MODEL ARCHITECTURE SPEC • ALIBABA CLOUD235B MoE
Qwen 3.8 235B-A22B MoE VRAM & inference speed
235B Total Params22B Active (MoE)94 Layers • 4 KV Heads32K Default Ctx (Max 128K)Ollama: qwen3.8:235b-a22b
Accurate memory requirements, KV cache footprint, and tokens/sec generation benchmarks for running Qwen 3.8 235B-A22B MoE across local GPUs, Apple Silicon Macs, and cloud instances.
✓ FITS IN VRAM149.84 GB (69%)~37 t/s
Qwen 3.8 235B-A22B MoE on Apple Mac Studio M3 Ultra (256GB Unified Memory)MoE (22B Active / 235B Total)
✓ Fits Comfortably in VRAM
✓ FITS IN VRAM
Required VRAM
149.84 GB
69% of hardware capacity (218 GB)Est. Decode Speed
~37 tok/s
22B active params @ 819 GB/sAvailable Headroom
+68.16 GB
Safe for OS display & full contextMemory Allocation & Headroom
149.84 GB allocated of 218 GB capacity
69% utilized
Weights: 141 GBMoE (22B Active)
KV: (32K): 3.15 GB
Overhead: 5.69 GB
Free Headroom: +68.16 GB
Architecture & Throughput Transparency: Single-user decode speed (~37 tok/s) is memory-bandwidth bound at batch size = 1 on Apple Mac Studio M3 Ultra (256GB Unified Memory) (819 GB/s). Prompt prefill / TTFT (~537 tok/s) is compute-bound and utilizes GPU Tensor cores. Actual latency varies with FlashAttention-2 and runtime engine (Ollama/vLLM/llama.cpp).
MoE Memory vs Speed Architecture: While decode throughput scales with the active parameter count (22B parameters streamed per token), all 235B parameters must be resident in VRAM so the dynamic router can route tokens to any expert layer.
Cloud Instance Options for Qwen 3.8 235B-A22B MoE
On-DemandDeploy this model on an on-demand cloud GPU without local hardware constraints or power usage:
RunPod Cloud GPUsBest 1-Click Ollama & vLLM
$0.38/hr (RTX 4090 Spot) – $2.39/hr (H100)Vast.ai GPU MarketplaceCheapest Dual RTX 3090 / 4090
$0.18/hr (RTX 3060) – $0.78/hr (2x RTX 3090)Lambda GPU CloudEnterprise 8x H100 Clusters
$2.49/hr (H100) – $19.92/hr (8x H100)Quantization Precision Matrix
32K ctx • KV: Q8_0| Quantization | Weights | KV Cache | Total VRAM | Est. Speed | Fit Status |
|---|---|---|---|---|---|
| FP1616 bpw | 470 GB | 3.15 GB | 490.35 GB (225%) | 0 t/s | OOM |
| Q8_08.5 bpw | 249.1 GB | 3.15 GB | 261.72 GB (120%) | ~5 t/s | PARTIAL |
| Q6_K6.56 bpw | 192.7 GB | 3.15 GB | 203.34 GB (93%) | ~28 t/s | FITS |
| Q5_K_M5.68 bpw | 166.85 GB | 3.15 GB | 176.59 GB (81%) | ~32 t/s | FITS |
| Q4_K_M4.8 bpw | 141 GB | 3.15 GB | 149.84 GB (69%) | ~37 t/s | FITS |
| FP44.25 bpw | 124.55 GB | 3.15 GB | 132.81 GB (61%) | ~42 t/s | FITS |
| IQ3_XS3.3 bpw | 96.35 GB | 3.15 GB | 103.62 GB (48%) | ~53 t/s | FITS |
Export Full Architecture & Infrastructure Spec
Download an instant Markdown architecture spec with exact memory math, hardware benchmarks, and runtime CLI flags.
Hardware Compatibility Matrix for Qwen 3.8 235B-A22B MoE
Standard Q4_K_M precision with 32K context window and Q8_0 KV cache
| Hardware System | VRAM Envelope | Required VRAM | Headroom | Est. Decode Speed | Compatibility | Test Fit |
|---|---|---|---|---|---|---|
| NVIDIA GeForce RTX 5090 (32GB GDDR7) Consumer GPU • 1792 GB/s | 32 GB | 149.84 GB | -117.84 GB | 0 (OOM) | ✕ OOM | Test Fit → |
| NVIDIA GeForce RTX 4090 (24GB GDDR6X) Consumer GPU • 1008 GB/s | 24 GB | 149.84 GB | -125.84 GB | 0 (OOM) | ✕ OOM | Test Fit → |
| 2x NVIDIA RTX 3090 Dual Rig (48GB Total VRAM) Multi-GPU Rig • 1648 GB/s | 48 GB | 150.54 GB | -102.54 GB | 0 (OOM) | ✕ OOM | Test Fit → |
| NVIDIA GeForce RTX 5080 (16GB GDDR7) Consumer GPU • 960 GB/s | 16 GB | 149.84 GB | -133.84 GB | 0 (OOM) | ✕ OOM | Test Fit → |
| NVIDIA RTX 4070 Ti Super (16GB) Consumer GPU • 672 GB/s | 16 GB | 149.84 GB | -133.84 GB | 0 (OOM) | ✕ OOM | Test Fit → |
| NVIDIA GeForce RTX 3060 (12GB) Consumer GPU • 360 GB/s | 12 GB | 149.84 GB | -137.84 GB | 0 (OOM) | ✕ OOM | Test Fit → |
| Apple MacBook Pro/Air M1 (8GB Unified Memory) Apple Silicon • 68 GB/s | 7 GB | 149.84 GB | -142.84 GB | 0 (OOM) | ✕ OOM | Test Fit → |
| Apple MacBook Pro/Air (16GB Unified Memory) Apple Silicon • 100 GB/s | 14 GB | 149.84 GB | -135.84 GB | 0 (OOM) | ✕ OOM | Test Fit → |
| Apple MacBook Pro M4 Pro (24GB Unified Memory) Apple Silicon • 273 GB/s | 20 GB | 149.84 GB | -129.84 GB | 0 (OOM) | ✕ OOM | Test Fit → |
| Apple Mac Mini M4 Pro (64GB Unified Memory) Apple Silicon • 273 GB/s | 54 GB | 149.84 GB | -95.84 GB | 0 (OOM) | ✕ OOM | Test Fit → |
| Apple Mac Studio M4 Max (128GB Unified Memory) Apple Silicon • 546 GB/s | 112 GB | 149.84 GB | -37.84 GB | ~5 tok/s | ! PARTIAL | Test Fit → |
| Apple Mac Studio M3 Ultra (256GB Unified Memory) Apple Silicon • 819 GB/s | 236 GB | 149.84 GB | +86.16 GB | ~37 tok/s | ✓ FITS | Test Fit → |
| AMD Strix Halo Mini-PC (Ryzen AI Max+ 395 128GB) Unified APU / Mini-PC • 256 GB/s | 110 GB | 149.84 GB | -39.84 GB | 0 (OOM) | ✕ OOM | Test Fit → |
| NVIDIA DGX Spark (128GB Grace-Blackwell Unified) AI Appliance • 273 GB/s | 118 GB | 149.84 GB | -31.84 GB | ~3 tok/s | ! PARTIAL | Test Fit → |
| NVIDIA H100 SXM (80GB HBM3 Cloud GPU) Datacenter Cloud GPU • 3350 GB/s | 80 GB | 149.84 GB | -69.84 GB | 0 (OOM) | ✕ OOM | Test Fit → |
| NVIDIA GeForce RTX 5070 Ti (16GB GDDR7) Consumer GPU • 896 GB/s | 16 GB | 149.84 GB | -133.84 GB | 0 (OOM) | ✕ OOM | Test Fit → |
| NVIDIA GeForce RTX 5070 (12GB GDDR7) Consumer GPU • 672 GB/s | 12 GB | 149.84 GB | -137.84 GB | 0 (OOM) | ✕ OOM | Test Fit → |
| NVIDIA GeForce RTX 4080 Super (16GB GDDR6X) Consumer GPU • 736 GB/s | 16 GB | 149.84 GB | -133.84 GB | 0 (OOM) | ✕ OOM | Test Fit → |
| AMD Radeon RX 7900 XTX (24GB GDDR6) Consumer GPU • 960 GB/s | 24 GB | 149.84 GB | -125.84 GB | 0 (OOM) | ✕ OOM | Test Fit → |
| AMD Radeon RX 7900 XT (20GB GDDR6) Consumer GPU • 800 GB/s | 20 GB | 149.84 GB | -129.84 GB | 0 (OOM) | ✕ OOM | Test Fit → |
| Apple Mac Studio M2 Ultra (192GB Unified Memory) Apple Silicon Workstation • 800 GB/s | 144 GB | 149.84 GB | -5.84 GB | ~8 tok/s | ! PARTIAL | Test Fit → |