A model trained on your work. Owned by your company.

Farka fine-tunes open AI models on your examples, measures the improvement on data the model never saw, and delivers a model you can run on infrastructure you control.

Open modelYour examplesYour specialist
15 30
correct of 30 unseen examplesQwen2.5 Coder 7B
0 30
correct of 30 unseen examplesGranite 4.1 8B
Free
quote before any GPU starts

Recorded on internal classification tests. See how each run was measured

See one run, from examples to a model you keep.

Upload examples, review the price, train, then compare the result with the untouched model on the same unseen data.

  1. Show it the job.

    robotics-process.jsonl

    • Photograph weld seam 7 for porosity review (cell 1)1. Move to the seam start pose …12345check
    • Assemble kit 19 from bins 3, 7 and 11 (cell 2)1. Read the kit manifest …12345check
    • Deliver a cart from ward 2 to the sterile store (cell 3)1. Confirm the cart is latched …12345check
    • Photograph weld seam 7 for porosity review (cell 6)1. Move to the seam start pose …12345check
    • Assemble kit 19 from bins 3, 7 and 11 (cell 7)1. Read the kit manifest …12345check
    • Deliver a cart from ward 2 to the sterile store (cell 8)1. Confirm the cart is latched …12345check
  2. Farka checks your examples.

    • format
    • quality
    • duplicates

    Set aside. Kept out of training.

  3. See the price before any GPU starts.

    1. Quote
    2. You review the price
    3. Training starts

    No GPU is running yet

  4. Farka trains it.

    Base model

    Trained on your examples

  5. Compare on examples it never saw.

    30 examples kept out of training

    • Already correct
    • Improved
    • Missed

    Before

    0/30

    After

    30/30

    Granite 4.1 8B. Recorded internal classification test.

  6. Keep the model.

    Your modelWeights export
    • adapter_model.safetensors
    • adapter_config.json
    • tokenizer.json
    • eval-report.json

    Light-tune weights export

Animated overview of a Farka run. Scores are from a recorded internal test.

One model for the work that matters.

A general model is good at a bit of everything. A specialist learns the multi-step work your team repeats every day: procedures, tool-using workflows, long documents and images, answered in exactly the format your systems expect.

TeamJobGoes inComes out
OperationsWrite procedures that respect safety limitsPhotograph weld seam 7 for porosity review (cell 1)Ordered steps and a measurable success check
SupportRun a multi-step workflow with your toolsA customer case and each tool resultThe next action at every step, within your policy
ComplianceReview long documents against a checklistA full contract and your review rulesEvery finding in your schema, tied to its clause
EngineeringRaise accuracy where answers can be checkedProblems with a verifiable answerAnswers trained against an automatic check
Field serviceRead inspection photosA photo and an instructionFindings in your exact format

Typical first projects, not customer results. The operations task is a real row from our robotics-process.jsonl test data.

When fine-tuning is the wrong answer, we will say so.

Good fit

  • A repeated job, single-step or multi-step, where you can say what a good answer is
  • Examples of the task and the answer you want
  • A strong prompted model that is close, but not reliable enough
  • A need to run the model on infrastructure you control

Not a fit

  • Answers that depend on live data. Use retrieval for that
  • Work where nobody can say what a correct answer is
  • A task that a simple rule or a good prompt already solves
  • A need for Farka to host your model. Hosted inference is not generally available

Three steps, each one reviewable.

You see the data checks, the price, and the evaluation before anything is delivered.

  1. 1

    Show it the job

    Upload examples of the task and the answers you want, chat files made for OpenAI or Foundry, or your manuals. Farka checks format, quality and duplicates before training.

  2. 2

    Farka trains it

    Pick a supported model and review the quote. Train from worked examples, multi-step workflows, images, or outcomes that can be checked automatically.

  3. 3

    Check it, then keep it

    Test the result on examples kept out of training before delivery. Download the model files or a package ready to run on your own infrastructure.

Improvement measured on unseen examples.

Compare before and after on the same examples kept out of training. These are internal classification tests, not a guarantee of results on your data.

Qwen2.5 Coder 7B

15/3030/30

Correct answers before and after training

View run details for Qwen2.5 Coder 7B

Light tune · 1 pass · 16 GB accelerator

270 training examples and 30 test examples. Internal classification task.

Base and trained model tested on the same 30 examples kept out of training.

Read the evidence

Granite 4.1 8B

0/3030/30

Correct answers before and after training

View run details for Granite 4.1 8B

Light tune · 3 passes · 16 GB accelerator

270 training examples and 30 test examples. Internal classification task.

Base and trained model tested on the same 30 examples kept out of training.

Read the evidence
See how each run was measured

Start from a proven open model.

From 7B-class specialists to frontier-scale models for reasoning, code and agents. Larger-model runs are planned per workload, with capacity confirmed before a run is scheduled.

Larger models

View capacity details

Larger-model runs use workload-specific GPU setups. Availability, hardware fit, and price depend on the complete training configuration and current capacity.

  • CatalogDeepSeek V4.1 Flash552B + 196B

    Backbone + Engram memory · 8B active prefill / 16B decode · Vision, text and agents

    View model details
  • CatalogDeepSeek V4 Flash284B

    13B active parameters · Coding and agents

    View model details
  • CatalogQwen3.627B

    Reasoning, code and agents

  • CatalogGemma 431B

    Images, text and assistants

Measured examples

View hardware details

16 GB accelerator memory class. Fit depends on the full training configuration and available capacity.

  • MeasuredGemma 4 E4B8B

    Before / after: 25/30 → 30/30

  • MeasuredGranite 4.18B

    Before / after: 0/30 → 30/30

  • MeasuredQwen2.5 Coder7B

    Before / after: 15/30 → 30/30

Larger models are catalog options, not measured results. The before-and-after scores come from recorded internal tests. Every new run must pass its own quality checks.

See all models and boundaries

Built for your security review.

Three things a company needs to know before any data moves.

Your data stays where you pin it

Pin a training run to the EU or the US and it runs nowhere else. A run that cannot be placed in-region fails rather than moving.

EU
Training, datasets, and artifacts stay on EU/EEA infrastructure.
US
Training, datasets, and artifacts stay on US infrastructure.
Auto
Cheapest capacity anywhere. No residency commitment, and we say so.
Explore data control Ask about EU hosting

The trained model is yours

Download standard safetensors files that load in common open-weight tools, or a package ready to run on infrastructure you control.

Read the docs

Common questions.

What does Farka do?
Farka trains specialized models from examples you provide or outcomes that can be checked automatically. Compare the model before and after training on examples kept out of training, then download it once the run passes its quality checks.
How is Farka billed?
By the GPU-second, against prepaid credits. Every run is quoted before any GPU starts, and credits reserved at launch are released if the run fails or is cancelled before settlement.
Who owns the fine-tuned model?
You do. Farka produces adapter weights, deployment bundles, and SDK artifacts for runs you paid for, and does not use your data to train general-purpose models.
Can a training run be kept in one country?
Yes. A run can be pinned to the European Union or the United States. Pinning is enforced rather than advisory: a run that cannot be placed in region fails instead of relocating.
What does the name Farka mean?
Farka is Albanian for “the forge”. A forge turns raw metal into a tool made for one job, and Farka does the same with open models: it shapes a general model into a specialist for your task.

Talk to an engineer before you commit.

Bring one workflow. Your request goes directly to Edi Hasaj, Farka’s founder, and you get an honest view on whether fine-tuning is the right fix.

What we can scope with you

  • Custom model or objective
  • Private or customer cloud
  • Tailored quality gates
  • Production handoff
Explore custom training
Start a review

Tell us what you need to train.

We’ll review the fit and reply to your work email.

Submissions go directly to Farka and are not stored.

Prefer to start on your own?

Create a workspace, upload a dataset, and quote your first run before any GPU starts.

Create your workspace