Catalog entries are options. Proof is earned per workflow.

Farka separates license review, hardware fit, live training, paired evaluation, tuned inference, export, billing, and teardown. A model is not called verified because it appeared in a list or accepted a quote.

Larger model options

Open models for reasoning, code, and agents. The training setup, capacity, and price depend on your workload. These catalog entries carry no completed-training or quality claim.

DeepSeek V4.1 Flash

Catalog · 552B + 196B

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

DeepSeek V4 Flash

Catalog · 284B

13B active parameters · Coding and agents

Qwen3.6

Catalog · 27B

Reasoning, code and agents

Gemma 4

Catalog · 31B

Images, text and assistants

Selected verified configurations

Exact model, training method, passes over the data, accelerator-memory class, and delivery combinations from collected runs. Farka uses different training approaches depending on the run: a light tune adds a small set of new weights to the base model, while a full tune updates every weight. Provider-native hardware remains attached to the audited run record.

01

Gemma 4 E4B

Light tune · 3 passes

25/30 → 30/3016 GB accelerator
02

Qwen2.5 Coder 7B

Light tune · 1 pass

15/30 → 30/3016 GB accelerator
03

SmolLM2 360M

Full tune · 3 passes

21/60 → 58/6016 GB accelerator
04

Granite 4.1 8B

Light tune · 3 passes

Workflow verified16 GB accelerator
05

DeepSeek R1 Distill Qwen 7B

Light tune · 3 passes

Workflow verified16 GB accelerator

Results are configuration-specific. They do not imply that every dataset, model size, or training recipe will improve.

Self-service

Curated, license-reviewed models verified in production. Farka selects the cheapest verified capacity with sufficient memory.

Enterprise capacity

Models outside the verified set, multi-accelerator, private-network, or customer-cloud workflows begin with a capacity and security review. Explore custom training.

License review

Commercial use and weight modification must be supported by publisher evidence. Unknown, non-commercial, or incompatible licenses fail closed.

Check your model and dataset against the proven boundary.