37% Experts Pruned, 88.4 HUMANEVAL (base 92.1)

Qwen3-Coder-30B-A3B-Instruct compacted via per-layer-normalized MoE expert pruning against the unmodified teacher.

• HUMANEVAL: 88.4 (base 92.1, Δ -3.7)
• HUMANEVAL+PLUS: 85.4 (base 88.4, Δ -3.0)

The first 30B-class coder that fits a 12 GB consumer GPU. Forged from Qwen/Qwen3-Coder-30B-A3B-Instruct by removing 35% of MoE experts via calibration-aware activation-frequency ranking on a held-out Python code corpus. Quantized to GGUF Q4_K_M / Q5_K_M / Q8_0 for llama.cpp / Ollama / LM Studio. Native 256K context preserved (the only Qwen3-Coder variant in this VRAM tier with 256K). Apache-2.0. 88.4 HumanEval against the unmodified base's 92.1 (Δ −3.7), both hardware-measured on RTX 5090 in the same eval pipeline. This is the only Qwen3-Coder-class artifact at the 12 GB tier. Cryptographic provenance via ForgeAlloy — the published per-problem JSONL outputs let anyone re-score against the alloy's resultHash without trusting the producer's claim.

Benchmarks

What Changed (Base → Forged)

Runs On

Quick Start

from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained("continuum-ai/qwen3-coder-30b-a3b-compacted-19b-256k",
    torch_dtype="auto", device_map="auto")
tokenizer = AutoTokenizer.from_pretrained("continuum-ai/qwen3-coder-30b-a3b-compacted-19b-256k")

inputs = tokenizer("def merge_sort(arr):", return_tensors="pt").to(model.device)
output = model.generate(**inputs, max_new_tokens=200)
print(tokenizer.decode(output[0], skip_special_tokens=True))

Methodology

Produced via MoE expert pruning, GGUF quantization. Full methodology, ablations, and per-stage rationale are in the methodology paper and the companion MODEL_METHODOLOGY.md in this repository. The pipeline ran as expert-activation-profile → expert-prune → quant → eval over 1 cycle on NVIDIA GeForce RTX 5090.

Limitations

• HumanEval is 3.7 points below the base anchor (88.4 vs 92.1, both hardware-measured). This residual gap is load-bearing: the calibration-aware metric closed it from −13.4 (router-gate-norm baseline, see priorMetricBaselines) but did not eliminate it. A v2 release will add KL-distillation compensation LoRA training to attempt to close the remaining gap; check this repo for updates.
• Calibration corpus was 300 Python code examples (~125K tokens). Generalization to non-Python languages is not separately measured. If your workload is dominated by Rust, Go, or other less-represented languages, the 3.7 point gap may differ on those workloads.
• Long-context generation (>32K tokens) has not been benchmarked. M-RoPE is preserved bit-exact from the source so the architectural capability is intact, but no held-out long-context evaluation has been run.
• MBPP+ and LiveCodeBench v6 are not yet measured. HumanEval/HumanEval+ alone establish the headline; the additional benchmarks will land in a v1.x release without a re-prune.

Chain of Custody

Scan the QR or verify online. Download the alloy file to verify independently.

Make Your Own

Forged with Continuum — a distributed AI world that runs on your hardware.

The Factory configurator lets you design and forge custom models visually — context extension, pruning, LoRA, quantization, vision/audio modalities. Pick your target devices, the system figures out what fits.

GitHub · All Models · Forge-Alloy

License

apache-2.0