Qwen1.5 MoE A2.7B — Hardware Requirements & GPU Compatibility
Chat
Qwen1.5 MoE A2.7B is a Mixture of Experts (MoE) model from Alibaba Cloud's Qwen 1.5 generation, with 14.3 billion total parameters but only 2.7 billion active parameters per forward pass. The MoE architecture allows it to deliver performance closer to dense 7B models while requiring less compute during inference, as only a subset of expert layers are activated for each token. The model supports a 32K token context window and requires VRAM proportional to its total parameter count for loading, despite lower compute cost per token. It is an interesting architectural variant for users exploring efficient inference and MoE models locally. Released under a custom Qwen license.
Runs great on a 24 GB GPU
Runs from 8 GB (Q2_K · 8K)Best on 24 GB (Q4_K_M · ~251 tok/s · 8K)Macs from 16 GB (Q4_K_M · ~58 tok/s)
Other
Reference GPU · RTX 4090 24 GB
Estimates, ±20% · how we estimate
Run Qwen1.5 MoE A2.7B
One recommended way. Every other method, OS and quant is in the step-by-step guide.Each way opens a step-by-step guide for your OS and quant.
Recommended
ollama run hf.co/TULLUS/Qwen1.5-MoE-A2.7B-GGUF:Q4_K_MQ4_K_M fits fully on the RTX 4090 · 13.5 GB free · 8K context
Download: qwen1.5-moe-a2.7b-q4_k_m.gguf9.50 GBTULLUS
Ways to runOther ways
each opens step-by-stepRun in cloud
Runs on RTX A4000 (16 GB) (5 GB headroom) · Q4_K_M
- Generation speed
- ~137 tok/s
- generation speed
- Cost per 1M output tokens
- $0.18
- per 1M output tokens
How Much VRAM Does Qwen1.5 MoE A2.7B Need?
Select a quantization level; the listed hardware adheres to it.
| Quantization | VRAM | +32K | On RTX 4090 (reference) | |
|---|---|---|---|---|
| 5.5 GB | — | 2-bit · i-quant medium — Importance-weighted 2-bit, medium | Fits: 8K context · ~344 tok/s | |
| 6.8 GB | — | 2-bit · K-quant — 2-bit quantization with K-quant improvements | Fits: 8K context · ~306 tok/s | |
| 7.1 GB | — | 3-bit · i-quant medium — Importance-weighted 3-bit, medium | Fits: 8K context · ~297 tok/s | |
| 7.7 GB | — | 3-bit medium — 3-bit medium quantization | Fits: 8K context · ~284 tok/s | |
| 8.4 GB | — | 4-bit · i-quant compact — Importance-weighted 4-bit, compact | Fits: 8K context · ~268 tok/s | |
| 9.3 GB | — | 4-bit medium — 4-bit medium quantization — most popular sweet spot | Fits: 8K context · ~251 tok/s | |
| 10.9 GB | — | 5-bit medium — 5-bit medium quantization — good quality/size tradeoff | Fits: 8K context · ~225 tok/s | |
| 12.5 GB | — | 6-bit, very good — 6-bit quantization, very good quality | Fits: 8K context · ~204 tok/s | |
| 15.0 GB | — | 8-bit, near-lossless — 8-bit quantization, near-lossless | Fits: 8K context · ~178 tok/s | |
| 29.3 GB | — | 16-bit · bfloat — Brain floating point 16 — preferred for training | Does not fit: needs 29.3 GB |
Selected Q4_K_M
4-bit medium quantization10.5GBat 8K context
- Weights
- 8.6 GB
- KV cache
- 1.6 GB
- Overhead
- 0.3 GB
Fits RTX 4090 (reference) · 13.5 GB free · ~251 tok/s
≥50 tok/s needs ≥144 GB/s · RTX 5090 ~355, RTX 3090 Ti ~251
Which Hardware Can Run Qwen1.5 MoE A2.7B?
Q4_K_M · 9.3 GBQwen1.5 MoE A2.7B (Q4_K_M) requires 9.3 GB of VRAM to load the model weights. For comfortable inference with headroom for KV cache and system overhead, 13+ GB is recommended. Using the full 8K context window can add up to 1.2 GB, bringing total usage to 10.5 GB. 40 GPUs can run it, including NVIDIA GeForce RTX 5090, NVIDIA GeForce RTX 3090 Ti, NVIDIA GeForce RTX 3080 Ti. 150 devices with unified memory can run Qwen1.5 MoE A2.7B, including Mac Studio (M3 Ultra, 256GB), Mac Studio (M3 Ultra, 512GB), Mac Mini M2 Pro (16 GB).
12 GB
12 GB tier →fits · 1.5 GB free · up to 8K ctx
10 GPUs
RTX 3080 Ti~235Simulate output on RTX 3080 TiRTX 5070~188Simulate output on RTX 5070RTX 4070~150Simulate output on RTX 407020 GB
fits · 9.5 GB free · up to 8K ctx
1 GPU
RX 7900 XT~119Simulate output on RX 7900 XT32–40 GB
fits · 21.5 GB free · up to 8K ctx
1 GPU · 3 Mac classes · 4 device families
RTX 5090~355Simulate output on RTX 5090M5 Max36 GB~98Simulate output on Mac Studio M5 Max (36 GB)M532 GB~48Simulate output on MacBook Air 13" M5 (32 GB)Orin~69Simulate output on NVIDIA Jetson AGX Orin 32GBSnapdragon X Elite (X1E-80-100)~45Simulate output on Snapdragon X Elite Copilot+ PC
By VRAM tier · typical GPUs and Macs at Q4_K_M
Alternatives to Qwen1.5 MoE A2.7B
Qwen family · 28 models →Similar models
6| Model | Score | VRAM | tok/s | Why it is similar |
|---|---|---|---|---|
| Apriel 1.6 15B ThinkerServiceNow-AI | — | 13.2 GB | ~67 | same size |
| Phi 4 ReasoningMicrosoft | — | 12.5 GB | ~69 | same size |
| Phi 3 Medium 128k InstructMicrosoft | — | 12 GB | ~72 | same size |
| JiRackUltra 14BCMSManhattan | — | 12.4 GB | ~69 | same size |
| Nemotron Terminal 14BNVIDIA | — | 11.9 GB | ~69 | same size |
| Ministral 3 14B Reasoning 2512Mistral AI | — | 12 GB | ~72 | same size |
Specifications
- Publisher
- Alibaba
- Family
- Qwen
- Parameters
- 14.3B
- Active parameters
- 2.7B per token
- Architecture
- Qwen2MoeForCausalLM
- Context Length
- 8,192 tokens
- Vocabulary Size
- 151,936
- Release Date
- 2024-02-29
- License
- Other
Frequently Asked Questions
How much VRAM does Qwen1.5 MoE A2.7B need?
Qwen1.5 MoE A2.7B requires 9.3 GB of VRAM at Q4_K_M, or 29.3 GB at BF16. Full 8K context adds up to 1.2 GB (10.5 GB total).
VRAM = Weights + KV Cache + Overhead
Weights = 14.3B × 4.8 bits ÷ 8 = 8.6 GB
KV Cache + Overhead ≈ 0.7 GB (at 2K context + ~0.3 GB framework)
Fit ratings and hardware model lists check this model with room for a 16K-token context, which needs a little more memory.
KV Cache + Overhead ≈ 1.9 GB (at full 8K context)
VRAM usage by quantization
Can NVIDIA GeForce RTX 4090 run Qwen1.5 MoE A2.7B?
Yes, at Q8_0 (15.0 GB) or lower. Higher quantizations like BF16 (29.3 GB) exceed the NVIDIA GeForce RTX 4090's 24 GB.
What's the best quantization for Qwen1.5 MoE A2.7B?
For Qwen1.5 MoE A2.7B, Q4_K_M (9.3 GB) offers the best balance of quality and VRAM usage. Q5_K_M (10.9 GB) provides better quality if you have the VRAM. The smallest option is IQ2_M at 5.5 GB.
VRAM requirement by quantization
★ Recommended — best balance of quality and VRAM usage.
Can I run Qwen1.5 MoE A2.7B on a Mac?
Yes — Mac Mini M2 Pro (16 GB) and 133 other Macs can run Qwen1.5 MoE A2.7B. Apple Silicon uses unified memory, so the model shares RAM with the system. At Q4_K_M you need at least 9.3 GB of usable unified memory (RAM minus macOS overhead).
Can I run Qwen1.5 MoE A2.7B locally?
Yes — Qwen1.5 MoE A2.7B can run locally on consumer hardware. Popular tools include Ollama, LM Studio, and llama.cpp.
How fast is Qwen1.5 MoE A2.7B?
At Q4_K_M, Qwen1.5 MoE A2.7B can reach ~194 tok/s on AMD Instinct MI350X. On NVIDIA GeForce RTX 4090: ~251 tok/s. Speed depends mainly on GPU memory bandwidth. Real-world results typically within ±20%.
tok/s = 1000 ÷ (active GB ÷ (bandwidth GB/s × efficiency) × 1000 + layers × routing ms)
Mixture-of-Experts: just the active experts get read for each token, along with a fixed routing cost per layer.
Example: NVIDIA B200 → 1.7 GB active ÷ (8000 × 0.65) = 0.336 ms, plus 24 layers × 0.055 ms = 1.320 ms, so 1000 ÷ 1.656 ms = ~604 tok/s
Estimated speed at Q4_K_M (9.3 GB)
Real-world results typically within ±20%. Speed depends on batch size, quantization kernel, and software stack.
What's the download size of Qwen1.5 MoE A2.7B?
At Q4_K_M, the download is about 8.59 GB. The full-precision BF16 version is 28.63 GB. The smallest option (IQ2_M) is 4.83 GB.
Which GPUs can run Qwen1.5 MoE A2.7B?
40 consumer GPUs can run Qwen1.5 MoE A2.7B at Q4_K_M. Top options include AMD Radeon RX 6800, AMD Radeon RX 6800 XT, AMD Radeon RX 6900 XT, AMD Radeon RX 6700 XT. 26 GPUs have plenty of headroom for comfortable inference.
Which devices can run Qwen1.5 MoE A2.7B?
150 devices with unified memory can run Qwen1.5 MoE A2.7B at Q4_K_M (9.3 GB), including AMD Ryzen AI 9 HX 370 (Strix Point) Laptop, ASUS Ascent GX10, Beelink GTR9 Pro (Ryzen AI Max+ 395, 128 GB), Framework Desktop (Ryzen AI Max+ 395, 128 GB). Apple Silicon Macs use unified memory shared between CPU and GPU, making them well-suited for local LLM inference.