# Local LLM Inventory

Last verified: 2026-06-28  
Host: `sol.system42.one` local archive

This page records the local language-model assets currently present on disk under `/home/david/.cache/models`, the local runtimes available on this machine, and whether each model family should be considered open-weight.

## Local runtimes

- `ollama` at `~/.local/bin/ollama`
- `llama-server` at `~/.local/bin/llama-server`
- `llama-cli` at `~/.local/bin/llama-cli`

Ollama is installed, but `ollama list` was empty at the time of verification. The local model inventory below is therefore based on GGUF files present on disk rather than Ollama-managed registrations.

## Installed model files

### 1. DeepSeek-R1-Distill-Qwen-7B

- File: `/home/david/.cache/models/DeepSeek-R1-Distill-Qwen-7B-Q4_K_M.gguf`
- Size: `4.4G`
- Format: GGUF
- Likely role: local reasoning / general chat

### 2. Llama-3.2-1B-Instruct

- File: `/home/david/.cache/models/Llama-3.2-1B-Instruct-Q4_K_M.gguf`
- Size: `771M`
- Format: GGUF
- Likely role: lightweight local chat, lower memory footprint

### 3. Qwen2.5-7B-Instruct

- File: `/home/david/.cache/models/Qwen2.5-7B-Instruct-Q4_K_M.gguf`
- Size: `163M`
- Format: GGUF
- Status note: the file has a valid `GGUF` header, but the size is much smaller than expected for a 7B Q4 model and should be treated as suspicious or incomplete until load-tested

### 4. Qwen3VL-8B-Thinking

- File: `/home/david/.cache/models/Qwen3VL-8B-Thinking-Q4_K_M.gguf`
- Size: `4.7G`
- Format: GGUF
- Likely role: multimodal reasoning when paired with the projection weights listed below

### 5. gemma-3-4b-it

- File: `/home/david/.cache/models/gemma-3-4b-it-Q4_K_M.gguf`
- Size: `2.4G`
- Format: GGUF
- Likely role: compact local instruction model

## Multimodal projection files

These are not separate LLMs. They are auxiliary weights used for multimodal vision-language pipelines.

- `/home/david/.cache/models/mmproj-Qwen3VL-8B-Thinking-F16.gguf` (`1.1G`)
- `/home/david/.cache/models/mmproj-model-f16.gguf` (`812M`)

## Open-weight status

### Open-weight with standard permissive licensing

#### DeepSeek-R1-Distill-Qwen-7B

- Classification: open-weight
- License position: published with available weights; model card indicates MIT for the DeepSeek release
- Source: <https://huggingface.co/deepseek-ai/DeepSeek-R1-Distill-Qwen-7B>

#### Qwen2.5-7B-Instruct

- Classification: open-weight
- License position: Apache 2.0
- Source: <https://huggingface.co/Qwen/Qwen2.5-7B-Instruct>

#### Qwen3-VL-8B-Thinking

- Classification: open-weight
- License position: Apache 2.0 on the official GGUF release
- Source: <https://huggingface.co/Qwen/Qwen3-VL-8B-Thinking-GGUF>

### Open-weight, but under custom or non-OSI terms

#### gemma-3-4b-it

- Classification: open-weight
- License position: Google publishes Gemma as open weights, but under Gemma-specific terms rather than a standard OSI open-source license
- Sources:
  - <https://ai.google.dev/gemma/docs/core/model_card_3>
  - <https://ai.google.dev/gemma/terms>

#### Llama-3.2-1B-Instruct

- Classification: publicly available weights
- License position: Meta `Llama 3.2 Community License`, not a standard open-source license
- Source: <https://huggingface.co/meta-llama/Llama-3.2-1B-Instruct>

## Summary

### Models on this machine that are clearly open-weight and permissively licensed

- `DeepSeek-R1-Distill-Qwen-7B`
- `Qwen2.5-7B-Instruct`
- `Qwen3VL-8B-Thinking`

### Models on this machine with public/open weights but custom licensing

- `gemma-3-4b-it`
- `Llama-3.2-1B-Instruct`

## Notes

- The inventory reflects files physically present on disk on 2026-06-28.
- Runtime availability does not guarantee that every file is intact or currently tuned for usable inference.
- The `Qwen2.5-7B-Instruct` GGUF should be considered the main candidate for re-download or integrity verification because its size is unexpectedly small.