---

license: apache-2.0

base_model: huihui-ai/Huihui-Qwen3-Coder-30B-A3B-Instruct-abliterated

tags:

- gguf

- uncensored

- abliterated

- code

- moe

- ollama

---

Huihui-Qwen3-Coder-30B-A3B-Instruct-abliterated-i1-GGUF — AI-conn fork

> This is a fork, not our model. The abliterated weights are by

> huihui-ai;

> this GGUF conversion is by

> mradermacher;

> the base model is Qwen/Qwen3-Coder-30B-A3B-Instruct.

> This fork exists so the fleet's copy cannot be altered or pulled upstream,

> and to carry the fit notes below. License: Apache-2.0 (per the source repo).

Uncensored coder pick: MoE, 30B total / ~3B active per token — but note that

all expert weights must sit in VRAM for GPU residency; "3B active" does

not shrink the memory bill.

Recommended file for a 20 GB card: Huihui-Qwen3-Coder-30B-A3B-Instruct-abliterated.i1-Q4_K_M.gguf (17.28 GiB).

Ollama pull:

ollama pull hf.co/AI-conn/Huihui-Qwen3-Coder-30B-A3B-Instruct-abliterated-i1-GGUF:i1-Q4_K_M

Prime fit (measured budget, RX 7900 XT — 19.98 GiB VRAM, 32 GB RAM)

• i1-Q4_K_M 17.28 GiB + small MoE KV + ~0.7 GiB reserve ≈ **18.3 GiB —

resident, tight**

• i1-IQ4_XS (15.24 GiB) is the roomy option if you want more context headroom

at slightly lower quant fidelity.

• Independent datapoint (community harness): the stock base scores 60.4% on

HumanEval Pro. No independent coding eval of any abliterated build exists

yet — our fleet settles coder rankings with a local bake-off, not cards.