---

license: apache-2.0

base_model: baidu/ERNIE-4.5-21B-A3B-PT

base_model_relation: quantized

pipeline_tag: text-generation

library_name: gguf

tags:

- gguf

- moe

- apex

- quantized

- ernie

- llama.cpp

---

ERNIE-4.5-21B-A3B — APEX GGUF

MoE-aware, mixed-precision APEX quantizations of

baidu/ERNIE-4.5-21B-A3B-PT

— Baidu's ernie4_5-moe: 28 layers (dense layer 0 + 27 MoE), **64 routed + 2

shared** experts, top-6 routing, ~21B total / ~3B active, Apache-2.0.

Results & which file to pick

Perplexity on wikitext-2-raw (test, 200×512-token windows), llama-perplexity.

Both quants are imatrix-guided (Bartowski calibration_datav3).

| File | Size | BPW | PPL | Δ vs bf16 | Fits (VRAM) |

|------|------|-----|-----|-----------|-------------|

| bf16 (reference) | 41 GB | 16.0 | 7.362 | — | — |

| APEX-i-quality | 15 GB | 5.63 | 7.352 | −0.14% (at parity) | 16 GB — best quality |

| APEX-i-compact | 11 GB | 4.11 | 7.640 | +3.77% | 12 GB — smallest usable |

• i-quality — recommended; effectively lossless — its PPL lands a hair

below the bf16 reference (within noise). Q6_K/Q5_K experts on a depth gradient,

Q8_0 shared, Q6_K attention. Runs on a 16 GB card.

• i-compact — Q4_K_M-class mid experts; a real but usable step (+3.8%) that

fits a 12 GB card.

ERNIE's routed-expert intermediate dim is 1536 (256-divisible), so i-quants

apply cleanly to every expert tensor — no IQ4_NL fallback needed.

Usage (llama.cpp)

llama-cli    -m ERNIE-4.5-21B-A3B-APEX-i-quality.gguf -ngl 999 -p "Hello"
llama-server -m ERNIE-4.5-21B-A3B-APEX-i-compact.gguf -ngl 999 --host 0.0.0.0 --port 8080

Requires a llama.cpp build supporting the ernie4_5-moe architecture.

Method

APEX = per-tensor-role, per-layer precision over stock

llama-quantize --tensor-type-file. ERNIE-4.5-21B is a standard MoE —

conventional ffn__exps / ffn__shexp tensor names, a single dense layer 0

(--dense-layers 1) — so it used the stock APEX config generator, no patching

(unlike the Kimi/Granite/Moonlight hybrids in this collection). Routed experts on

a layer-depth gradient; shared experts protected at Q8_0/Q6_K; router left high.

Baseline: bartowski's bf16 GGUF.

imatrix from Bartowski calibration_datav3. See REPRODUCE.md and configs/.

Attribution & licenses

See LICENSE (Apache-2.0) and NOTICE.

• Base: Baidu (@baidu) — ERNIE-4.5-21B-A3B-PT (Apache-2.0)
• bf16 GGUF: bartowski (@bartowski) — source
• Engine: llama.cpp (@ggml-org) (MIT)
• APEX: Ettore Di Giacinto / LocalAI (@mudler) — localai-org/apex-quant (MIT)
• Calibration: Bartowski (@bartowski) — calibration_datav3

Unofficial community quantization; not affiliated with or endorsed by Baidu.