---
license: mit
language:
- en
- code
library_name: gguf
pipeline_tag: text-generation
tags:
- svelte
- sveltekit
- svelte-5
- runes
- code-generation
- gguf
- qwen3
- lora
base_model: Qwen/Qwen3-4B
base_model_relation: finetune
---

# Svelte Coder 4B (v0.9.0)

A Svelte 5 / SvelteKit 2 specialist coding model — **4B variant**.
Free to use under MIT. Built by [rockypod](https://rockypod.com) on a
homelab RTX 3090 Ti using continuous retrieval-augmented fine-tuning
(RAFT) and a correction-stream methodology.

This is the **4B variant** for edge hardware. For best benchmark
results, use the [14B variant](https://huggingface.co/rockypod/svelte-coder)
when the hardware allows.

**[14B (recommended)](https://huggingface.co/rockypod/svelte-coder)** ·
**[8B (mid-tier)](https://huggingface.co/rockypod/svelte-coder-8b)** ·
**[GitHub — exam, integration guides, transparency](https://github.com/rockypod/svelte-coder)**

## Benchmark

| Instrument | Score |
|---|---|
| 30Q spot exam | **79.3%** (34.5 / 43.5 weighted) |
| 204Q in-scope (rescored) | 67.81% (131 / 190 raw) |

For comparison, the 14B variant scores 100% / 70.11% on the same
instruments. The 4B trades capability for accessibility on edge
hardware. See the
[main README](https://huggingface.co/rockypod/svelte-coder/blob/main/README.md)
for the full two-exams discussion.

## Hardware requirements

- **VRAM:** ~3 GB (Q4_K_M GGUF), runs on entry-level GPUs and
  Apple Silicon devices with limited memory
- **Context length:** 8192
- **Recommended use case:** edge hardware where even the 8B variant
  is too large; iOS/Android via llama.cpp; constrained Linux servers

## Files

- `svelte-coder-v0.9.0-4b-q4_k_m.gguf` — 4-bit quantized weights (~3 GB)

## Usage

### Ollama

```bash
ollama pull rockypod/svelte-coder:4b
ollama run rockypod/svelte-coder:4b "Write a Svelte 5 counter with $state and $derived"
```

### LM Studio / llama.cpp

Download `svelte-coder-v0.9.0-4b-q4_k_m.gguf` and load with the
production parameters: temperature 0.2, num_ctx 8192, num_predict 1500,
repeat_penalty 1.5. Use the ChatML template:

```
<|im_start|>system
You are SvelteCoder, an expert Svelte 5 / SvelteKit 2 coding assistant. Answer the question with complete, production-quality code.<|im_end|>
<|im_start|>user
Your question<|im_end|>
<|im_start|>assistant
<think>
```

## Limitations specific to the 4B

- **Svelte 4 echo trap is most frequent on this variant.** The 4B has
  the least capacity to override Qwen3-4B's pretrained Svelte 4
  reflexes. T1 (Runes) and T4 (WCAG/ARIA) fix-this-snippet questions
  show 1/3 pass rates on the 30Q spot. Review output for `export let`,
  `on:click`, `<slot>` patterns when modernizing Svelte 4 code, and
  prefer the 8B or 14B if Svelte 4 conversion is a primary use case.
- **Hard reasoning weaker than larger variants.** T6 multi-step
  refactors are weaker on the 4B than on the 8B or 14B. Use the larger
  variants for architectural decisions or complex refactors.
- All other limitations from the [main README](https://huggingface.co/rockypod/svelte-coder/blob/main/README.md)
  apply.

## Apple Silicon note

MLX builds for Apple Silicon are not included in v0.9.0 for the 8B and
4B variants. Apple Silicon users are recommended to use the 14B variant,
which includes MLX 4-bit weights.

## License & Attribution

**Fine-tuning work licensed under the MIT License** — see
[LICENSE](LICENSE) in the GitHub repo.

**Base model and teacher model are licensed under Apache 2.0** — see
[LICENSE-APACHE](LICENSE-APACHE) and [NOTICE](NOTICE):

- Base: [Qwen3-4B](https://huggingface.co/Qwen/Qwen3-4B) — © Alibaba Cloud
- Teacher: [Qwen3-Coder-Next 80B](https://huggingface.co/Qwen/Qwen3-Coder-Next) — © Alibaba Cloud

The 4B Svelte Coder weights are a derivative work of Qwen3-4B,
fine-tuned via LoRA adapters on the v1.5 Svelte 5 / SvelteKit 2
specialist dataset (1,508 entries).