ai.canshu.com
中文
VRAM Calculator Advisor & TCO API Pricing Compare Models Leaderboard GPUs

Search models, GPUs, and tools

↑↓ Navigate · Enter Open · ★ Watchlist prioritized ai.canshu.com
Q4_K_M: Smooth Fit 17.92 GB / 61.4 GB Usable

Gemma-4-26B-A4B-IT (MoE) on 华为昇腾 Ascend 910B 64GB (HBM2e): VRAM Requirement & Speed Estimate

Yes. At Q4_K_M 4-bit quantization with a 16K context window, Gemma-4-26B-A4B-IT (MoE) requires 17.92 GB of VRAM (15.76 GB weights + 0.55 GB KV cache + 1.62 GB overhead), fitting within 昇腾 910B 64G's 61.4 GB usable memory at ~320.3 tokens/s.

Concurrency (Batch):
MoE 26B / A4B·30L (d=2816)·16Q/2KV·Ctx 256K·Text+Image+Video
Runnable Models →
GPUs
64 GB VRAM·1200 GB/s·Safe Usable 61.4 GB
Smooth Fit17.92 / 61.4 GB~320.3 tok/s
3. Weight Quantization
4. Per-Request Context16K (16,384 tok)
Verified: 2026-11-20·Roofline bandwidth/compute estimate (Batch=1)
Smooth Fit
17.92/ 61.4 GB usable

Physical 64 GB · Load 29.2%

Theoretical Decode Speed
~320.3tokens/s

MoE: reads only 2.42 GB active weights/step

Max Context at Batch=1
256K/ 256K

Enable Q8/Q4 KV quantization to scale batch concurrency

Weights: 15.76 GBKV Cache (1×16K): 0.55 GBOverhead: 1.62 GB
🚀 Blazing Fast (Instant CoT)~1.6s for a 500-token response
Reference Launch CLI:
OLLAMA_CONTEXT_LENGTH=16384 ollama run gemma4:26b-a4b

昇腾 910B 64G Fits Gemma-4-26B-A4B-IT (MoE) · Cost & TCO

Approx. $13,200 (~$206/GB VRAM) · 310W TDP

📊 All Quantization Levels for Gemma-4-26B-A4B-IT (MoE) on 昇腾 910B 64G

Quantization Total VRAM (16K) Verdict Est. Speed Max Context
FP16 / BF16 54.89 GB Tight Fit (Watch Context) ~96.1 tok/s 256K
FP8 / Q8_0 30.02 GB Smooth Fit ~191 tok/s 256K
GGUF Q6_K 23.59 GB Smooth Fit ~243.2 tok/s 256K
GGUF Q5_K_M 20.71 GB Smooth Fit ~277.1 tok/s 256K
Q4_K_M / AWQ 17.92 GB Smooth Fit ~320.3 tok/s 256K
GGUF IQ4_XS 15.93 GB Smooth Fit ~360.4 tok/s 256K
GGUF Q3_K_M 14.81 GB Smooth Fit ~387.9 tok/s 256K
GGUF Q2_K 12.29 GB Smooth Fit ~420 tok/s 256K

💡 Recommended Alternative GPUs & Macs for Gemma-4-26B-A4B-IT (MoE)

Frequently Asked Questions

Can 华为昇腾 Ascend 910B 64GB (HBM2e) run Gemma-4-26B-A4B-IT (MoE) locally?

Yes. At Q4_K_M 4-bit quantization with a 16K context window, Gemma-4-26B-A4B-IT (MoE) requires 17.92 GB of VRAM (15.76 GB weights + 0.55 GB KV cache + 1.62 GB overhead), fitting within 昇腾 910B 64G's 61.4 GB usable memory at ~320.3 tokens/s.

What is the best quantization for Gemma-4-26B-A4B-IT (MoE) on 昇腾 910B 64G?

The highest-quality quantization that fits cleanly on 昇腾 910B 64G is FP16 / BF16 (Lossless Full Precision), using 54.89 GB VRAM at ~96.1 tokens/s with up to 256K context.

How much VRAM does Gemma-4-26B-A4B-IT (MoE) need in full FP16 precision?

In unquantized FP16/BF16 precision (16K context), Gemma-4-26B-A4B-IT (MoE) requires 54.89 GB VRAM (52 GB for weights alone). Because it uses a Mixture-of-Experts (MoE) architecture, all 26B parameters must sit in memory, while only 4B active parameters are read per token.