Gemma 3 4B IT OlderGated

Released2025-03-12Size4.3B tensorsRoleAssistantLicensegemmaContextunknown

Where it runs 8K context · ≥ 5 t/s floor

Setup	Best plan	Memory	Generate
8 GB GPU + 32 GB RAM	Q8_0llama.cpp CUDAFull GPU	KV unknown	Unknown
12 GB GPU + 32 GB RAM	FP16llama.cpp CUDAFull GPU	KV unknown	Unknown
16 GB GPU + 64 GB RAM	FP16llama.cpp CUDAFull GPU	KV unknown	Unknown
24 GB GPU + 64 GB RAM	FP16llama.cpp CUDAFull GPU	KV unknown	Unknown
32 GB GPU + 128 GB RAM	FP16llama.cpp CUDAFull GPU	KV unknown	Unknown
2× 24 GB GPUs + 128 GB RAM	FP16llama.cpp CUDAFull GPU	KV unknown	Unknown
96 GB workstation GPU + 256 GB RAM	FP16llama.cpp CUDAFull GPU	KV unknown	Unknown
Mac, 24 GB	FP16llama.cpp MetalUnified memory	KV unknown	Unknown
Mac, 48 GB	FP16llama.cpp MetalUnified memory	KV unknown	Unknown
Mac, 128 GB	FP16llama.cpp MetalUnified memory	KV unknown	Unknown
Mac, 512 GB	FP16llama.cpp MetalUnified memory	KV unknown	Unknown
DGX Spark 128 GB	FP16llama.cpp CUDAUnified memory	KV unknown	Unknown
Ryzen AI Max+ 395, 128 GB	FP16llama.cpp VulkanUnified memory	KV unknown	Unknown
CPU only, 64 GB DDR5-5600 dual channel	FP16llama.cpp CPUCPU only	KV unknown	Unknown

Same plan engine as the planner, with its default reserves and RAM. How plans are chosen.

Files exact Hugging Face bytes

Artifact	Size	Files	Revision	Source
FP16 ggml-org/gemma-3-4b-it-GGUF	7.23 GiB	1	d097622374	HF files
Q8_0 ggml-org/gemma-3-4b-it-GGUF	3.85 GiB	1	d097622374	HF files
Q5_K_M bartowski/google_gemma-3-4b-it-GGUF	2.64 GiB	1	71506238f9	HF files
Q4_K_M ggml-org/gemma-3-4b-it-GGUF	2.32 GiB	1	d097622374	HF files
Native safetensors BF16 google/gemma-3-4b-it	8.01 GiB	2	093f9f388b	HF files

Not found in the representative quant search (not proven unavailable): MLX 8-bit, MLX 4-bit, EXL3, AWQ, GPTQ. Complete selected file sets; projectors, MTP/draft and lookup files are listed separately and budgeted only by plans that use them. File bytes are not allocated memory.

Quality

No Artificial Analysis evaluation mapped to this checkpoint.

Arena text rating

1,291

CI 1282–1300 · 4,171 votes · published 2026-10-15 · dataset

Explicit name identity; exact hosted revision/local quant equivalence not verified

Publisher-reported benchmarks

None extracted from the model card.

AA and Arena describe the hosted model in the stated mode, not each quantized file; scores are never inherited by base, pruned, abliterated or quantized variants and never averaged together.

How to run it

No model-specific recipe in the research. The planner binds the backend’s documented command template (llama.cpp, Ollama, LM Studio, MLX, vLLM) to your plan: open a plan’s details.

Recipes are source examples (settings are the publisher’s or author’s, not measured optima) and were not executed by LocalFit. A downloadable file does not prove your backend build supports the architecture.

Reported speeds 3

Hardware	Quant	Backend	Placement	Prompt t/s	Gen t/s	Date	Source
Intel(R) Iris(R) Xe Graphics	Q4_K_M	llama.cpp Vulkan	Full GPU	254	14	2026-02-19	llama.cpp #10879
Intel(R) Graphics	Q8_0	llama.cpp Vulkan	Full GPU	281	8.6	2025-08-07	llama.cpp #10879
AMD Radeon Graphics	Q4_0	llama.cpp Vulkan	Full GPU	156	18	2025-04-22	llama.cpp #10879

Reported by the linked sources, not reproduced by LocalFit. Stream pieces/s and method-unknown rates are flagged and cannot pass a tokens/s floor. All rows