How to use from
Hermes Agent
Start the llama.cpp server
# Install llama.cpp:
brew install llama.cpp
# Start a local OpenAI-compatible server:
llama serve -hf jaimef21/crystal-qwen-v3-30b-gguf
Configure Hermes
# Install Hermes:
curl -fsSL https://hermes-agent.nousresearch.com/install.sh | bash
hermes setup
# Point Hermes at the local server:
hermes config set model.provider custom
hermes config set model.base_url http://127.0.0.1:8080/v1
hermes config set model.default jaimef21/crystal-qwen-v3-30b-gguf
Run Hermes
hermes
Quick Links

crystal-qwen-v3-30b-gguf

LoRA-fine-tuned Qwen3-Coder-30B-A3B-Instruct specialised for the Crystal programming language. Quantized to Q8_0 GGUF for local inference via Ollama or llama.cpp.

This is v3 โ€” a full-pipeline retrain on a substantially larger and cleaner corpus than v2 (crystal-qwen3.6-30b-gguf).

Quick start (Ollama)

# Download the GGUF + Modelfile from this repo, then:
ollama create crystal-qwen-v3 -f Modelfile
ollama run crystal-qwen-v3 "Write a Crystal class Item with name and quantity that serializes to JSON."

What v3 changes vs v2

v2 (3.6) v3
CPT tokens ~3 M ~50 M (top-500 Crystal GitHub repos + stdlib + book + RFCs)
SFT pairs ~600 ~3 000 (mined + Claude-Haiku-4.5 augmented, compile-gated)
DPO pairs 37 74 (chosen = idiomatic Crystal; rejected = Ruby-style or wrong)
LoRA rank / ฮฑ 32 / 64 64 / 128
Quantization Q4_K_M Q8_0
Targets attn only attn + MLP + MoE experts.gate_up_proj/down_proj

Eval

Held-out eval (30 natural-language Crystal tasks; idiom score + crystal build --no-codegen compile gate). Higher is better.

Model Idiom Compile pass Total
crystal-qwen-v3 (this model) +76 26/28 (93 %) +206
jaimef21/crystal-qwen3.6-30b (v2) +67 21/28 (75 %) +172
Qwen3-Coder-30B-A3B-Instruct (base) +71 21/28 (75 %) +176

v3 beats both prior baselines on every axis. v2 actually slightly trailed the base model on idiom (+67 vs +71) โ€” v3 is the first checkpoint of this lineage that is unambiguously better than vanilla Qwen3-Coder at Crystal.

Caveat: v3 ships at Q8_0 while v2 and base measurements above are Q4_K_M, so some compile-gate fidelity gap is attributable to less quantization noise.

Files

  • crystal-qwen-v3-30b.gguf โ€” 32 GB, Q8_0 (8.51 BPW)
  • Modelfile โ€” Ollama Modelfile (chatml template + Crystal system prompt)
  • README.md โ€” this file

Reproducing

Full reproduction pipeline (data scraping โ†’ training on RunPod H200 โ†’ GGUF quantization โ†’ eval) is documented at: github.com/jaimef/crystal-lora โ€” see REPRODUCE.md.

License

Inherits Qwen's Apache 2.0 license from the base model.

Downloads last month
6
GGUF
Model size
31B params
Architecture
qwen3moe
Hardware compatibility
Log In to add your hardware

We're not able to determine the quantization variants.

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support

Model tree for jaimef21/crystal-qwen-v3-30b-gguf

Adapter
(54)
this model