Neotoi Coder v3.2 — 8B
A Rust / Dioxus 0.7 specialist fine-tuned from Qwen3-8B (8.2B parameters, 6.95B non-embedding) using RAFT (Retrieval-Augmented Fine-Tuning). Optimized for production-quality Dioxus 0.7 components with Tailwind v4 and WCAG 2.2 AAA accessibility.
This is the 8B variant of the v3.2 release. Companion repos: 4B (rockypod/neotoi-coder-4b) · 15B family hub (rockypod/neotoi-coder)
v3.2 Exam Results — 114Q Dioxus 0.7 Spec Exam
160.0 / 164.0 weighted | 111 / 114 raw | 97.56%
| Tier | Name | Cnt | Raw | Wtd | /Max | Rate | Floor | Status |
|---|---|---|---|---|---|---|---|---|
| T1 | Fundamentals | 12 | 12 | 12.0 | 12.0 | 100.0% | 82% | ✅ |
| T2 | RSX Syntax | 12 | 11 | 11.0 | 12.0 | 91.7% | 82% | ✅ |
| T3 | Signal Hygiene | 12 | 12 | 12.0 | 12.0 | 100.0% | 82% | ✅ |
| T4 | WCAG / ARIA | 15 | 15 | 22.5 | 22.5 | 100.0% | 82% | ✅ |
| T5 | use_resource | 8 | 8 | 12.0 | 12.0 | 100.0% | 82% | ✅ |
| T6 | Hard Reasoning | 10 | 10 | 20.0 | 20.0 | 100.0% | 88% | ✅ |
| T7 | Primitives + CSS | 13 | 13 | 19.5 | 19.5 | 100.0% | 82% | ✅ |
| T8 | GlobalSignal / i18n | 8 | 7 | 10.5 | 12.0 | 87.5% | 82% | ✅ |
| T9 | Static Navigator | 6 | 6 | 9.0 | 9.0 | 100.0% | 82% | ✅ |
| T10 | Dioxus 0.7.4 | 6 | 6 | 12.0 | 12.0 | 100.0% | 88% | ✅ |
| T11 | Server Functions | 4 | 4 | 6.0 | 6.0 | 100.0% | 82% | ✅ |
| T12 | Format Compliance (NEW) | 6 | 6 | 12.0 | 12.0 | 100.0% | 88% | ✅ |
| T13 | SyncStore (NEW) | 2 | 1 | 1.5 | 3.0 | 50.0% | 82% | ⚠️ |
| Total | 114 | 111 | 160.0 | 164.0 | 97.56% | — | — |
- Publication bar (90%): PASS
- Release bar (95%): PASS
- Tier floors: FAIL (T13 only — structural: 2 questions, single real miss = 50%)
3 misses: q022 (T2, rsx! macro), q087 (T8, use_signal), q113 (T13, tokio::spawn)
The T13 SyncStore floor failure is structural — only 2 questions in the tier means any single real miss equals a floor failure regardless of difficulty.
v3.2 vs v3.1 (8B)
| Metric | v3.1 8B | v3.2 8B |
|---|---|---|
| Score | 144.5/144.5 (100.0%) | 160.0/164.0 (97.56%) |
| Exam | 103Q, max 144.5, 11 tiers | 114Q, max 164.0, 13 tiers |
| T4 WCAG / ARIA | 100.0% | 100.0% ✅ |
| T7 Primitives + CSS | 100.0% | 100.0% ✅ (15B was 92.3%) |
| T12 Format Compliance | — | 100.0% ✅ (15B was 83.3%) |
| T13 SyncStore | — | 50.0% ⚠️ |
| Dioxus surface | 0.7.0–0.7.4 | 0.7.0–0.7.9 |
| Dataset | 4,880 rows, 43 topics | 5,287 rows, 57 topics |
The 8B v3.2 outscores the 15B v3.2 (97.56% vs 95.12%) — consistent with the v3.1 pattern where the 8B outperformed the larger model.
Version History
| Version | Base (params) | Score | Exam | Dataset |
|---|---|---|---|---|
| v3.0 15B | Qwen3-Coder-14B (14.8B) | 124.0/144.5 (85.8%) | 103Q weighted | 4,535 |
| v3.1 15B | Qwen3-Coder-14B (14.8B) | 137.0/144.5 (94.81%) | 103Q weighted | 4,880 |
| v3.1 8B | Qwen3-8B (8.2B) | 144.5/144.5 (100.00%) | 103Q weighted | 4,880 |
| v3.1 4B | Qwen3-4B (4.0B) | 143.5/144.5 (99.31%) | 103Q weighted | 4,880 |
| v3.2 15B | Qwen3-Coder-14B (14.8B) | 156.0/164.0 (95.12%) | 114Q weighted | 5,287 |
| v3.2 8B | Qwen3-8B (8.2B) | 160.0/164.0 (97.56%) | 114Q weighted | 5,287 |
Files
neotoi-coder-v3.2-8b-q4_k_m_patched.gguf— current Q4_K_M +qwen3.thinking=truepatch (~4.68 GB)neotoi-coder-v3.1-8b-q4_k_m_patched.gguf— v3.1 archive
Install
Ollama
ollama pull rockypod/neotoi-coder:8b
ollama run rockypod/neotoi-coder:8b "Write a Dioxus 0.7 counter with use_signal"
LM Studio
Download neotoi-coder-v3.2-8b-q4_k_m_patched.gguf from this repo (~4.68 GB).
See integration/lm_studio.md in the GitHub repo
for prompt template setup.
llama.cpp
./llama-cli -m neotoi-coder-v3.2-8b-q4_k_m_patched.gguf -ngl 99 --temp 0.2 \
-p "<|im_start|>user\nYour question<|im_end|>\n<|im_start|>assistant\n<think>"
Model Details
- Base model: Qwen/Qwen3-8B (8.2B total, 6.95B non-embedding)
- Method: RAFT with LoRA adapters (Unsloth)
- Dataset: 5,287 curated Dioxus 0.7 examples across 57 topics (T1–T57)
- Scope: Rust + Dioxus 0.7.0–0.7.9 + Tailwind v4 + WCAG 2.2 AAA
- Quantization: Q4_K_M (~4.68 GB)
- Thinking tokens: patched (
qwen3.thinking = true)
Training
| Field | Value |
|---|---|
| Steps | 2,644 |
| Epochs | 4 |
| LoRA rank | 16 (alpha 16, dropout 0) |
| Target modules | q_proj, k_proj, v_proj, o_proj, gate_proj, up_proj, down_proj |
| Sequence length | 8192 |
| Precision | bf16 + 4-bit base |
| Hardware | RTX 3090 Ti (24 GB) |
What's New in v3.2
- Full Dioxus 0.7 series coverage (0.7.0–0.7.9): Scoped CSS + CSS modules
(0.7.3), SyncStore +
use_store_sync(0.7.2),onauxclick/onscrollendevents (0.7.3), server-only extractors +serde_qs,inertattribute + web panic resilience (0.7.6),IntoAttributeValuefor&T,Action::PartialEq - Format compliance training — fenced-code-only outputs, no orphan
</think>, no prose preamble - Preserve-and-append training — edits to
.ftlcatalogs,Cargo.toml, route enums add without replacing - WCAG / ARIA corrections — T55 correction set ensures
rsx!macro is never dropped on ARIA-heavy components - 5,287 training examples across 57 topics (up from 4,880 / 43 in v3.1)
Enabling Thinking Mode
This model emits Qwen3 native <think>...</think> blocks. Thinking is on
by default with the _patched.gguf quants on inference backends that
honor qwen3.thinking.
Transparency
- Weights: HuggingFace — rockypod/neotoi-coder-8b
- Family hub (8B / 4B / 15B comparison): rockypod/neotoi-coder
- Exam runner, grader, per-question results: GitHub — rockypod/neotoi-coder
- Ollama:
ollama pull rockypod/neotoi-coder:8b
The training dataset is not redistributed — see the GitHub repo for the data-generation pipeline.
License
Fine-tuned weights: Neotoi Coder Community License v1.0 — commercial use of outputs permitted, weight redistribution prohibited, mental health deployment requires written permission. See LICENSE.
Base model: Qwen3-8B — Apache 2.0 © Alibaba Cloud.
Credits
- Unsloth — 2× faster fine-tuning
- Qwen3-8B — base model
- Dioxus — the framework this model specializes in
- Claude Code — dataset pipeline and training infrastructure
Built on a homelab RTX 3090 Ti in Washington State.