Supergemma4 26B Uncensored MLX 4bit v2 — Hardware Requirements & GPU Compatibility
ChatReasoningFunctions
Supergemma4 26B Uncensored MLX 4bit v2 is a 25.2B-parameter open language model from Jiunsong in the Gemma family. It supports a context window of up to 262,144 tokens. At BF16 it needs about 51.11 GB of VRAM — see which GPUs and Macs can run it below.
20.8K downloads
254 likesSpecifications
- Publisher
- Jiunsong
- Family
- Gemma
- Parameters
- 25.2B
- Architecture
- Gemma4ForConditionalGeneration
- Context Length
- 262,144 tokens
- Vocabulary Size
- 262,144
- Release Date
- 2026-04-12
- License
- Gemma Terms
Get Started
How Much VRAM Does Supergemma4 26B Uncensored MLX 4bit v2 Need?
Select a quantization to see compatible GPUs below.
| Quantization | Bits | VRAM | + Context | File Size | Quality |
|---|---|---|---|---|---|
| BF16 | 16.00 | 51.1 GB | 95.1 GB | 50.47 GB | Brain floating point 16 — preferred for training |
Which GPUs Can Run Supergemma4 26B Uncensored MLX 4bit v2?
BF16 · 51.1 GBSupergemma4 26B Uncensored MLX 4bit v2 (BF16) requires 51.1 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 67+ GB is recommended. Using the full 262K context window can add up to 44.0 GB, bringing total usage to 95.1 GB. No single GPU has enough memory — multi-GPU or cluster setups are needed.
Which Devices Can Run Supergemma4 26B Uncensored MLX 4bit v2?
BF16 · 51.1 GB8 devices with unified memory can run Supergemma4 26B Uncensored MLX 4bit v2, including NVIDIA DGX H100, NVIDIA DGX A100 640GB, Mac Studio M4 Max (64 GB).
Runs great
— Plenty of headroom
Decent
— Enough memory, may be tightRelated Models
Frequently Asked Questions
- How much VRAM does Supergemma4 26B Uncensored MLX 4bit v2 need?
Supergemma4 26B Uncensored MLX 4bit v2 requires 51.1 GB of VRAM at BF16. Full 262K context adds up to 44.0 GB (95.1 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 25.2B × 16 bits ÷ 8 = 50.5 GB
KV Cache + Overhead ≈ 0.6 GB (at 2K context + ~0.3 GB framework)
KV Cache + Overhead ≈ 44.6 GB (at full 262K context)
VRAM usage by quantization
BF1651.1 GBBF16 + full context95.1 GB- Can NVIDIA GeForce RTX 5090 run Supergemma4 26B Uncensored MLX 4bit v2?
No — Supergemma4 26B Uncensored MLX 4bit v2 requires at least 51.1 GB at BF16, which exceeds the NVIDIA GeForce RTX 5090's 32 GB of VRAM.
- Can I run Supergemma4 26B Uncensored MLX 4bit v2 on a Mac?
Supergemma4 26B Uncensored MLX 4bit v2 requires at least 51.1 GB at BF16, which exceeds the unified memory of most consumer Macs. You would need a Mac Studio or Mac Pro with a high-memory configuration.
- Can I run Supergemma4 26B Uncensored MLX 4bit v2 locally?
Yes — Supergemma4 26B Uncensored MLX 4bit v2 can run locally on consumer hardware. At BF16 quantization it needs 51.1 GB of VRAM. Popular tools include Ollama, LM Studio, and llama.cpp.
- How fast is Supergemma4 26B Uncensored MLX 4bit v2?
At BF16, Supergemma4 26B Uncensored MLX 4bit v2 can reach ~57 tok/s on AMD Instinct MI300X. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = (bandwidth GB/s ÷ model GB) × efficiency
Example: AMD Instinct MI300X → 5300 ÷ 51.1 × 0.55 = ~57 tok/s
Estimated speed at BF16 (51.1 GB)
~57 tok/s~43 tok/s~35 tok/sReal-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
- What's the download size of Supergemma4 26B Uncensored MLX 4bit v2?
At BF16, the download is about 50.47 GB.
- Which GPUs can run Supergemma4 26B Uncensored MLX 4bit v2?
No single consumer GPU has enough VRAM to run Supergemma4 26B Uncensored MLX 4bit v2 at BF16 (51.1 GB). Multi-GPU or professional hardware is required.
- Which devices can run Supergemma4 26B Uncensored MLX 4bit v2?
8 devices with unified memory can run Supergemma4 26B Uncensored MLX 4bit v2 at BF16 (51.1 GB), including Mac Pro M2 Ultra (192 GB), Mac Studio M2 Ultra (192 GB), Mac Studio M4 Max (128 GB), Mac Studio M4 Max (64 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.