Skip to content

Kojen vs · Fine-tuning

Kojen vs Unsloth

Unsloth is a free, open-source library (and local Studio app) for the fastest, most memory-efficient fine-tuning you can run on your own GPU, covering LoRA, QLoRA, full fine-tuning, DPO and GRPO across many open models; Kojen is a hosted dojo whose text LoRA trains adapters on two 8B models in early access, on rented GPUs you pay for by the second, next to evaluations, red-teaming and prompt tuning. If you have a GPU and want control over training, Unsloth is the better tool; if you want to find out whether you need fine-tuning at all, and test the result, start with Kojen.

Unsloth is far more capable at training weights. Kojen's case is everything around the training run.

By Will SchulzUpdated 6 sources

Side-by-side comparison

"Partial" means the feature exists with real limits, given in the note. "Not verified" means we couldn't confirm it from the vendor's own material.

Kojen vs Unsloth: features and pricing, checked November 2026
Checked Nov 15, 2026UnslothKojen
What it isFine-tuning

Unsloth AI (independent)

Training dojo

Kojen (that's us)

PricingOpen source: free. Pro and Enterprise tiers (multi-GPU, multi-node) are listed without public prices. SourceFree to start, no card needed; then pay as you go. GPU training (LoRA) is paid in credits billed by the second at the GPU's live price. Team and volume pricing by email. Source
Open sourceYes, Apache-2.0 core; some parts AGPL-3.0 RepoNo
Evaluations on datasets

Run a model or app over a set of test cases and score the results.

No
Yes

Test cases scored for accuracy, hallucination, bias and latency

LLM-as-judge scoring

A model grades outputs against a rubric or reference.

No
Yes

Judge kept out of the target's model family in simulations

pass@k / pass^k

Several samples per case, reported as any-success (pass@k) or all-success (pass^k).

No
Yes

1–10 samples per case

Confidence intervals

Rates reported with an interval or a significance test, not just a point estimate.

No
Yes

Wilson and bootstrap intervals on every rate

Prompt optimization

Automatically rewrites instructions or examples to raise a metric.

No
Yes

Prompt tuning plus search across 11 hosted models

Fine-tuning / LoRA

Trains model weights or adapters.

Yes

LoRA, QLoRA, full FT, DPO, GRPO; vision and TTS

Partial

Text LoRA on Qwen3-8B / Llama-3.1-8B, early access; adapters can't be served yet; full fine-tuning and RLHF are previews

Red-teaming

Generates adversarial attacks (prompt injection, jailbreaks, data leaks) and scores them.

No
Yes

Prompt injection, jailbreak, data exfiltration, harmful content, PII leak

Model endpoints

Serves a model or tuned version behind an API you call in production.

Partial

Local OpenAI-compatible serving in Studio; no hosted endpoints

Yes

OpenAI-compatible, SSE streaming, tool calls forwarded

Production tracing

Logs and traces live traffic from your app.

No
Partial

Request logs, latency, errors and feedback for Kojen endpoints only; no tracing SDK

Official MCP server

An MCP server from the vendor that lets agents drive the product.

No

Local models can be used from MCP clients; no Unsloth MCP server

Yes

ai.kojen/kojen in the MCP Registry, OAuth sign-in

Self-host

You can run it on your own infrastructure.

Yes

Runs on your own GPUs

No

Hosted only

Training weights: Unsloth is ahead

Unsloth supports LoRA, QLoRA, full fine-tuning, pretraining, DPO and GRPO reinforcement learning, plus vision and text-to-speech models, and exports to GGUF and other formats. Its kernels cut memory enough to fine-tune mid-size models on a single consumer GPU, and the core is Apache-2.0.

Kojen's text LoRA trains a real adapter on Qwen3-8B or Llama-3.1-8B-Instruct from rows of input and expected output, and reports real train and eval loss. It is limited to accounts with GPU training enabled, runs one job at a time, and the adapter can't yet be served or evaluated in Kojen. Full fine-tuning and RLHF in Kojen are previews that train no weights.

Everything around the run: Kojen's case

Fine-tuning is the expensive lever. Kojen's Forge first tries the cheaper ones: tuning the system prompt against your test cases, and searching 11 hosted models for the one that holds up. Those run for real today with no GPUs. Evaluations report each rate with a confidence interval and pass^k, and red-team simulations attack the result before it ships behind an OpenAI-compatible endpoint.

When you do train, Kojen rents the GPU for you on Modal, Runpod, Vast.ai or Lambda, picking the cheapest in stock by default, prices credits from the GPU's live hourly price, locks that price when the run starts, and bills only the time used. With Unsloth you choose, rent and manage the GPU yourself, which is cheaper if you already own one.

Unsloth at a glance

An open-source library and local app (Unsloth Studio) for fast, memory-efficient fine-tuning and RL: LoRA, QLoRA, full fine-tuning, DPO and GRPO on a single GPU or more, with export to GGUF and local OpenAI-compatible serving.

Strengths

  • The fastest, least memory-hungry way to fine-tune on one GPU, with a huge community.
  • Covers everything from QLoRA to full fine-tuning and GRPO reinforcement learning.
  • Free and runs entirely on hardware you control.

Limits

  • You supply, run and pay for the GPUs.
  • No evaluations, red-teaming or hosted deployment; paid-tier prices aren't published.

Kojen at a glance

A hosted dojo for AI agents: LLM-judged evaluations with confidence intervals and pass@k / pass^k, red-team and stress simulations, prompt tuning and model search, LoRA training in early access, and OpenAI-compatible endpoints, all driveable over MCP and REST.

Strengths

  • Measures and improves in one loop: evaluations, red-teaming, prompt tuning, model search and deployment share the same project and test cases.
  • Honest statistics by default: every rate has a confidence interval, tuned prompts must beat the current one in a paired test, and pass@k / pass^k come from several samples per case.
  • Agent-native: the full loop is exposed over an MCP server and a REST API with an OpenAPI spec, with the same 24 tools as the in-app agent.

Limits

  • Young product: weight training is early access on 8B models, LoRA adapters can't be served yet, and full fine-tuning and RLHF are previews.
  • No tracing SDK for apps running elsewhere, so it is not a replacement for an observability platform.
  • Closed source and hosted only; no self-hosting.

When to choose Unsloth, and when to choose Kojen

Choose Unsloth if…

  • You have your own GPU, or are happy renting one, and want maximum control over training.
  • You need methods or models Kojen doesn't offer: QLoRA, full fine-tuning, GRPO, vision, larger models.
  • You want to export weights (GGUF and others) and run them locally.
  • You want a free, open-source tool.

Choose Kojen if…

  • You aren't sure fine-tuning is needed and want to try prompt tuning and model choice first.
  • You want evaluations with confidence intervals and red-teaming around every version.
  • You don't want to manage GPUs: Kojen rents the cheapest one in stock and bills by the second.
  • You want the winning version served behind an OpenAI-compatible endpoint.

The short version: For training weights, Unsloth is the better and cheaper tool. Kojen is for deciding whether training is worth it, testing whatever you end up with, and shipping it. Many teams will prototype in Kojen and move heavy training to Unsloth.

Frequently asked questions

Is Unsloth free?
The open-source library is free; its core is Apache-2.0 and some parts, such as the Studio UI, are AGPL-3.0. Unsloth lists Pro and Enterprise tiers for multi-GPU and multi-node training without public prices.
Which models can Kojen fine-tune?
Text LoRA trains on Qwen3-8B or Llama-3.1-8B-Instruct (Llama on Modal only), and audio LoRA on ACE-Step 1.5. Both are early access for accounts with GPU training enabled.
How much GPU memory does LoRA need?
Roughly 2 bytes per parameter for the frozen 16-bit weights, so about 16 GB for an 8B model before activations; QLoRA's 4-bit weights need about a quarter of that. Kojen's free VRAM calculator works it out for any size.
Can I serve an Unsloth model in Kojen?
No. Kojen evaluates, red-teams and serves its own hosted models and the versions you tune on them; it doesn't accept uploaded weights or call models hosted elsewhere.
Should I fine-tune or prompt-tune first?
Usually prompt-tune first: it is cheaper, faster and reversible. Fine-tune when you have hundreds of good examples and a measured gap prompting can't close.

Keep reading

Sources

Checked November 2026.

  1. [1]unslothai/unsloth. GitHub.
  2. [2]Unsloth documentation. Unsloth.
  3. [3]Pricing. Unsloth.
  4. [4]Kojen for Agents: REST API & MCP Server. Kojen.
  5. [5]llms.txt. Kojen.
  6. [6]Pricing. Kojen.

Try it on your own test cases.

Run an LLM-judged evaluation with confidence intervals, a red-team pass and a prompt-tuning run on your agent. Free to start, no card needed, then pay as you go.

Start training free