validation.
  3. Test on real production inputs (shadow traffic if possible).
  4. Compare cost / latency to base model + RAG.

All three platforms give you enough logs and endpoints to run this. Tinker publishes Cookbook recipes with example eval loops.

The Real Decision Framework

Choose Tinker if:

  • You want to fine-tune Inkling or the Thinking Machines curated catalog.
  • Integrated managed workflow matters more than the lowest per-token price.
  • You’re a mid-market or enterprise buyer.
  • You may want to move the fine-tuned weights off-platform later.

Choose OpenAI Model Customization if:

  • You must fine-tune a GPT model specifically.
  • You’re already deep in OpenAI’s Response / Assistants API stack.
  • Portability doesn’t matter — you’ll run inference on OpenAI regardless.

Choose Together AI if:

  • You want the widest open-weight model catalog.
  • Cost per training token matters most.
  • You have the ML engineering to run more of the workflow yourself.
  • You need multi-cloud or cloud-agnostic deployment.

Choose a cloud provider’s fine-tune if:

  • You’re deeply committed to AWS, Azure, or GCP.
  • IAM / security / networking integration with your existing cloud stack is critical.
  • You’ll accept a narrower model catalog for tighter cloud integration.

Choose DIY (Modal / Anyscale / RunPod / your own GPUs) if:

  • You have serious ML engineering.
  • You have specific method / hyperparameter requirements that platforms don’t expose.
  • You’re operating at a scale where the platform premium is meaningful.

Bottom Line

Fine-tuning is entering an interesting moment. Tinker’s launch alongside Inkling gives Thinking Machines a coherent enterprise pitch — open-weight base + managed customization + portable weights. OpenAI keeps the closed-weight fine-tune market with best-in-class base models. Together AI keeps the price-competitive open-weight market.

Which one to use depends on:

  • What base model — closed frontier (OpenAI) or open-weight (Tinker / Together AI).
  • What matters more — workflow quality (Tinker) or cost per token (Together AI).
  • Where inference will run — vendor-hosted only (OpenAI) or portable (Tinker / Together AI).

For most enterprises starting fine-tuning in July 2026: try Together AI first for cost-effective proof-of-concept, then evaluate Tinker if you want tighter Inkling / enterprise-managed workflow, or move to OpenAI if closed-weight capability wins the tradeoff.

The right long-term answer for many enterprises is probably both a closed-weight and an open-weight fine-tuning workflow — closed-weight for the highest-stakes reasoning, open-weight for cost-optimized specialized behavior.

Sources