ExtGemma4 40 5B — Hardware Requirements & GPU Compatibility
ChatReasoning
ExtGemma4 40 5B is a 39.5B-parameter open language model from TOTORONG in the Gemma 4 family. It supports a context window of up to 262,144 tokens. At BF16 it needs about 80.94 GB of VRAM — see which GPUs and Macs can run it below.
ExtGemma4 40 5B runs on a 96 GB+ Mac or a workstation GPU
Best local pick → Mac Studio (M3 Ultra, 96GB) · BF16 · 82.7 GB · ~7 tok/s
Macs from 96 GB (BF16 · ~7 tok/s)
Gemma Terms
Not set up for it?
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run ExtGemma4 40 5B
One recommended way. Every other method, OS and quant is in the step-by-step guide.
Recommended
vllm serve TOTORONG/extGemma4-40_5B --max-model-len 16384 --tensor-parallel-size 8vLLM repo with its own quantisation
Download: Weights on Hugging Face
Other ways
each opens step-by-stepRun in cloud
Fits on H200 (141 GB) (58 GB headroom) · BF16
- Generation speed
- ~39 tok/s
- generation speed
- Cost per 1M output tokens
- $25.90
- per 1M output tokens
How Much VRAM Does ExtGemma4 40 5B Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | Runs on | |
|---|---|---|---|---|
| 80.9 GB | 85.6 GB | 16-bit · bfloat — Brain floating point 16 — preferred for training | Runs on RTX PRO 6000 Blackwell Max-Q Workstation Edition 96 GB |
Selected BF16
Brain floating point 1682.7GBat 16K context
- Weights
- 78.9 GB
- KV cache
- 3.5 GB
- Overhead
- 0.3 GB
Server-class: ~82.7 GB at 16K context · one RTX PRO 6000 Blackwell Max-Q Workstation Edition 96 GB
Which Hardware Can Run ExtGemma4 40 5B?
BF16 · 80.9 GBExtGemma4 40 5B (BF16) requires 80.9 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 106+ GB is recommended. Using the full 262K context window can add up to 44.1 GB, bringing total usage to 125.0 GB. No consumer GPU has enough memory. 30 devices with unified memory can run ExtGemma4 40 5B, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Studio (M3 Ultra, 96GB).
96 GB
fits · 13.3 GB free · up to 64K ctx
3 GPUs · 2 Mac classes
RTX PRO 6000 Blackwell Max-Q Workstation Edition~14Simulate output on RTX PRO 6000 Blackwell Max-Q Workstation EditionRTX PRO 6000 Blackwell Workstation Edition~14Simulate output on RTX PRO 6000 Blackwell Workstation EditionRTX PRO 6000 Blackwell Server Edition~13Simulate output on RTX PRO 6000 Blackwell Server EditionM3 Ultra*96 GB~7Simulate output on Mac Studio (M3 Ultra, 96GB)M3 Max*96 GB~3Simulate output on MacBook Pro 14" M3 Max (96 GB)128–144 GB
fits · 45.3 GB free · up to 256K ctx
4 GPUs · 2 Mac classes · 4 device families
GH200 Grace Hopper Superchip~39Simulate output on GH200 Grace Hopper SuperchipH200 NVL~39Simulate output on H200 NVLH200 SXM~39Simulate output on H200 SXMM2 Ultra*128 GB~7Simulate output on Mac Pro M2 Ultra (128 GB)M5 Max128 GB~5Simulate output on Mac Studio M5 Max (128 GB)GB10~2Simulate output on ASUS Ascent GX10Jetson T5000 (Thor)~2Simulate output on NVIDIA Jetson AGX Thor Developer Kit
By VRAM tier · typical GPUs and Macs at BF16
Consumer GPUs (up to 32 GB) and Macs up to 64 GB: too small for this model (needs ≥ 82.7 GB)8 GB tier →12 GB tier →16 GB tier →24 GB tier →Macs →
Alternatives to ExtGemma4 40 5B
Gemma 4 family · 35 models →Similar models
3| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| Phi 3.5 MoE InstructMicrosoft | — | 23.9 GB | ~141 | same size |
| Falcon 40B InstructTII UAE | — | 23.8 GB | ~33 | same size |
| K2 Horizon MoVA 36B A4BIFM | — | 23.4 GB | ~162 | same size |
Specifications
- Publisher
- TOTORONG
- Family
- Gemma 4
- Parameters
- 39.5B
- Architecture
- Gemma4ForConditionalGeneration
- Context Length
- 262,144 tokens
- Vocabulary Size
- 262,144
- Release Date
- 2026-07-12
- License
- Gemma Terms
Frequently Asked Questions
How much VRAM does ExtGemma4 40 5B need?
ExtGemma4 40 5B requires 80.9 GB of VRAM at BF16. Full 262K context adds up to 44.1 GB (125.0 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 39.5B × 16 bits ÷ 8 = 78.9 GB
KV Cache + Overhead ≈ 2 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
KV Cache + Overhead ≈ 46.1 GB (at full 262K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 5090 run ExtGemma4 40 5B?
No — ExtGemma4 40 5B requires at least 80.9 GB at BF16, which exceeds the NVIDIA GeForce RTX 5090's 32 GB of VRAM.
Can I run ExtGemma4 40 5B on a Mac?
Yes — Mac Studio (M3 Ultra, 96GB) and 20 other Macs can run ExtGemma4 40 5B. Apple Silicon uses unified memory, so the model shares RAM with the system. At BF16 you need at least 80.9 GB of usable unified memory (RAM minus macOS overhead).
Can I run ExtGemma4 40 5B locally?
Yes — ExtGemma4 40 5B can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is ExtGemma4 40 5B?
At BF16, ExtGemma4 40 5B can reach ~59 tok/s on AMD Instinct MI350X. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = (bandwidth GB/s ÷ model GB) × efficiency
Example: NVIDIA B200 → 8000 ÷ 80.9 × 0.65 = ~64 tok/s
Estimated speed at BF16 (80.9 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of ExtGemma4 40 5B?
At BF16, the download is about 78.94 GB.
Which GPUs can run ExtGemma4 40 5B?
No single consumer GPU has enough VRAM to run ExtGemma4 40 5B at BF16 (80.9 GB). Multi-GPU or professional hardware is required.
Which devices can run ExtGemma4 40 5B?
30 devices with unified memory can run ExtGemma4 40 5B at BF16 (80.9 GB), including ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB), GMKtec EVO-X2 (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.