Models
All models available in the Llanite registry, with hardware requirements. Use llanite set <stack> model <model> to swap the model for any stack.
Qwen3
| Model | Params | Disk | RAM | Description |
|---|---|---|---|---|
qwen3:4b | 4B | ~2.6 GB | 4 GB min · 8 GB rec | Smallest Qwen3 variant. Quick responses, low RAM footprint. |
qwen3:8b | 8B | ~5.2 GB | 8 GB min · 16 GB rec | Balanced general-purpose model. Good code and reasoning quality. |
qwen3:14b | 14B | ~9.3 GB | 16 GB min · 32 GB rec | Strong coding and reasoning. Includes extended thinking mode. |
qwen3:32b | 32B | ~19.9 GB | 24 GB min · 48 GB rec | Top-tier Qwen3 for complex tasks. Requires 32 GB+ RAM. |
Qwen 3.5
| Model | Params | Disk | RAM | Description |
|---|---|---|---|---|
qwen3.5:4b | 4B | ~2.8 GB | 4 GB min · 8 GB rec | Compact Qwen 3.5 with 262K context. Fast, low memory. |
qwen3.5:9b | 9B | ~5.8 GB | 8 GB min · 16 GB rec | Strong mid-tier. 262K context, hybrid thinking, 201 languages. |
qwen3.5:27b | 27B | ~17 GB | 24 GB min · 48 GB rec | High-quality dense model. Near-frontier performance at 27B scale. |
Qwen2.5 Coder
| Model | Params | Disk | RAM | Description |
|---|---|---|---|---|
qwen2.5-coder:7b | 7B | ~4.7 GB | 8 GB min · 16 GB rec | Coding-specialist model. Default for local-coder stack. |
qwen2.5-coder:14b | 14B | ~9.3 GB | 16 GB min · 32 GB rec | Larger coding model for complex refactors and architecture tasks. |
qwen2.5-coder:32b | 32B | ~19.9 GB | 24 GB min · 48 GB rec | Best-in-class open coding model. Fits 32 GB RAM at Q4. |
Meta Llama
| Model | Params | Disk | RAM | Description |
|---|---|---|---|---|
llama3.1:8b | 8B | ~4.9 GB | 8 GB min · 16 GB rec | Meta's open general-purpose model. Great for private chat stacks. |
llama3.1:70b | 70B | ~43 GB | 48 GB min · 80 GB rec | Frontier-class open model. Requires significant RAM. |
Google Gemma
| Model | Params | Disk | RAM | Description |
|---|---|---|---|---|
gemma3:4b | 4B | ~3 GB | 6 GB min · 12 GB rec | Fast and lightweight. 128K context, fits on low-RAM machines. |
gemma3:12b | 12B | ~8 GB | 12 GB min · 24 GB rec | Strong general performance with 128K context window. |
gemma3:27b | 27B | ~17 GB | 32 GB min · 48 GB rec | Top Gemma 3 performance for 32 GB machines. |
gemma4:9b | 9B | ~6 GB | 12 GB min · 24 GB rec | Latest Gemma. Vision-capable, strong benchmarks at 9B scale. |
Mistral
| Model | Params | Disk | RAM | Description |
|---|---|---|---|---|
mistral:7b | 7B | ~4.1 GB | 8 GB min · 16 GB rec | Fast, capable general model. Efficient for chat and Q&A. |
mixtral:8x7b | 56B | ~26.1 GB | 32 GB min · 48 GB rec | Mixture-of-experts model. Strong quality with selective activation. |
DeepSeek
| Model | Params | Disk | RAM | Description |
|---|---|---|---|---|
deepseek-coder-v2:16b | 16B | ~9.1 GB | 16 GB min · 32 GB rec | Strong coding model with reasoning capabilities. |
StarCoder
| Model | Params | Disk | RAM | Description |
|---|---|---|---|---|
starcoder2:7b | 7B | ~4.3 GB | 8 GB min · 16 GB rec | Code completion specialist. Trained on 600+ programming languages. |
starcoder2:15b | 15B | ~9.1 GB | 16 GB min · 32 GB rec | Larger code completion model. Better at multi-file context. |
Not sure which model to use?
Run llanite doctor <stack> to see which models fit your machine, or check the Model Fit docs for an explanation of the fit tiers.
For agentic and tool-calling workloads, the Berkeley Gorilla Leaderboard benchmarks models specifically on function-calling accuracy — useful if your stack relies heavily on tools.