# Qwen3 Coder Models — June 20, 2026

## Models Added

### 1. Qwen3-Coder-30B-A3B Q4_K_M
- **Repo:** unsloth/Qwen3-Coder-30B-A3B-Instruct-GGUF
- **File:** `Qwen3-Coder-30B-A3B-Instruct-Q4_K_M.gguf` (18 GB)
- **Arch:** qwen3moe, 30.5B total / 3.3B active, 48 layers, 128 experts/8 active
- **GQA:** 32 Q heads / 4 KV heads
- **Context:** 262K native
- **Recommended flags:** `--temp 1.0 --top-p 0.95 --min-p 0.01 --top-k 40` (per unsloth guide)
- **Benchmark:** 34.6 t/s, 7.9 GB VRAM, 37s cold start at 131K ctx
- **No MTP/Eagle3** — no speculative decoding variants exist

### 2. Qwen3-Coder-Next Q4_K_M
- **Repo:** unsloth/Qwen3-Coder-Next-GGUF
- **File:** `Qwen3-Coder-Next-Q4_K_M.gguf` (45 GB)
- **Arch:** qwen3next (hybrid DeltaNet+Attention+MoE), **80B total** / 3B active, 48 layers
- **512 experts** / 10 active / 1 shared
- **GQA:** 16 Q heads / 2 KV heads
- **Context:** 262K native
- **Recommended flags:** same as 30B-Coder
- **Benchmark:** 33.1 t/s, **4.4 GB VRAM** (very efficient!), 133s cold start
- **No MTP/Eagle3** — no speculative decoding variants exist
- Per unsloth docs: needs >45 GB RAM (but with --n-cpu-moe, only 4.4 GB VRAM needed)

### 3. Qwen3-Coder-Next-REAP-48B Q4_K_XL (downloading/partial)
- **Repo:** lovedheart/Qwen3-Coder-Next-REAP-48B-A3B-GGUF
- **File:** `Qwen3-Coder-Next-REAP-48B-A3B-Q4_K_XL.gguf` (31 GB)
- **Same qwen3next arch** as Coder-Next but REAP-pruned (308 experts from 512, 40% pruning)
- **No Q4_K_M available** — only Q4_K_XL
- Slow CDN (~9 MB/s from lovedheart repo)
- Same flags as Coder-Next

### 4. Qwen3-Coder-Next-REAP-40B Q4_K_M (WORKING — added 2026-06-22)
- **Repo:** mradermacher/Qwen3-Coder-Next-REAP-40B-A3B-i1-GGUF
- **File:** `Qwen3-Coder-Next-REAP-40B-A3B.i1-Q4_K_M.gguf` (24 GB on disk, 23.2 GB download)
- **Arch:** qwen3next (hybrid DeltaNet+Attention+MoE), **40B total / 3B active**, 48 layers
- **256 experts** / 10 active / 1 shared expert
- **12 full-attention layers** (2 KV heads, 256 head_dim) + **36 linear-attention layers**
- **KV cache:** Only the 12 full-attention layers need it — 12,288 bytes/token at q8_0 = ~3 GB at 262K full context
- **Linear attention state:** Fixed-size recurrent state (~18 MB total for all 36 layers)
- **Native context:** 262,144 tokens
- **n-cpu-moe tuning:** --n-cpu-moe 38 keeps 10 layers' experts on GPU (38 on CPU)
- **VRAM:** 10,666 MiB at full 262K context with q8_0 KV
- **Benchmark:** 38.2 t/s gen, 72.3 t/s prompt
- **Critical config:** `--batch-size 128 --ubatch-size 32 --no-warmup` required to fit compute graph buffer (~708 MiB at 262K)
- **No MTP/Eagle3** — no speculative decoding variants
- **CDN speed:** Fast (~37 MB/s from mradermacher via HF)
- **Aliases in llama-swap:** `reap-coder`, `coder-reap`, `qwen-reap`

## Performance Notes

All three coder models lack speculative decoding heads (no MTP/Eagle3). Speed is ~33-35 t/s regardless of total parameter count because active params are ~3B and CPU offload is PCIe-bound. The primary benefit of Coder-Next (80B) over Coder-30B (30B) is **quality**, not speed.