Alibaba QwenGeneralLicense: Apache 2.0
Qwen3.8 27B
Alibaba's newest open mid-size model and one of the strongest dense models you can run on a single 24GB GPU or 32GB+ Mac.
Qwen3.8-27B (released August 2026) is a dense 27.8B-parameter model with a 256k context window and a hybrid attention design, where only a quarter of its layers use full attention, which keeps long-context memory use down. It is a strong all-rounder for coding, reasoning and writing.
Parameter Count27.8B
Context Window256k tokens
Min VRAM (Q4)18 GB
Recommended RAM24 GB+
⚡ Key Architectural Strengths
✓Dense 27.8B model with 256k context
✓Hybrid attention keeps the KV cache smaller at long context
✓Apache 2.0 licence
VRAM & Quantization Memory Footprint
Exact file weights, minimum VRAM to boot, and recommended headroom for extended context windows.
| Quantization Tier | Weight Size | Min VRAM/RAM | Recommended VRAM | Recommended Use Case |
|---|---|---|---|---|
| Q4_K_M (Recommended) | 16.7 GB | 18.2 GB | 20 GB | 4-bit weights. The usual size/quality sweet spot for local use. |
| Q8_0 (High Precision) | 29.5 GB | 31.5 GB | 33.9 GB | 8-bit weights. Close to full quality at about half the size of BF16. |
| BF16 / FP16 (Unquantized) | 55.6 GB | 58.6 GB | 62.2 GB | Unquantized weights, mainly for fine-tuning or GPU servers. |
⚡ Need to calculate exact KV Cache with 16k or 32k context?
Simulate exact VRAM allocations across various context lengths and precision depths.
1-Click Local Run Commands
Copy and paste directly into your terminal or local server engine.
Ollama CLI:
ollama run qwen3.8:27bLM Studio CLI:
lms get qwen3.8-27bvLLM Production Server:
vllm serve Qwen/Qwen3.8-27Bllama.cpp Binary:
./llama-server -m Qwen3.8-27B-Q4_K_M.gguf -c 16384Hardware Compatibility & Speed Estimates
Estimated against 55+ Apple Silicon, Qualcomm, and NVIDIA hardware profiles.
| Device / Hardware Name | Platform | RAM/VRAM | Fit Status | Est. Tokens/Sec | Notes |
|---|---|---|---|---|---|
| Apple MacBook Air 13-inch (M4, 2025) | macOS | 16 GB | Unsupported | — | Not enough memory to load the 4-bit weights (~16.7GB) plus overhead. |
| HP Omnibook 5 OLED | Windows ARM | 32 GB | Tight Fit | 4.7 t/s | Fits with a modest context (4k-8k). Close other memory-heavy apps. Speed is an estimate. |
| Apple MacBook Pro 16-inch (M4 Max) | macOS | 64 GB | Comfortable | 21.3 t/s | Fits fully in GPU-usable memory (~48GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Apple Mac mini (M4, 16GB) | macOS | 16 GB | Unsupported | — | Not enough memory to load the 4-bit weights (~16.7GB) plus overhead. |
| Apple Mac mini (M4 Pro, 24GB) | macOS | 24 GB | Unsupported | — | Not enough memory to load the 4-bit weights (~16.7GB) plus overhead. |
| Apple Mac Studio (M2 Ultra 192GB) | macOS | 192 GB | Comfortable | 31.1 t/s | Fits fully in GPU-usable memory (~144GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Microsoft Surface Laptop 7 | Windows ARM | 16 GB | Unsupported | — | Not enough memory to load the 4-bit weights (~16.7GB) plus overhead. |
| Lenovo ThinkPad T14s Gen 5 (Snapdragon) | Windows ARM | 32 GB | Tight Fit | 4.7 t/s | Fits with a modest context (4k-8k). Close other memory-heavy apps. Speed is an estimate. |
| ASUS Zenbook S 16 (Ryzen AI 9 HX 370) | Windows x86 | 32 GB | Tight Fit | 4.2 t/s | Fits with a modest context (4k-8k). Close other memory-heavy apps. Speed is an estimate. |
| Dell XPS 13 (Intel Core Ultra 7 258V) | Windows x86 | 32 GB | Tight Fit | 4.8 t/s | Fits with a modest context (4k-8k). Close other memory-heavy apps. Speed is an estimate. |
| ASUS ROG Zephyrus G16 (RTX 4070 Laptop) | Windows x86 | 32 GB | CPU Offload | 3.5 t/s | Exceeds the 8GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| Razer Blade 16 (RTX 4090 Mobile 16GB) | Windows x86 | 32 GB | CPU Offload | 3.5 t/s | Exceeds the 16GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| Framework Laptop 13 (Ryzen 7 7840U) | Linux | 64 GB | Comfortable | 3.5 t/s | Fits fully in GPU-usable memory (~38.4GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Lenovo Legion Pro 7i (RTX 4080 12GB) | Windows x86 | 32 GB | CPU Offload | 3.5 t/s | Exceeds the 12GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| Dell XPS 14 (Core Ultra 7) | Windows x86 | 16 GB | Unsupported | — | Not enough memory to load the 4-bit weights (~16.7GB) plus overhead. |
| Lenovo ThinkPad E14 (Budget Setup) | Windows x86 | 8 GB | Unsupported | — | Not enough memory to load the 4-bit weights (~16.7GB) plus overhead. |
| Acer Swift 14 AI (Snapdragon X Plus) | Windows ARM | 16 GB | Unsupported | — | Not enough memory to load the 4-bit weights (~16.7GB) plus overhead. |
| Custom Workstation PC (RTX 4090 24GB) | Linux | 64 GB | Comfortable | 39.2 t/s | Fits fully in VRAM (~24GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Legacy Budget Desktop (GTX 1650 4GB) | Windows x86 | 16 GB | CPU Offload | 1.7 t/s | Exceeds the 4GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| Apple MacBook Pro 16-inch (M4 Max 128GB) | macOS | 128 GB | Comfortable | 21.3 t/s | Fits fully in GPU-usable memory (~96GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Apple MacBook Pro 14-inch (M4 Pro 48GB) | macOS | 48 GB | Comfortable | 10.6 t/s | Fits fully in GPU-usable memory (~36GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Apple MacBook Pro 14-inch (M4 Base 24GB) | macOS | 24 GB | Unsupported | — | Not enough memory to load the 4-bit weights (~16.7GB) plus overhead. |
| Apple MacBook Air 15-inch (M3 16GB) | macOS | 16 GB | Unsupported | — | Not enough memory to load the 4-bit weights (~16.7GB) plus overhead. |
| Apple MacBook Air 13-inch (M3 8GB Base) | macOS | 8 GB | Unsupported | — | Not enough memory to load the 4-bit weights (~16.7GB) plus overhead. |
| Apple MacBook Air 13-inch (M2 16GB) | macOS | 16 GB | Unsupported | — | Not enough memory to load the 4-bit weights (~16.7GB) plus overhead. |
| Apple MacBook Air (M1 16GB Legacy) | macOS | 16 GB | Unsupported | — | Not enough memory to load the 4-bit weights (~16.7GB) plus overhead. |
| Apple MacBook Pro 16-inch (M3 Max 128GB) | macOS | 128 GB | Comfortable | 15.6 t/s | Fits fully in GPU-usable memory (~96GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Apple MacBook Pro 14-inch (M3 Pro 36GB) | macOS | 36 GB | Comfortable | 5.8 t/s | Fits fully in GPU-usable memory (~24.1GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Apple Mac Studio (M1 Ultra 64GB) | macOS | 64 GB | Comfortable | 31.1 t/s | Fits fully in GPU-usable memory (~48GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Samsung Galaxy Book4 Edge (Snapdragon X Elite) | Windows ARM | 16 GB | Unsupported | — | Not enough memory to load the 4-bit weights (~16.7GB) plus overhead. |
| Lenovo Yoga Slim 7x Copilot+ | Windows ARM | 32 GB | Tight Fit | 4.7 t/s | Fits with a modest context (4k-8k). Close other memory-heavy apps. Speed is an estimate. |
| Microsoft Surface Pro 11 Copilot+ (OLED) | Windows ARM | 32 GB | Tight Fit | 4.7 t/s | Fits with a modest context (4k-8k). Close other memory-heavy apps. Speed is an estimate. |
| Dell XPS 13 Copilot+ (Snapdragon X Elite) | Windows ARM | 16 GB | Unsupported | — | Not enough memory to load the 4-bit weights (~16.7GB) plus overhead. |
| Custom Workstation PC (RTX 4080 Super 16GB) | Linux / Windows | 32 GB | CPU Offload | 3.7 t/s | Exceeds the 16GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| Custom AI Workstation (RTX 4070 Ti Super 16GB) | Linux / Windows | 32 GB | CPU Offload | 3.7 t/s | Exceeds the 16GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| Desktop PC (RTX 4070 Super 12GB) | Windows / Linux | 32 GB | CPU Offload | 3.5 t/s | Exceeds the 12GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| Budget AI Workstation (RTX 4060 Ti 16GB) | Linux / Windows | 32 GB | CPU Offload | 2 t/s | Exceeds the 16GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| Mainstream Desktop (RTX 4060 8GB) | Windows / Linux | 32 GB | CPU Offload | 2 t/s | Exceeds the 8GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| Used AI Powerhouse PC (RTX 3090 24GB) | Linux / Windows | 64 GB | Comfortable | 36.4 t/s | Fits fully in VRAM (~24GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Budget Workhorse PC (RTX 3060 12GB) | Windows / Linux | 16 GB | CPU Offload | 2 t/s | Exceeds the 12GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| Lenovo Legion 5i (RTX 4060 Laptop 8GB) | Windows x86 | 16 GB | CPU Offload | 3.5 t/s | Exceeds the 8GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| ASUS TUF Gaming A15 (RTX 4060 Laptop 8GB) | Windows x86 | 16 GB | CPU Offload | 3.5 t/s | Exceeds the 8GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| Acer Predator Helios Neo 16 (RTX 4070 Laptop) | Windows x86 | 32 GB | CPU Offload | 3.5 t/s | Exceeds the 8GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| HP Victus 16 (RTX 4050 Laptop 6GB) | Windows x86 | 16 GB | CPU Offload | 3.5 t/s | Exceeds the 6GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| Dell Alienware m16 R2 (RTX 4070 8GB) | Windows x86 | 32 GB | CPU Offload | 3.5 t/s | Exceeds the 8GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| ASUS ROG Zephyrus G16 (AMD AI 9 + RTX 4070) | Windows x86 | 32 GB | CPU Offload | 4.7 t/s | Exceeds the 8GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| Lenovo IdeaPad Pro 5 (Ryzen 7 8845HS) | Windows / Linux | 32 GB | Tight Fit | 3.6 t/s | Fits with a modest context (4k-8k). Close other memory-heavy apps. Speed is an estimate. |
| Custom Workstation PC (Radeon RX 7900 XTX 24GB) | Linux / Windows | 64 GB | Comfortable | 37.4 t/s | Fits fully in VRAM (~24GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Desktop PC (Radeon RX 7800 XT 16GB) | Linux / Windows | 32 GB | CPU Offload | 2.2 t/s | Exceeds the 16GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| Budget Workstation (Radeon RX 7600 XT 16GB) | Linux / Windows | 32 GB | CPU Offload | 2 t/s | Exceeds the 16GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| ASUS Zenbook S 14 (Lunar Lake Core Ultra 7) | Windows x86 | 32 GB | Tight Fit | 4.8 t/s | Fits with a modest context (4k-8k). Close other memory-heavy apps. Speed is an estimate. |
| Lenovo Yoga Slim 7i Aura (Core Ultra 7 256V) | Windows x86 | 16 GB | Unsupported | — | Not enough memory to load the 4-bit weights (~16.7GB) plus overhead. |
| Custom PC (Intel Arc A770 16GB) | Linux / Windows | 32 GB | CPU Offload | 3.5 t/s | Exceeds the 16GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| Homelab AI Server (Dual RTX 3090 48GB VRAM) | Linux (Ubuntu Server) | 128 GB | Comfortable | 36.4 t/s | Fits fully in VRAM (~48GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Lenovo ThinkPad X1 Carbon Gen 12 | Windows x86 / Linux | 32 GB | Tight Fit | 4.2 t/s | Fits with a modest context (4k-8k). Close other memory-heavy apps. Speed is an estimate. |
| Flagship AI Workstation (RTX 5090 32GB) | Linux / Windows | 64 GB | Comfortable | 69.7 t/s | Fits fully in VRAM (~32GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| High-End AI Rig (RTX 5080 16GB) | Linux / Windows | 64 GB | CPU Offload | 3.7 t/s | Exceeds the 16GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| Next-Gen AI Laptop (Ryzen AI Max+ 395 128GB) | Linux / Windows | 128 GB | Comfortable | 10 t/s | Fits fully in GPU-usable memory (~96GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Apple Mac mini (M6, 16GB, 2026) | macOS | 16 GB | Unsupported | — | Not enough memory to load the 4-bit weights (~16.7GB) plus overhead. |
| Apple Mac mini (M6, 32GB, 2026) | macOS | 32 GB | Comfortable | 6.6 t/s | Fits fully in GPU-usable memory (~21.4GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Apple Mac mini (M5 Pro, 48GB, 2026) | macOS | 48 GB | Comfortable | 11.9 t/s | Fits fully in GPU-usable memory (~36GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Apple Mac mini (M5 Pro, 64GB, 2026) | macOS | 64 GB | Comfortable | 11.9 t/s | Fits fully in GPU-usable memory (~48GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Apple MacBook Pro 14-inch (M5, 32GB, 2025) | macOS | 32 GB | Comfortable | 6 t/s | Fits fully in GPU-usable memory (~21.4GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Apple Mac Studio (M5 Max, 128GB, 2026) | macOS | 128 GB | Comfortable | 23.9 t/s | Fits fully in GPU-usable memory (~96GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Apple Mac Studio (M5 Ultra, 256GB, 2026) | macOS | 256 GB | Comfortable | 46.7 t/s | Fits fully in GPU-usable memory (~192GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Apple Mac Studio (M5 Ultra, 512GB, 2026) | macOS | 512 GB | Comfortable | 46.7 t/s | Fits fully in GPU-usable memory (~384GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Apple Mac Studio (M3 Ultra, 512GB, 2025) | macOS | 512 GB | Comfortable | 31.9 t/s | Fits fully in GPU-usable memory (~384GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| NVIDIA DGX Spark (GB10, 128GB) | DGX OS (Linux) | 128 GB | Comfortable | 10.6 t/s | Fits fully in GPU-usable memory (~110GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Desktop with NVIDIA RTX 5070 Ti (16GB) | Windows / Linux | 32 GB | CPU Offload | 3.7 t/s | Exceeds the 16GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| Desktop with NVIDIA RTX 5060 Ti (16GB) | Windows / Linux | 32 GB | CPU Offload | 3.5 t/s | Exceeds the 16GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| Desktop with AMD Radeon RX 9070 XT (16GB) | Windows / Linux | 32 GB | CPU Offload | 3.7 t/s | Exceeds the 16GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| Desktop with AMD Radeon RX 9060 XT (16GB) | Windows / Linux | 32 GB | CPU Offload | 3.5 t/s | Exceeds the 16GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| Desktop with Intel Arc B580 (12GB) | Windows / Linux | 32 GB | CPU Offload | 3.5 t/s | Exceeds the 12GB of VRAM, so some layers run from system RAM, which is much slower. Estimate only. |
| ASUS Zenbook A16 (Snapdragon X2 Elite Extreme, 48GB) | Windows 11 (ARM) | 48 GB | Comfortable | 8.9 t/s | Fits fully in GPU-usable memory (~28.8GB) with room for a longer context. Speed is a bandwidth-based estimate. |
| Snapdragon X2 Elite Laptop (32GB) | Windows 11 (ARM) | 32 GB | Tight Fit | 5.3 t/s | Fits with a modest context (4k-8k). Close other memory-heavy apps. Speed is an estimate. |
Guides that cover Qwen3.8 27B
Step-by-step setup, memory sizing and quantization choices from the BFWR guide library.
- How to Run Qwen3.8 27B LocallyHardware, VRAM and step-by-step setup.Read the guide
- How to Run Qwen3.8-Flash-Next Locally180B with 6B active: which build fits 128GB, real speeds, licence.Read the guide
- Strata: Running Qwen3.8-Flash-Next on a 12GB Graphics CardThe engine that splits a 125B model across GPU, RAM and SSD.Read the guide
Not sure if Qwen3.8 27B fits your specific setup?
Our automated recommender calculates your exact hardware limits and suggests the best model weights for coding, reasoning, or fast conversational chat.