An IQ2_M (imatrix) quant of Qwen/Qwen3-Coder-30B-A3B-Instruct at ~10 GB — runs on a normal laptop, stays a capable agentic coder at 2-bit. Weights & training © Qwen (Apache-2.0); independent quant + eval, full credit. (Other quants: unsloth — this one adds an independent agentic eval, the exact test set, and configurable-context Modelfiles.)

📦 The quant

Qwen3-Coder-30B-A3B-Instruct-IQ2_M.gguf — ~10 GB, imatrix-quantized (2-bit "I-quant": smallest size that stays sharp; needs the bundled imatrix, a touch more CPU than K-quants). 30B total / 3B active (MoE → fast inference despite the size).

🧮 Will it fit? (RAM/VRAM cheat-sheet)

Total ≈ 10 GB weights + KV-cache (grows with context):

you have	context you can run
12 GB	~8–16K
16 GB	~32–64K
24 GB	~128K
32 GB+	full native (262K)

🧠 Context size (262K native)

Configurable — Modelfiles at 8K / 32K / 128K / 256K included; full guide in CONTEXT.md.

bash

ollama create qc-128k -f Modelfile-128k && ollama run qc-128k
# or per-session:  >>> /set parameter num_ctx 262144

🚀 How to run it

bash

ollama run hf.co/KikoCis/Qwen3-Coder-30B-A3B-Instruct-IQ2_M-GGUF   # default 32K
llama-server -m Qwen3-Coder-30B-A3B-Instruct-IQ2_M.gguf -c 131072 --jinja

Recommended sampling: temperature 0.7, top_p 0.8. Agentic use: this model emits the <function=...> (OpenHands/Qwen-Coder) tool format. In OpenHands / Qwen-Agent it's native. In a Claude-Code-style harness, inject tool defs as TEXT in that format — relying on the native tools param yields empty output (that fix took it from 0 → solving tasks here).

⚠️ Good to know

Strengths: agentic coding — at IQ2 it still does clean multi-step tool use + writes correct code (verified, see below).
Limits: 2-bit is lossy vs the fp16; for max fidelity use a higher quant. It's a coder, English-centric.
The agentic tool-format note above matters — without it, it looks "broken" when it isn't.

📊 Evaluation methodology (so the number is reproducible)

Local agentic SWE-bench Verified probe, Claude Code + agent-bridge harness (terminal-bench/Docker), temp 0.7 / top_p 0.8, tool format openhands.

probe	score	what it is
`swe-mix` (ranking)	3/30	6 continuity anchors + 24 discriminating (15× "1-4 h" + 9× "15 min-1 h", 10 repos). Exact instances in `swe_mix.json`.
`django-6` (legacy/easy)	4/6	the original easy probe (saturated)

On swe-mix, the 3 solved are the easier "anchor" tasks; the "1–4 h" instances are out of reach for a 10 GB model in a 20-min agent budget (expected). This ranks models head-to-head on the same set — not 1:1 with the official SWE-bench leaderboard. Per-instance pass/fail and the imatrix are reproducible from the artifacts here.

🗒️ Changelog

2026-06-28 — v1: IQ2_M release + context Modelfiles + agentic eval.

📚 Credit & license

Model, weights, training: © Qwen — Qwen3-Coder-30B-A3B-Instruct. IQ2_M imatrix quant + context Modelfiles + eval: KikoCis. Apache-2.0 (same as upstream). No weights modified.