A Rust / Dioxus 0.7 specialist fine-tuned via RAFT on Qwen3 (3 sizes), optimized for Tailwind v4 + WCAG 2.2 AAA accessibility and dioxus-primitives discipline. Built with CRAFT on a homelab RTX 3090 Ti. No cloud GPUs.

Neotoi Coder

A Rust / Dioxus 0.7 specialist fine-tuned via RAFT on Qwen3 (3 sizes), optimized for Tailwind v4 + WCAG 2.2 AAA accessi bility and dioxus-primitives discipline.

Variants

Tag	Base	Params	Q4_K_M size	Spec-exam score	Pick when…
`rockypod/neotoi-coder:8b`	Qwen3-8B	8.2B	4.68 GB	¹⁰³⁄₁₀₃ (100.00%)	You want the ceiling fundamentals sc
ore
`rockypod/neotoi-coder:4b`	Qwen3-4B	4.0B	2.33 GB	¹⁰²⁄₁₀₃ (99.31%)	Disk / RAM constrained or you want ~40% f
aster generation
`rockypod/neotoi-coder:15b`	Qwen3-Coder-14B	14.8B	8.4 GB	137.0/144.5 (94.81%)	Legacy 14B, broadest coverage

Both :8b and :4b clear the 90% publication bar and the 95% release bar with all per-tier floors (82% / 88%) PASS.

Quick start

ollama pull rockypod/neotoi-coder:8b
ollama run rockypod/neotoi-coder:8b

Or use the smaller 4B:

ollama pull rockypod/neotoi-coder:4b

Recommended runtime parameters (already baked into the Modelfile):

Parameter	Value
`num_ctx`	8192
`temperature`	0.2
`num_predict`	2000
`repeat_penalty`	1.1
`stop`	`<\|im_end\|>`

The Qwen3 ChatML template prefills <think> on the assistant turn so the model emits its native chain-of-thought block before answering.

What it knows

Dioxus 0.7 RSX brace syntax (never function-call style)
use_signal, use_resource with the canonical three-arm match
r#for on labels only, never inputs
WCAG 2.2 AAA: aria_labelledby, aria_describedby, live regions, role="alert", role="dialog"
dioxus-primitives — no manual ARIA on managed components
styles!() macro and native CSS modules
Tailwind v4 utility classes and semantic tokens
GlobalSignal patterns (LANG / THEME), EN/VI i18n, dark-mode toggling via document::eval
Dioxus 0.7.4 APIs: WritableResultExt, WebSocket Stream+Sink, server-fn extractors

What it does not know

Playwright / E2E testing (out of scope)
Non-Dioxus web frameworks (use rockypod/svcoder for SvelteKit)
Backends or databases beyond what server functions cover

Spec-exam scorecards

Re-graded 2026-04-26 with the patched run_grade_v31.py (Q87 accepts LANG() / THEME() GlobalSignal accessor calls) .

Tier	Name (max wt / Q)	8B raw	8B wtd	4B raw	4B wtd	14B v3.1 wtd
T1	Fundamentals (12 / 12)	12	12.0	11	11.0	12.0
T2	RSX Syntax (12 / 12)	12	12.0	12	12.0	10.0
T3	Signal Hygiene (12 / 12)	12	12.0	12	12.0	11.0
T4	WCAG / ARIA (21 / 14)	14	21.0	14	21.0	16.5
T5	use_resource (12 / 8)	8	12.0	8	12.0	12.0
T6	Hard Reasoning (20 / 10)	10	20.0	10	20.0	20.0
T7	Primitives + CSS (18 / 12)	12	18.0	12	18.0	18.0
T8	GlobalSignal / i18n (12 / 8)	8	12.0	8	12.0	12.0
T9	Static Navigator (9 / 6)	6	9.0	6	9.0	9.0
T10	Dioxus 0.7.4 (12 / 6)	6	12.0	6	12.0	12.0
T11	Server Functions (4.5 / 3)	3	4.5	3	4.5	4.5
Total	(144.5 / 103)	103	144.5	102	143.5	137.0
Percent			100.00%		99.31%	94.81%

Tier floors (82% on weight 1.0/1.5, 88% on weight 2.0): all PASS for all three variants.

The only miss across the new releases is the 4B’s Q8 (T1 RSX conversion) — generation truncated mid-<think> block.

Transparency

HuggingFace weights (per variant):
- 8B → https://huggingface.co/rockypod/neotoi-coder-8b
- 4B → https://huggingface.co/rockypod/neotoi-coder-4b
- 14B → https://huggingface.co/rockypod/neotoi-coder (family hub)
Exam runner, patched grader, per-question results: https://github.com/rockypod/neotoi-coder
The training dataset itself is not redistributed (Tailwind v4 reference material is treated as a competence input , not a shipped artifact).

License

NeoToi Coder Community License v1.0. Commercial use of model outputs permitted. Weight redistribution prohibited. Menta l health deployment requires written permission.