rockypod/ svelte-coder:latest

7 2 days ago

A Svelte 5 / SvelteKit 2 specialist coding model in three sizes. Free to use under MIT. Built on a homelab RTX 3090 Ti using cretrieval-augmented fine-tuning (RAFT).

4b 8b 14b 
ollama run rockypod/svelte-coder

Details

2 days ago

04fc5c27d630 · 9.0GB ·

model
qwen3
·
14.8B
·
Q4_K_M
template
{{- if .System }}<|im_start|>system {{ .System }}<|im_end|> {{ end }}<|im_start|>user {{ .Prompt }}<
system
You are SvelteCoder, an expert Svelte 5 / SvelteKit 2 coding assistant. Answer the question with com
params
{ "num_ctx": 8192, "num_predict": 1500, "repeat_penalty": 1.5, "temperature": 0.2 }

Readme

Svelte Coder

A Svelte 5 / SvelteKit 2 specialist coding model in three sizes. Free to use under MIT. Built by rockypod on a homelab RTX 3090 Ti using continuous retrieval-augmented fine-tuning (RAFT).

Quickstart

# 14B — recommended default
ollama run rockypod/svelte-coder

# 8B — for hardware where 14B doesn't fit
ollama run rockypod/svelte-coder:v0.9.0-8b

# 4B — edge hardware
ollama run rockypod/svelte-coder:v0.9.0-4b

Sizes

Tag Params Disk VRAM When to pick
:latest / :v0.9.0 14B 8.4 GB ~10 GB Recommended. Best benchmark scores.
:v0.9.0-8b 8B 5.0 GB ~6 GB Mid-tier GPUs, 16 GB VRAM laptops.
:v0.9.0-4b 4B 2.5 GB ~3 GB Edge devices, entry-level GPUs.

Benchmark (v0.9.0)

Variant 30Q spot 204Q in-scope (rescored)
14B (recommended) 100% 70.11%
8B 82.8% 74.68%
4B 79.3% 67.81%

The 30Q spot exam is the cleaner instrument — pick by 30Q, not 204Q. The 204Q has known keyword-matching grader artifacts. The “8B beats 14B on 204Q” inversion is grader noise, not real capability — see the GitHub repo for the full transparency write-up.

What it does well

  • Svelte 5 runes — $state, $derived, $effect, $props, $bindable
  • SvelteKit 2 — +page.server.ts actions, load(), redirects, error handling, route groups, hooks
  • Production patterns — accessibility (WCAG / ARIA), Playwright end-to-end tests, D3 visualizations, Svelte Flow diagrams, DaisyUI components

What it doesn’t do well

  • Svelte 4 → Svelte 5 conversion is weakest on the 4B and weak on the 8B. The pretrained Svelte 4 reflexes (export let, on:click, <slot>) leak through more often on smaller variants. Use the 14B for Svelte 4 conversion work.
  • Multi-step architectural reasoning is weaker on the 4B. Use the 8B or 14B for refactors.

Production parameters

The Modelfile ships with these defaults — they’re load-bearing:

PARAMETER temperature 0.2
PARAMETER num_ctx 8192
PARAMETER num_predict 1500
PARAMETER repeat_penalty 1.5

Use a chat client that respects the Modelfile (Ollama CLI, Continue, Zed, LM Studio with the included template). The OpenAI-compatible /v1 endpoint silently drops num_ctx — use /api/chat if you need to override context length over HTTP.

License & attribution

  • Fine-tuning work: MIT — see LICENSE
  • Base model: Qwen3-Coder-14B / Qwen3-8B / Qwen3-4B — Apache 2.0, © Alibaba Cloud
  • Teacher model (synthetic data): Qwen3-Coder-Next 80B — Apache 2.0, © Alibaba Cloud
  • See NOTICE for the full attribution.

Links