Yes — with 142.0 GB to spare
Qwen3.5 2B at Q4_K_M fits your M2 Ultra · 192 GB entirely in unified memory at 8K context, at an estimated 350 tokens per second. There is room for its full 256K window.
Phone-class, and multimodal. Replaces Llama 3.2 3B as the "it runs on anything" answer.
Hardware requirements What hardware do I need to run Qwen3.5 2B? →At Q4_K_M with an 8K context: a 6 GB GPU (RTX 3060 Laptop) or a 8 GB Mac (M3). Minimum, recommended, high-quality and long-context tiers, with all 114 devices checked at every quantisation.
The VRAM budget
Quantisation ladder
| Quant | Weights | Total @ 8K | Max context | Tok/s | Quality | Fit |
|---|---|---|---|---|---|---|
| F16 | 4.2 GB | 4.9 GB | 256K | 106 | Reference | Long context |
| Q8_0 | 2.2 GB | 2.9 GB | 256K | 199 | −0.1% ppl | Long context |
| Q6_K | 1.7 GB | 2.4 GB | 256K | 258 | −0.4% ppl | Long context |
| Q5_K_M | 1.5 GB | 2.2 GB | 256K | 298 | −0.8% ppl | Long context |
| Q4_K_M | 1.3 GB | 2.0 GB | 256K | 350 | −1.9% ppl | Recommended |
| Q3_K_M | 1.0 GB | 1.7 GB | 256K | 433 | −5.4% ppl | Long context |
| Q2_K | 0.9 GB | 1.6 GB | 256K | 505 | −15% ppl | Long context |
Quality is the published perplexity delta against f16 weights. Max context assumes an f16 KV cache; q8_0 roughly doubles it. Only 6 of its 24 blocks keep a per-token KV cache; the rest are linear-attention, Mamba or convolution blocks with a fixed-size state.
How to run it
$ pip install mlx-lm $ mlx_lm.generate --model mlx-community/Qwen3.5-2B-4bit \ --max-tokens 512 --prompt "Hello"
Apple's own array framework. The fastest path on Apple Silicon. More on MLX.
Related models on an M2 Ultra · 192 GB
| Model | Total | Tok/s est. | |
|---|---|---|---|
| Qwen3.5 4B 4.66B · same family |
Q4_K_M · 3.5 GB | 171 | Runs great |
| Qwen3.5 0.8B 0.87B · same family |
Q4_K_M · 1.2 GB | 914 | Runs great |
| Qwen3.5 9B 9.65B · same family |
Q4_K_M · 6.3 GB | 82 | Runs great |
| Qwen3.5 122B-A10B 125B MoE · same family |
Q4_K_M · 71.1 GB | 37 | Runs great |
| LFM2.5 2.6B 2.7B · similar size |
Q4_K_M · 2.2 GB | 294 | Runs great |
| Qwen3 1.7B 1.72B · similar size |
Q4_K_M · 2.4 GB | 462 | Runs great |
| Llama 3.2 3B Instruct 3.21B · similar size |
Q4_K_M · 3.3 GB | 248 | Runs great |
Qwen3.5 2B on other hardware
| Device | Total | Tok/s est. | |
|---|---|---|---|
| GeForce RTX 3060 Laptop 5.3 GB usable · 336 GB/s |
2.0 GB | 159 | Runs great |
| GeForce RTX 4060 Laptop 7.0 GB usable · 256 GB/s |
2.0 GB | 121 | Runs great |
| M3 · 8 GB 5.4 GB usable · 100 GB/s |
2.0 GB | 44 | Runs great |
| GeForce RTX 3060 12 GB 10.6 GB usable · 360 GB/s |
2.0 GB | 171 | Runs great |
| M4 · 16 GB 10.7 GB usable · 120 GB/s |
2.0 GB | 53 | Runs great |
| GeForce RTX 4090 22.4 GB usable · 1008 GB/s |
2.0 GB | 478 | Runs great |
| Radeon RX 7900 XTX 22.4 GB usable · 960 GB/s |
2.0 GB | 455 | Runs great |
| M3 Max · 64 GB 48.0 GB usable · 400 GB/s |
2.0 GB | 175 | Runs great |