Training
Teach a model a behaviour, format or domain style.
A LoRA or QLoRA adapter is trained on your instruction dataset. Instead of rewriting every weight, a small low-rank matrix is learned alongside the frozen base — which is why a 14B model can be customised on a single consumer GPU in an afternoon.
- The model must answer in a consistent structure every time
- You need a tone or house style it will not hold through prompting
- You have several hundred or more good input/output pairs
- A task-specific skill needs to survive without a long system prompt
- Installing facts that change — retrain cost is real
- Fixing behaviour you have not yet tried to prompt for
Private company agent
An internal assistant that knows your documents, answers in your format, can act on internal systems, and never sends a byte outside the building.
Support responder
Trained on resolved tickets for structure and tone, retrieving live product documentation for the facts.
Qwen2.5 7B Instruct
7.6B · 32k · Apache 2.0
The default starting point for local work on 8–12 GB cards. Strong instruction following and reliable tool-call formatting for its size.
Qwen2.5 14B Instruct
14.8B · 32k · Apache 2.0
The sweet spot for 24 GB cards. Meaningfully stronger reasoning than 7B while still fine-tunable locally with QLoRA.
Qwen2.5 32B Instruct
32.8B · 32k · Apache 2.0
Approaches 70B quality at half the memory. Runs on a single 24 GB card at Q4_K_M with a modest context window.
Qwen2.5 72B Instruct
72.7B · 32k · Qwen License
Frontier-adjacent open weights. Needs a workstation, a multi-GPU rig or a large unified-memory Mac.
Qwen2.5 Coder 7B Instruct
7.6B · 32k · Apache 2.0
The practical local copilot. Supports fill-in-the-middle, so it works as an inline completion model rather than only a chat assistant.
Qwen2.5 Coder 32B Instruct
32.8B · 32k · Apache 2.0
The strongest open code model that still fits one 24 GB card. The reason a lot of people buy a 4090.