Qwen
ChatCodingAI codingReasoningMulti-purpose397B
Chat

Qwen3.5-397B (MoE)

Qwen · Apache 2.0

Mixture of Experts: 397B total parameters, 30B active per token.

Frontier-scale MoE with 397B total parameters and roughly 30B active per token. Requires an enormous memory footprint even when quantized, so local runs are limited to GB300-class hardware or aggressive offloading setups.

Parameters
397B
Architecture
MoE (30B active)
Context
262,144 tokens
Released
2026-02-24
Engines
llama.cpp, vLLM
Builder Tools
Continue, LM Studio, Open WebUI

Parameters

397B

VRAM

230 GB

Context

256K

Formats

3

GPUs

43

Qwen3.5-397B (MoE) (397B) requires 230 GB VRAM at recommended quality (Q4_K_M). On NVIDIA Grace Blackwell Ultra GB300, expect approximately 120 tok/s at Q4_K_M.

Source: OwnRig methodology

VRAM (Recommended)

230 GB

Quantization

Q4_K_M

File Size

220 GB

Max Context

256K tokens

Primary Use

Chat

Memory

VRAM Requirements

QualityQuantizationVRAMFile Size
recommendedQ4_K_M230 GB220 GB
efficientQ3_K_M175 GB168 GB
compressedQ2_K140 GB134 GB
Scaling

Context Length Impact

KV cache VRAM at Q4_K_M quality. Longer context = more memory.

ContextKV CacheTotal VRAM
2K819 MB230.8 GBexceeds 24 GB
4K1.6 GB231.6 GBexceeds 24 GB
8K3.2 GB233.2 GBexceeds 24 GB
16K6.4 GB236.4 GBexceeds 24 GB
32K12.8 GB242.8 GBexceeds 24 GB
64K25.6 GB255.6 GBexceeds 24 GB
128K51.2 GB281.2 GBexceeds 24 GB

Compatible GPUs

43 devices
NVIDIA Grace Blackwell Ultra GB300Q4_K_M120 tok/sExcellent
Apple M4 Ultra (192GB)Q3_K_M44 tok/sGood
Apple M4 Max (128GB Unified)Q2_K8 tok/sMarginal
Apple M3 Pro (18GB Unified)Q2_KNot viable
Apple M4 (16GB Unified)Q2_KNot viable
Apple M4 Max (36GB Unified)Q2_KNot viable
Apple M4 Max (64GB Unified)Q2_KNot viable
Apple M4 Pro (24GB Unified)Q2_KNot viable
Apple M4 Pro (48GB)Q2_KNot viable
NVIDIA GeForce RTX 3060 12GBQ2_KNot viable
NVIDIA GeForce RTX 3080 10GBQ2_KNot viable
NVIDIA GeForce RTX 3090Q2_KNot viable
NVIDIA GeForce RTX 4060 8GBQ2_KNot viable
NVIDIA RTX 4060 Laptop (40-60W)Q2_KNot viable
NVIDIA GeForce RTX 4060 Ti 16GBQ2_KNot viable
NVIDIA RTX 4070 Laptop (80-115W)Q2_KNot viable
NVIDIA GeForce RTX 4070 SuperQ2_KNot viable
NVIDIA GeForce RTX 4070 Ti 12GBQ2_KNot viable
NVIDIA GeForce RTX 4070 Ti SuperQ2_KNot viable
NVIDIA RTX 4080 Laptop (120-150W)Q2_KNot viable
NVIDIA GeForce RTX 4080 SuperQ2_KNot viable
NVIDIA RTX 4090 Laptop (150-175W)Q2_KNot viable
NVIDIA GeForce RTX 4090Q2_KNot viable
NVIDIA GeForce RTX 5080Q2_KNot viable
NVIDIA GeForce RTX 5090Q2_KNot viable
AMD Radeon RX 7600Q2_KNot viable
AMD Radeon RX 7900 XTXQ2_KNot viable
AMD Radeon Pro W7900Q2_KNot viable
NVIDIA RTX PRO 6000 BlackwellQ2_KNot viable
NVIDIA RTX PRO 6000 Blackwell Max-QQ2_KNot viable
AMD Radeon RX 9070Q2_KNot viable
Apple M1 (8GB Unified)Q2_KNot viable
Apple M1 (16GB Unified)Q2_KNot viable
Apple M1 Pro (16GB Unified)Q2_KNot viable
Apple M2 (8GB Unified)Q2_KNot viable
Apple M2 (16GB Unified)Q2_KNot viable
Apple M2 Pro (16GB Unified)Q2_KNot viable
Apple M3 (8GB Unified)Q2_KNot viable
Apple M3 (16GB Unified)Q2_KNot viable
AMD Radeon RX 9060 XT 16GBQ2_KNot viable
AMD Radeon RX 9060 XT 8GBQ2_KNot viable
NVIDIA GeForce RTX 5060 8GBQ2_KNot viable
NVIDIA GeForce RTX 5060 Ti 16GBQ2_KNot viable

Showing 43 of 43 entries

Builder Context

Qwen3.5-397B (MoE) is commonly used with Continue, LM Studio, Open WebUI. For an AI coding workflow, pair it with an embedding model like nomic-embed-text for local RAG.

FAQ

Frequently Asked Questions

How much VRAM does Qwen3.5-397B (MoE) need?
Qwen3.5-397B (MoE) requires 230 GB VRAM at recommended quality (Q4_K_M). At lower quality settings, it can fit in as little as 140 GB.
What is the best GPU for Qwen3.5-397B (MoE)?
The NVIDIA Grace Blackwell Ultra GB300 delivers the best performance for Qwen3.5-397B (MoE), achieving 120 tok/s at Q4_K_M with an excellent rating.
Can I run Qwen3.5-397B (MoE) on an RTX 4060 Ti?
Qwen3.5-397B (MoE) at Q2_K requires 230 GB VRAM, which exceeds the RTX 4060 Ti's 16 GB. Consider a lower quantization or a GPU with more VRAM.
What quantization should I use for Qwen3.5-397B (MoE)?
For the best quality, use Q4_K_M (230 GB VRAM). If your GPU has limited VRAM, Q2_K (140 GB) is the most efficient option with acceptable quality.
Is Qwen3.5-397B (MoE) good for coding?
Yes. Qwen3.5-397B (MoE) is used with Continue, LM Studio, Open WebUI for local AI coding. For the best coding experience, pair it with an embedding model for local RAG.
All models

Data confidence: estimated. Source

VRAM requirements are calculated from model parameters and may vary by inference engine, context length, and batch size. Performance estimates are based on community benchmarks and should be verified for your specific configuration.Qwen is a trademark of its respective owner. OwnRig is not affiliated with or endorsed by the model creator.