jarvis-qwen3-4b is an open-weight model from Coen Hao, released under Apache License 2.0. Its published files total 2.5 GB.
The local brain of Jarvis, a Mac assistant that lives in the notch. Qwen3 4B, fine-tuned with LoRA (MLX) and exported as GGUF Q4KM for llama.cpp. It is a function bot: it doesn't answer from its own knowledge.
Model Card
By Coen Hao, published under apache-2.0, revision 4175a0d5a2db.
The local brain of Jarvis, a Mac assistant that lives in the notch. Qwen3 4B, fine-tuned with LoRA (MLX) and exported as GGUF Q4KM for llama.cpp. It is a function bot: it doesn't answer from its own knowledge. It does four jobs through Jarvis's tools and hands everything else to a cloud model by replying ESCALATE: A tool call is one line of JSON, e.g. {"tool": "timerstart", "arguments": {"duration": "10 minutes"}}. Facts, writing, news, calendar and Notion questions, and anything that changes the user's data get ESCALATE. In Jarvis: put Jarvis-4B-Q4KM.gguf in ~/Library/Application Support/Jarvis/Models/, then pick Jarvis 4B (trained) in Settings › Connections › Model. It expects Jarvis's…
Read Coen Hao's full model card
The local brain of Jarvis, a Mac assistant that lives in the notch. Qwen3 4B, fine-tuned with LoRA (MLX) and exported as GGUF Q4_K_M for llama.cpp.
It is a function bot: it doesn't answer from its own knowledge. It does four jobs through Jarvis's tools and hands
everything else to a cloud model by replying ESCALATE:
| Job | Tool |
|---|---|
| Timers by length ("10 minutes on the clock") | timer_start (length in words) |
| Opening apps, loosely worded ("pull up spotify") | mac_open_app |
| Quick Dock on or off | mac_quick_dock |
| Arithmetic and units | math_calculate |
A tool call is one line of JSON, e.g. {"tool": "timer_start", "arguments": {"duration": "10 minutes"}}. Facts,
writing, news, calendar and Notion questions, and anything that changes the user's data get ESCALATE.
Use
In Jarvis: put Jarvis-4B-Q4_K_M.gguf in ~/Library/Application Support/Jarvis/Models/, then pick
Jarvis 4B (trained) in Settings › Connections › Model. It expects Jarvis's own system prompt (Persona.localTools
plus the tool list), with thinking off (/no_think).
Training
- Base: Qwen3-4B (MLX 4-bit), LoRA on 16 layers, learning rate 1e-4, 2,000 of 3,000 iterations (the better checkpoint).
- Data: 3,061 generated conversations (version v15): Jarvis's four jobs about 60%, handing on about 40%, including near-misses that look like its jobs but aren't ("what time is it in Tokyo"). No real user data.
Evaluation
Held-out wordings, 34 kinds × 10 cases, two seeds, replaying Jarvis's tool loop:
| Model | Seed 99 | Seed 123 | Confidently wrong |
|---|---|---|---|
| This model (4B, v15 @ 2000) | 340/340 | 340/340 | 0 |
| Previous Jarvis 8B (v12) | 230/340 | 232/340 | 96 / 94 |
The eval comes from the same generator as the training data (different wordings), so real use will score lower.
Limits
Trained for Jarvis's prompt and tools only; as a general chat model it will mostly answer ESCALATE.
Identity and Version
- Repository
- refriderator/jarvis-qwen3-4b
- Publisher
- Coen Hao
- Task
- Not stated by the source
- Modality
- Other
- Library
- Not stated by the source
- Parameters
- Not stated by the source
- Languages
- en
- Revision
- 4175a0d5a2db91fcfa714306d24d5787f8891d6c
- First published
- 2026-10-18
- Last updated
- 2026-10-18
Files and Weights
3 files, 2.5 GB in total. The weights are 1 file totalling 2.5 GB in gguf.
Every file
| File | Type | Size | SHA-256 |
|---|---|---|---|
| Jarvis-4B-Q4_K_M.gguf | Weights | 2.5 GB | 98b52288d1ec |
| README.md | Documentation | 2.1 KB | — |
| .gitattributes | Repository | 1.6 KB | — |
License and Download
- License
- apache-2.0
- Access
- Open weights, no gate
- Download size
- 2.5 GB
Released by Coen Hao through its official repository on Hugging Face. Read the license.
Built From
- Derived from Qwen/Qwen3-4B
- Quantized from Qwen/Qwen3-4B
Memory Requirements
| Precision | Weights in memory |
|---|---|
| As published | 2.5 GB |
Weights only, from the published parameter count; the key-value cache and runtime add to this.
Questions About jarvis-qwen3-4b
Can I use jarvis-qwen3-4b commercially?
Yes. jarvis-qwen3-4b is released under Apache License 2.0. The Apache License 2.0 is a permissive open-source license. It permits commercial use, modification and redistribution. It requires keeping the license and copyright notices and any NOTICE file, stating significant changes, and it includes an express patent grant from contributors.