gemma-4-26B-A4B-it ยท GGUF
GGUF builds of google/gemma-4-26B-A4B-it for llama.cpp / Ollama / LM Studio. Converted by axteam with llama.cpp (converted from the BF16 original; the chat template is embedded in the GGUF).
| File | Quant | Size | Notes |
|---|---|---|---|
gemma-4-26B-A4B-it-GGUF-BF16.gguf |
BF16 | 50.5 GB | unquantized 16-bit (reference; for re-quantizing or max quality) |
gemma-4-26B-A4B-it-GGUF-Q4_K_M.gguf |
Q4_K_M | 16.8 GB | 4-bit, the usual size/quality trade-off โ the most common choice |
gemma-4-26B-A4B-it-GGUF-Q8_0.gguf |
Q8_0 | 26.9 GB | 8-bit, closest to BF16 (larger) |
Q4_K_M is the usual default (size/quality trade-off). Q8_0 stays closest to BF16. BF16 is the unquantized reference. Measured agreement with the BF16 file is in the table below. Architecture: Gemma 4 ยท Mixture-of-Experts (multimodal; the GGUF text model runs the language side).
Quantization method
These GGUF files are post-training quantization (PTQ) made with llama.cpp โ the BF16 weights are quantized block-wise (llama.cpp k-quants: per-block scales with mixed bit-width allocation, a refined round-to-nearest), with no retraining. This is a different toolchain from the AWQ / NVFP4 vLLM builds, and it is not QAT (quantization-aware training, which fine-tunes the weights). Q8_0 stays closest to BF16, Q4_K_M is the usual size/quality trade-off (see the measured agreement below), and BF16 is the unquantized reference for re-quantizing.
Measured on GPU (llama.cpp)
llama.cpp b12005 (commit e0f142d, CUDA build) on an NVIDIA B200, 2026-11-02, all layers on the GPU, 16 windows of 2,048 tokens from the WikiText-2 test set and from a Korean Wikipedia sample. Raw-text perplexity of this instruction-tuned checkpoint is far above the base model's (the BF16 file scores 10,111 on WikiText-2 without its chat template), so it says little about quantization loss; each quantized file is compared with the BF16 file on the same tokens instead: mean KL divergence (lower = closer) and how often both pick the same top-1 token, for each quantized file.
| File | Mean KLD vs BF16 (EN / KO) | Top-1 agreement with BF16 (EN / KO) |
|---|---|---|
gemma-4-26B-A4B-it-GGUF-Q4_K_M.gguf |
3.1405 / 3.2336 | 43.0% / 42.2% |
gemma-4-26B-A4B-it-GGUF-Q8_0.gguf |
0.8216 / 0.9642 | 71.6% / 70.4% |
These figures are measured on the same raw text without the chat template, on which the BF16 file itself is far off (WikiText-2 perplexity 10,111), so they show how far the quantized files' predictions drift in that off-template setting โ not chat quality. (The same quant types on the base
gemma-4-26B-A4Bchange its WikiText-2 perplexity by Q8_0 +0.03% and Q4_K_M +4.60% vs its BF16 file.)
Requirements
| llama.cpp | build โฅ b8637 โ architecture gemma4 needs โฅ b8637, and the commands on this card use -fa on (โฅ b6325). Run here on 2026-11-02 with b12005 (commit e0f142d, CUDA build) on an NVIDIA B200: every file in this repo loaded with all layers on the GPU, each quantized file's agreement with the BF16 file was measured (see Measured on GPU above) and two short chat questions (capital of France / of Korea) were answered correctly with every file. |
| Memory | the chosen GGUF file (sizes in the table) plus KV cache for the context (-c); --cache-type-k/v q8_0 -fa on halves the KV cache. |
| Input | text only โ this repo has no mmproj (vision projector) file. |
| Ollama / LM Studio | not tested here. |
Run with llama.cpp
# download one file (e.g. Q4_K_M) then:
llama-server -m gemma-4-26B-A4B-it-GGUF-Q4_K_M.gguf -c 8192
# or a single answer and exit (-st):
llama-cli -m gemma-4-26B-A4B-it-GGUF-Q4_K_M.gguf -p "๋ํ๋ฏผ๊ตญ์ ์๋๋?" -st
Reduce KV-cache memory (optional, run-time)
KV-cache type is a run-time choice in llama.cpp (not baked into the file). Halve it with:
llama-server -m gemma-4-26B-A4B-it-GGUF-Q4_K_M.gguf --cache-type-k q8_0 --cache-type-v q8_0 -fa on
Ollama
ollama run hf.co/axteam/gemma-4-26B-A4B-it-GGUF:Q4_K_M
Recommended sampling (per Google):
temperature=1.0,top_p=0.95,top_k=64.
License
Quantized derivative of Google's Gemma 4 โ governed by the Gemma 4 license. Weights, tokenizer, and chat template belong to Google DeepMind.
ํ๊ตญ์ด ์์ฝ
google/gemma-4-26B-A4B-it์ GGUF ์์ํ ๋น๋ (llama.cppยทOllamaยทLM Studio์ฉ) ยท ์ ์ axteam
๊ฐ์
| ํญ๋ชฉ | ๋ด์ฉ |
|---|---|
| ์๋ณธ ๋ชจ๋ธ | google/gemma-4-26B-A4B-it |
| ํ์ | GGUF (llama.cpp) |
| ๋ณํ ๋๊ตฌ | llama.cpp convert_hf_to_gguf.py โ llama-quantize |
| ๋ชจ๋ธ ๊ตฌ์กฐ | Gemma 4 ยท MoE (์๋ณธ์ ๋ฉํฐ๋ชจ๋ฌ, GGUF๋ ์ธ์ด ๋ชจ๋ธ ๋ถ๋ถ๋ง ์คํ) |
| ํ์ผ ์ | 3๊ฐ |
์์ํ ์์ธ
- ์์ํ ์ ์ฐจ: PTQ(ํ์ต ํ ์์ํ, ์ฌํ์ต ์์), llama.cpp ๋ธ๋ก ๋จ์ ์์ํ(k-quant
Q*_K,Q8_0) - chat template: GGUF ํ์ผ์ ๋ด์ฅ
- BF16 ์ผ์น๋(์ค์ธก, 2026-11-02): ์์ํ ํ์ผ๋ง๋ค BF16 ํ์ผ๊ณผ์ ์ผ์น๋(๊ฐ์ ํ
์คํธ ์์ด/ํ๊ตญ์ด 2,048ํ ํฐร16, llama.cpp KL divergence ํ๊ท ยทtop-1 ํ ํฐ ์ผ์น์จ, B200 GPU) โ
gemma-4-26B-A4B-it-GGUF-Q4_K_M.ggufKLD 3.1405/3.2336, top-1 ์ผ์น 43.0%/42.2% ยทgemma-4-26B-A4B-it-GGUF-Q8_0.ggufKLD 0.8216/0.9642, top-1 ์ผ์น 71.6%/70.4% (chat template ์๋ ๊ฐ์ ํ ์คํธ์์ ์ฐ ๊ฐ โ BF16 ํ์ผ ์์ฒด๋ WikiText-2 PPL์ด 10,111๋ก ํฌ๊ฒ ๋ฒ์ด๋๋ ์กฐ๊ฑด์ด๋ผ ์ฑํ ํ์ง์ด ์๋๋ผ ๊ทธ ์กฐ๊ฑด์์ ์์ธก์ด BF16๊ณผ ์ผ๋ง๋ ๋ฌ๋ผ์ง๋์ง๋ฅผ ๋ํ๋) โ ์ฐธ๊ณ : ๊ฐ์ ์์ํ ํ์์ basegemma-4-26B-A4B์ ์ ์ฉํ๋ฉด WikiText-2 PPL์ด BF16 ํ์ผ ๋๋น Q8_0 +0.03%ยทQ4_K_M +4.60% ๋ณํจ
์ ๊ณต ํ์ผ
| ํ์ผ | ํฌ๊ธฐ | ์ค๋ช |
|---|---|---|
gemma-4-26B-A4B-it-GGUF-BF16.gguf |
50.5 GB | ์์ํํ์ง ์์ 16๋นํธ โ ๊ธฐ์ค ํ์ผ (์ฌ์์ํ์ฉยท์ต๊ณ ํ์ง) |
gemma-4-26B-A4B-it-GGUF-Q4_K_M.gguf |
16.8 GB | 4๋นํธ โ ํฌ๊ธฐยทํ์ง ๊ท ํ, ๊ฐ์ฅ ๋ง์ด ์ฐ๋ ์ ํ |
gemma-4-26B-A4B-it-GGUF-Q8_0.gguf |
26.9 GB | 8๋นํธ โ BF16์ ๊ฐ์ฅ ๊ฐ๊น์, ์ฉ๋ ํผ |
์๊ตฌ์ฌํญ
- llama.cpp: b8637 ์ด์ (
gemma4๊ตฌ์กฐ ์ง์ b8637๋ถํฐ, ์์ ๋ช ๋ น์-fa on์ต์ b6325๋ถํฐ) - ์คํ ํ์ธ: 2026-11-02 b12005(์ปค๋ฐ e0f142d) CUDA ๋น๋๋ก NVIDIA B200์์ ์ด ๋ฆฌํฌ์ ๋ชจ๋ ํ์ผ์ ์ ์ธต GPU์ ์ฌ๋ ค ์คํ โ ์์ํ ํ์ผ๊ณผ BF16 ํ์ผ์ ์ผ์น๋ ์ธก์ (์ ์์ด ๋ณธ๋ฌธ ํ), ๋ชจ๋ ํ์ผ์์ ์งง์ ์ง๋ฌธ 2๊ฐ(ํ๋์คยท๋ํ๋ฏผ๊ตญ ์๋)์ ๋ฐ๋ฅด๊ฒ ๋ตํจ
- ๋ฉ๋ชจ๋ฆฌ: ์ ํํ GGUF ํ์ผ ํฌ๊ธฐ + ์ปจํ
์คํธ(
-c)๋งํผ์ KV ์บ์ (KV ์บ์ ์ค์ด๋ ๋ฐฉ๋ฒ์ ์๋ ์ฌ์ฉ๋ฒ ์ฐธ๊ณ ) - ์
๋ ฅ: ํ
์คํธ ์ ์ฉ โ ๋น์ ์ธ์ฝ๋ยทํฌ์์ธต ํ์ผ(
mmproj)์ด ์์ด ์ด๋ฏธ์ง ์ ๋ ฅ ๋ถ๊ฐ - OllamaยทLM Studio: ์ ์ ์ธก์์ ์คํ ๋ฏธํ์ธ
์ฌ์ฉ๋ฒ
- ํ์ผ ๋ฐ๊ธฐ:
hf download axteam/gemma-4-26B-A4B-it-GGUF gemma-4-26B-A4B-it-GGUF-Q4_K_M.gguf --local-dir .(์๋-m๋ช ๋ น์ ํ์, ๋ค๋ฅธ ์์ํ๋ ํ์ผ๋ช ์Q4_K_M๋ง ๋ณ๊ฒฝ) - ์๋ฒ ์คํ:
llama-server -m gemma-4-26B-A4B-it-GGUF-Q4_K_M.gguf -c 8192 - ํ ๋ฒ ์ง๋ฌธ (๋ต๋ณ ํ ์ข
๋ฃ):
llama-cli -m gemma-4-26B-A4B-it-GGUF-Q4_K_M.gguf -p "๋ํ๋ฏผ๊ตญ์ ์๋๋?" -st - Ollama:
ollama run hf.co/axteam/gemma-4-26B-A4B-it-GGUF:Q4_K_M - KV ์บ์ ์ ๊ฐ(์ ํ):
--cache-type-k q8_0 --cache-type-v q8_0 -fa onโ KV ์บ์ ์ฝ ์ ๋ฐ, GGUF ํ์ผ๊ณผ ์๊ด์์ด ์คํ ์ ์ง์ - ๊ถ์ฅ ์ํ๋ง:
temperature=1.0,top_p=0.95,top_k=64(Google ๊ถ์ฅ๊ฐ)
๋ผ์ด์ ์ค
- ๋ผ์ด์ ์ค: Apache-2.0 (Gemma 4 ๋ผ์ด์ ์ค ์๋ด) โ ์๋ณธ ๋ชจ๋ธ์ ๋ผ์ด์ ์ค๋ฅผ ๊ทธ๋๋ก ๋ฐ๋ฆ
- ๊ถ๋ฆฌ: ์๋ณธ ๊ฐ์ค์นยทํ ํฌ๋์ด์ ยทchat template์ Google DeepMind ์์ , ์ด ๋ฆฌํฌ๋ ์์ํ ํ์๋ฌผ