Qwen Models: Local AI from 0.5B to 2.4T
Qwen is the widest open-weight family in the catalog, from 0.5B models that run on iPhones up to the Qwen3.8 2.4T A95B cloud MoE. For local hardware the flagship is Qwen3.8 27B: dense, Apache 2.0, with native vision-language support, 262K context extensible to 1M, and MTP decoding. It runs with `ollama run qwen3.8:27b` on a 24GB machine. Above it, Qwen3.8-Flash-Next is a 125B total / 6B active MoE whose ~111GB Q4 GGUF runs through llama.cpp on 192GB machines, and the 2.4T A95B stays API-only in practice. The mid range covers Qwen3 8B (the 16GB-machine pick) and Qwen3 14B, then steps down to Qwen3.5 4B and the phone-class 0.5B models.
- developer
- Alibaba Cloud
- local models
- 40
- API-only models
- 2
- size range
- 0.5B-235B
- RAM range
- 2-192 GB
Table 1Local builds
All Qwen models
40 local models| Model | Size | Quant | VRAM | Min RAM | Best For | Quality | Ollama |
|---|---|---|---|---|---|---|---|
| Qwen2.5 0.5B Instruct | 0.5B | Q4_K_M | 0.8 GB | 2 GB | Chat, Mobile | 42 | |
| Qwen3.5 0.8B Instruct | 0.8B | Q4_K_M | 0.8 GB | 2 GB | Chat, Mobile | 56 | |
| Qwen2.5 1.5B Instruct | 1.5B | Q4_K_M | 1.5 GB | 4 GB | Chat, Translation | 55 | |
| Qwen3.5 2B Instruct | 2B | Q4_K_M | 1.8 GB | 4 GB | Chat, Edge tasks | 70 | |
| Qwen2.5 3B Instruct | 3B | Q4_K_M | 2.5 GB | 4 GB | Chat, Coding | 62 | |
| Qwen3.5 4B Instruct | 4B | Q4_K_M | 3.5 GB | 6 GB | Coding, Agents, Multimodal | 83 | |
| Qwen3.5 4B Instruct (Q8) | 4B | Q8_0 | 4.3 GB | 8 GB | Coding, Agents, Multimodal | 85 | |
| Qwen2.5 7B Instruct | 7B | Q4_K_M | 5.5 GB | 8 GB | Chat, Coding | 74 | |
| Qwen2.5 Coder 7B | 7B | Q4_K_M | 5.5 GB | 8 GB | Coding | 78 | |
| Qwen3 8B | 8B | Q4_K_M | 6.5 GB | 12 GB | Chat, Coding | 87 | |
| Qwen3 8B (Q8) | 8B | Q8_0 | 8.1 GB | 16 GB | Chat, Coding | 89 | |
| Qwen3.5 9B Instruct | 9B | Q4_K_M | 7 GB | 12 GB | Quality, Coding, Reasoning | 91 | |
| Qwen3.5 9B Instruct (Q8) | 9B | Q8_0 | 10.7 GB | 16 GB | Quality, Coding, Reasoning | 93 | |
| Qwen2.5 14B Instruct | 14B | Q4_K_M | 11 GB | 16 GB | Coding, Chat | 82 | |
| Qwen2.5 Coder 14B | 14B | Q4_K_M | 11 GB | 16 GB | Coding | 85 | |
| Qwen3 14B | 14B | Q4_K_M | 11 GB | 16 GB | Coding, Quality | 89 | |
| Qwen3 14B (Q8) | 14B | Q8_0 | 15.9 GB | 24 GB | Coding, Quality | 91 | |
| Qwen3.5 27B Instruct | 27B | Q4_K_M | 16 GB | 24 GB | Chat, Coding, Complex reasoning | 91 | |
| Qwen3.5 27B Instruct (Q8) | 27B | Q8_0 | 27.1 GB | 48 GB | Chat, Coding, Complex reasoning | 93 | |
| Qwen3.6 27B | 27B | Q4_K_M | 18 GB | 32 GB | Coding, Quality, Long context | 94 | |
| Qwen3.8 27B | 27B | Q4_K_M | 16.5 GB | 24 GB | Coding, Agent, Vision, Long context | 97 | |
| Qwen3.8 27B (Q8) | 27B | Q8_0 | 27.1 GB | 48 GB | Coding, Agent, Vision, Long context | 99 | |
| Qwen3.6 27B (Q8) | 27B | Q8_0 | 30 GB | 48 GB | Coding, Quality, Long context | 96 | |
| Qwen3 30B | 30B | Q4_K_M | 22 GB | 32 GB | Quality, Coding | 88 | |
| Qwen3 30B (Q8) | 30B | Q8_0 | 30.3 GB | 48 GB | Quality, Coding | 90 | |
| Qwen3 Coder 30B-A3B | 30.5B | Q4_K_M | 18.6 GB | 32 GB | Coding, Agentic coding, Long context | 89 | |
| Qwen3 Coder 30B-A3B (Q8) | 30.5B | Q8_0 | 32.5 GB | 48 GB | Coding, Agentic coding, Long context | 91 | |
| Qwen3 32B | 32B | Q4_K_M | 20.2 GB | 32 GB | Quality, Coding | 90 | |
| Qwen3 32B (Q8) | 32B | Q8_0 | 35.1 GB | 64 GB | Quality, Coding | 92 | |
| Qwen3.5 35B-A3B Instruct | 35B | Q4_K_M | 20 GB | 32 GB | Reasoning, Coding, Agent scenarios | 93 | |
| Qwen3.6 35B-A3B | 35B | Q4_K_M | 22 GB | 32 GB | Reasoning, Coding, Agents | 93 | |
| Qwen3.6 35B-A3B (Q8) | 35B | Q8_0 | 38.7 GB | 64 GB | Reasoning, Coding, Agents | 95 | |
| Qwen3.5 35B-A3B Instruct (Q8) | 35B | Q8_0 | 38.7 GB | 64 GB | Reasoning, Coding, Agent scenarios | 95 | |
| Qwen3-Next 80B-A3B | 80B | Q4_K_M | 50.4 GB | 72 GB | Chat, Coding, Long Context | 94 | |
| Qwen3-Next 80B-A3B (Q8) | 80B | Q8_0 | 84.8 GB | 128 GB | Chat, Coding, Long Context | 96 | |
| Qwen3.5 122B-A10B Instruct | 122B | Q4_K_M | 72 GB | 96 GB | Frontier-level reasoning, Complex tasks | 97 | |
| Qwen3.8-Flash-Next | 125B | Q4_K_XL | 111.3 GB | 192 GB | Agentic coding, Reasoning, Multimodal | 92 | N/A |
| Qwen3.8-Flash-Next (IQ4_XS) | 125B | IQ4_XS | 93.7 GB | 128 GB | Agentic coding, Reasoning, Multimodal | 91 | N/A |
| Qwen3.8-Flash-Next (Q3) | 125B | Q3_K_XL | 90 GB | 128 GB | Agentic coding, Reasoning, Multimodal | 89 | N/A |
| Qwen3 235B A22B | 235B | Q4_K_M | 130 GB | 192 GB | Quality, Reasoning | 90 |
4-bit builds
Qwen model sizes in GB
26 buildsEvery local Qwen model at 4-bit, from 0.5B to 235B parameters. Load is our estimate of the memory the model takes once loaded (est.). The file size is the same quant as listed on Hugging Face, where we track it.
| Model | Params | Quant | Q4 load (est.) | File on HF | Min RAM |
|---|---|---|---|---|---|
| Qwen2.5 0.5B Instruct | 0.5B | Q4_K_M | 0.8 GB | - | 2 GB |
| Qwen3.5 0.8B Instruct | 0.8B | Q4_K_M | 0.8 GB | - | 2 GB |
| Qwen2.5 1.5B Instruct | 1.5B | Q4_K_M | 1.5 GB | - | 4 GB |
| Qwen3.5 2B Instruct | 2B | Q4_K_M | 1.8 GB | - | 4 GB |
| Qwen2.5 3B Instruct | 3B | Q4_K_M | 2.5 GB | - | 4 GB |
| Qwen3.5 4B Instruct | 4B | Q4_K_M | 3.5 GB | 2.81 GB | 6 GB |
| Qwen2.5 7B Instruct | 7B | Q4_K_M | 5.5 GB | - | 8 GB |
| Qwen2.5 Coder 7B | 7B | Q4_K_M | 5.5 GB | - | 8 GB |
| Qwen3 8B | 8B | Q4_K_M | 6.5 GB | 4.68 GB | 12 GB |
| Qwen3.5 9B Instruct | 9B | Q4_K_M | 7 GB | 5.75 GB | 12 GB |
| Qwen2.5 14B Instruct | 14B | Q4_K_M | 11 GB | - | 16 GB |
| Qwen2.5 Coder 14B | 14B | Q4_K_M | 11 GB | - | 16 GB |
| Qwen3 14B | 14B | Q4_K_M | 11 GB | 8.38 GB | 16 GB |
| Qwen3.5 27B Instruct | 27B | Q4_K_M | 16 GB | 16.75 GB | 24 GB |
| Qwen3.6 27B | 27B | Q4_K_M | 18 GB | 16.75 GB | 32 GB |
| Qwen3.8 27B | 27B | Q4_K_M | 16.5 GB | 15.33 GB | 24 GB |
| Qwen3 30B | 30B | Q4_K_M | 22 GB | 17.35 GB | 32 GB |
| Qwen3 Coder 30B-A3B | 30.5B | Q4_K_M | 18.6 GB | - | 32 GB |
| Qwen3 32B | 32B | Q4_K_M | 20.2 GB | - | 32 GB |
| Qwen3.5 35B-A3B Instruct | 35B | Q4_K_M | 20 GB | 20.75 GB | 32 GB |
| Qwen3.6 35B-A3B | 35B | Q4_K_M | 22 GB | 20.75 GB | 32 GB |
| Qwen3-Next 80B-A3B | 80B | Q4_K_M | 50.4 GB | 45.17 GB | 72 GB |
| Qwen3.5 122B-A10B Instruct | 122B | Q4_K_M | 72 GB | - | 96 GB |
| Qwen3.8-Flash-Next | 125B | Q4_K_XL | 111.3 GB | - | 192 GB |
| Qwen3.8-Flash-Next (IQ4_XS) | 125B | IQ4_XS | 93.7 GB | - | 128 GB |
| Qwen3 235B A22B | 235B | Q4_K_M | 130 GB | - | 192 GB |
By memory
Best local Qwen model by RAM
For each amount of RAM, the Qwen build with the highest ModelFit quality score whose minimum RAM fits. On a tie, the larger model wins.
| RAM | Best Qwen model | Quant | Load (est.) | Min RAM | Quality |
|---|---|---|---|---|---|
| 8 GB | Qwen3.5 4B Instruct (Q8) | Q8_0 | 4.3 GB | 8 GB | 85 |
| 16 GB | Qwen3.5 9B Instruct (Q8) | Q8_0 | 10.7 GB | 16 GB | 93 |
| 24 GB | Qwen3.8 27B | Q4_K_M | 16.5 GB | 24 GB | 97 |
| 32 GB | Qwen3.8 27B | Q4_K_M | 16.5 GB | 24 GB | 97 |
| 64 GB | Qwen3.8 27B (Q8) | Q8_0 | 27.1 GB | 48 GB | 99 |
| 128 GB | Qwen3.8 27B (Q8) | Q8_0 | 27.1 GB | 48 GB | 99 |
Key features
- Qwen3.8 27B local flagship (Apache 2.0, vision)
- Cloud line tops out at the 2.4T A95B MoE
- Qwen3.8-Flash-Next: 125B MoE from 192GB RAM
- Mid range: Qwen3 8B and 14B on 12-16GB Macs
- Phone-class models from 0.5B, 2GB RAM
Table 2
Cloud / API-only models
2 APIWeights are open but far beyond consumer hardware. These run as APIs. Each page lists the local alternatives.
| Model | Size | Access | Best For |
|---|---|---|---|
| Qwen3.7-Plus | Undisclosed | API only | Multimodal, Agentic, Vision |
| Qwen3.8 2.4T A95B | 2400B | API only | Frontier reasoning, Agentic coding |
Table 3
Device compatibility
Which Qwen models can run on each device class, based on minimum RAM requirements.
| Model | iPhone | Air | Pro | Studio | Mini | |
|---|---|---|---|---|---|---|
| Qwen2.5 0.5B Instruct (0.5B) | Excellent | Excellent | Excellent | Excellent | Excellent | Excellent |
| Qwen3.5 0.8B Instruct (0.8B) | Excellent | Excellent | Excellent | Excellent | Excellent | Excellent |
| Qwen2.5 1.5B Instruct (1.5B) | Excellent | Excellent | Excellent | Excellent | Excellent | Excellent |
| Qwen3.5 2B Instruct (2B) | Excellent | Excellent | Excellent | Excellent | Excellent | Excellent |
| Qwen2.5 3B Instruct (3B) | Excellent | Excellent | Excellent | Excellent | Excellent | Excellent |
| Qwen3.5 4B Instruct (4B) | Possible | Possible | Excellent | Excellent | Excellent | Possible |
| Qwen3.5 4B Instruct (Q8) (4B) | Possible | Possible | Excellent | Excellent | Excellent | Possible |
| Qwen2.5 7B Instruct (7B) | Possible | Possible | Excellent | Excellent | Excellent | Possible |
| Qwen2.5 Coder 7B (7B) | Possible | Possible | Excellent | Excellent | Excellent | Possible |
| Qwen3 8B (8B) | Possible | Possible | Possible | Excellent | Possible | Possible |
| Qwen3 8B (Q8) (8B) | No | Possible | Possible | Excellent | Possible | Possible |
| Qwen3.5 9B Instruct (9B) | Possible | Possible | Possible | Excellent | Possible | Possible |
| Qwen3.5 9B Instruct (Q8) (9B) | No | Possible | Possible | Excellent | Possible | Possible |
| Qwen2.5 14B Instruct (14B) | No | Possible | Possible | Excellent | Possible | Possible |
| Qwen2.5 Coder 14B (14B) | No | Possible | Possible | Excellent | Possible | Possible |
| Qwen3 14B (14B) | No | Possible | Possible | Excellent | Possible | Possible |
| Qwen3 14B (Q8) (14B) | No | Possible | Possible | Possible | Possible | Possible |
| Qwen3.5 27B Instruct (27B) | No | Possible | Possible | Possible | Possible | Possible |
| Qwen3.5 27B Instruct (Q8) (27B) | No | No | Possible | Possible | Possible | No |
| Qwen3.6 27B (27B) | No | Possible | Possible | Possible | Possible | Possible |
| Qwen3.8 27B (27B) | No | Possible | Possible | Possible | Possible | Possible |
| Qwen3.8 27B (Q8) (27B) | No | No | Possible | Possible | Possible | No |
| Qwen3.6 27B (Q8) (27B) | No | No | Possible | Possible | Possible | No |
| Qwen3 30B (30B) | No | Possible | Possible | Possible | Possible | Possible |
| Qwen3 30B (Q8) (30B) | No | No | Possible | Possible | Possible | No |
| Qwen3 Coder 30B-A3B (30.5B) | No | Possible | Possible | Possible | Possible | Possible |
| Qwen3 Coder 30B-A3B (Q8) (30.5B) | No | No | Possible | Possible | Possible | No |
| Qwen3 32B (32B) | No | Possible | Possible | Possible | Possible | Possible |
| Qwen3 32B (Q8) (32B) | No | No | Possible | Possible | Possible | No |
| Qwen3.5 35B-A3B Instruct (35B) | No | Possible | Possible | Possible | Possible | Possible |
| Qwen3.6 35B-A3B (35B) | No | Possible | Possible | Possible | Possible | Possible |
| Qwen3.6 35B-A3B (Q8) (35B) | No | No | Possible | Possible | Possible | No |
| Qwen3.5 35B-A3B Instruct (Q8) (35B) | No | No | Possible | Possible | Possible | No |
| Qwen3-Next 80B-A3B (80B) | No | No | Possible | Possible | No | No |
| Qwen3-Next 80B-A3B (Q8) (80B) | No | No | Possible | Possible | No | No |
| Qwen3.5 122B-A10B Instruct (122B) | No | No | Possible | Possible | No | No |
| Qwen3.8-Flash-Next (125B) | No | No | No | Possible | No | No |
| Qwen3.8-Flash-Next (IQ4_XS) (125B) | No | No | Possible | Possible | No | No |
| Qwen3.8-Flash-Next (Q3) (125B) | No | No | Possible | Possible | No | No |
| Qwen3 235B A22B (235B) | No | No | No | Possible | No | No |
RAM requirements
Fig. 1Estimated load per model, with the minimum RAM
Qwen2.5 0.5B Instruct0.8 GB · min 2 GB
Qwen3.5 0.8B Instruct0.8 GB · min 2 GB
Qwen2.5 1.5B Instruct1.5 GB · min 4 GB
Qwen3.5 2B Instruct1.8 GB · min 4 GB
Qwen2.5 3B Instruct2.5 GB · min 4 GB
Qwen3.5 4B Instruct3.5 GB · min 6 GB
Qwen3.5 4B Instruct (Q8)4.3 GB · min 8 GB
Qwen2.5 7B Instruct5.5 GB · min 8 GB
Qwen2.5 Coder 7B5.5 GB · min 8 GB
Qwen3 8B6.5 GB · min 12 GB
Qwen3 8B (Q8)8.1 GB · min 16 GB
Qwen3.5 9B Instruct7 GB · min 12 GB
Qwen3.5 9B Instruct (Q8)10.7 GB · min 16 GB
Qwen2.5 14B Instruct11 GB · min 16 GB
Qwen2.5 Coder 14B11 GB · min 16 GB
Qwen3 14B11 GB · min 16 GB
Qwen3 14B (Q8)15.9 GB · min 24 GB
Qwen3.5 27B Instruct16 GB · min 24 GB
Qwen3.5 27B Instruct (Q8)27.1 GB · min 48 GB
Qwen3.6 27B18 GB · min 32 GB
Qwen3.8 27B16.5 GB · min 24 GB
Qwen3.8 27B (Q8)27.1 GB · min 48 GB
Qwen3.6 27B (Q8)30 GB · min 48 GB
Qwen3 30B22 GB · min 32 GB
Qwen3 30B (Q8)30.3 GB · min 48 GB
Qwen3 Coder 30B-A3B18.6 GB · min 32 GB
Qwen3 Coder 30B-A3B (Q8)32.5 GB · min 48 GB
Qwen3 32B20.2 GB · min 32 GB
Qwen3 32B (Q8)35.1 GB · min 64 GB
Qwen3.5 35B-A3B Instruct20 GB · min 32 GB
Qwen3.6 35B-A3B22 GB · min 32 GB
Qwen3.6 35B-A3B (Q8)38.7 GB · min 64 GB
Qwen3.5 35B-A3B Instruct (Q8)38.7 GB · min 64 GB
Qwen3-Next 80B-A3B50.4 GB · min 72 GB
Qwen3-Next 80B-A3B (Q8)84.8 GB · min 128 GB
Qwen3.5 122B-A10B Instruct72 GB · min 96 GB
Qwen3.8-Flash-Next111.3 GB · min 192 GB
Qwen3.8-Flash-Next (IQ4_XS)93.7 GB · min 128 GB
Qwen3.8-Flash-Next (Q3)90 GB · min 128 GB
Qwen3 235B A22B130 GB · min 192 GB
Fig. 2Max model size by memory
Largest dense Q4 model that fits each memory tier
8 GB9B
12 GB12B
16 GB14B
24 GB29.3B
32 GB33.1B
36 GB33.1B
48 GB33.1B
64 GB70B
72 GB70B
96 GB123B
128 GB123B
192 GB123B
256 GB123B
512 GB123B
billion parametersGet told when a better model fits your MacBook Air M5 16 GB
Get one email when a new open-weight model is a better fit for this machine.
For: MacBook Air M5 16 GB
You'll also get the weekly ModelFit email. By signing up you agree to our Privacy Policy. Unsubscribe from either email anytime, in one click.
Config reports
| Source | Machine | tok/s | Quant | Runtime | Context | Note |
|---|---|---|---|---|---|---|
| Speed test | NVIDIA GeForce RTX 3090Qwen2.5 1.5B Instruct | 306.4 | Q4_K_M | Ollama | - | 1 run |
| Speed test | NVIDIA GeForce RTX 4070 TiQwen2.5 1.5B Instruct | 260.1 | Q4_K_M | Ollama | - | 1 run |
| Speed test | Apple M5 Max 128 GBQwen2.5 1.5B Instruct | 251.0 | Q4_K_M | Ollama | - | 3 runs (median) |
| Speed test | Apple M4 Pro 48 GBQwen2.5 1.5B Instruct | 166.7 | Q4_K_M | Ollama | - | 1 run |
| Speed test | NVIDIA GeForce RTX 5080Qwen3.5 4B Instruct | 163.8 | Q4_K_M | Ollama | - | 1 run |
| Speed test | NVIDIA GeForce RTX 5090Qwen3.8 27B | 150.0 | Q4_K_M | Ollama | - | 1 run |
| Speed test | NVIDIA GeForce RTX 3060Qwen2.5 1.5B Instruct | 134.6 | Q4_K_M | Ollama | - | 1 run |
| Speed test | Apple M3 Ultra 60-core GPU 96 GBQwen3.5 9B Instruct | 126.2 | Q4_K_M | Ollama | - | 1 run |
| Speed test | Apple M5 Ultra 64-core GPU 96 GBQwen3.5 9B Instruct | 125.0 | Q4_K_M | Ollama | - | 1 run |
| Speed test | Apple M1 Max 64 GBQwen2.5 1.5B Instruct | 115.6 | Q4_K_M | Ollama | - | 1 run |
| Speed test | NVIDIA GeForce RTX 5060 TiQwen3.5 4B Instruct | 104.2 | Q4_K_M | Ollama | - | 1 run |
| Speed test | NVIDIA GeForce RTX 4070Qwen3.5 4B Instruct | 101.4 | Q4_K_M | Ollama | - | 1 run |
| Speed test | Apple M2 Max 38-core GPU 64 GBQwen3.6 35B-A3B | 101.0 | Q4_K_M | Ollama | - | 1 run |
| Speed test | Apple M5 Ultra 64-core GPU 96 GBQwen3.8 27B | 99.4 | Q4_K_M | Ollama | - | 1 run |
| Speed test | NVIDIA GeForce RTX 3070Qwen3.5 4B Instruct | 96.8 | Q4_K_M | Ollama | - | 2 runs (median) |
| Speed test | NVIDIA GeForce RTX 4080Qwen3.5 9B Instruct | 92.5 | Q4_K_M | Ollama | - | 2 runs (median) |
| Speed test | Apple M4 Pro 20-core GPU 48 GBQwen3.6 35B-A3B | 80.5 | Q4_K_M | Ollama | - | 4 runs (median) |
| Speed test | Apple M4 16 GBQwen2.5 1.5B Instruct | 71.3 | Q4_K_M | Ollama | - | 1 run |
| Speed test | Apple M5 Max 40-core GPU 128 GBQwen3.8 27B | 71.3 | Q4_K_M | Ollama | - | 1 run |
| Speed test | Apple M5 Max 40-core GPU 64 GBQwen3.8 27B | 70.9 | Q4_K_M | Ollama | - | 1 run |
| Speed test | NVIDIA GeForce RTX 5060Qwen3.5 4B Instruct | 69.8 | Q4_K_M | Ollama | - | 1 run |
| Speed test | NVIDIA GeForce RTX 4070Qwen3.5 9B Instruct | 67.7 | Q4_K_M | Ollama | - | 1 run |
| Speed test | NVIDIA GeForce RTX 5060 TiQwen3.5 9B Instruct | 67.1 | Q4_K_M | Ollama | - | 1 run |
| Speed test | Apple M1 Pro 32 GBQwen2.5 1.5B Instruct | 65.9 | Q4_K_M | Ollama | - | 3 runs (median) |
| Speed test | Apple M1 Max 24-core GPU 64 GBQwen3.5 4B Instruct | 52.1 | Q4_K_M | Ollama | - | 1 run |
| Speed test | Apple M1 8 GBQwen2.5 1.5B Instruct | 51.1 | Q4_K_M | Ollama | - | 2 runs (median) |
| Speed test | Apple M4 Max 40-core GPU 64 GBQwen3.8 27B | 49.7 | Q4_K_M | Ollama | - | 2 runs (median) |
| Speed test | Apple M2 Ultra 60-core GPU 128 GBQwen3.8 27B | 43.0 | Q4_K_M | Ollama | - | 1 run |
| Speed test | Apple M5 Pro 20-core GPU 48 GBQwen3.8 27B | 37.8 | Q4_K_M | Ollama | - | 2 runs (median) |
| Speed test | NVIDIA GeForce RTX 4060 TiQwen3.5 9B Instruct | 34.0 | Q4_K_M | Ollama | - | 2 runs (median) |
| Speed test | Apple M2 Pro 16-core GPU 16 GBQwen3.5 9B Instruct | 32.7 | Q4_K_M | Ollama | - | 1 run |
| Speed test | Apple M4 Pro 20-core GPU 48 GBQwen3.8 27B | 31.7 | Q4_K_M | Ollama | - | 2 runs (median) |
| Speed test | Apple M1 Pro 16-core GPU 16 GBQwen3.5 9B Instruct | 29.9 | Q4_K_M | Ollama | - | 1 run |
| Speed test | Apple M4 10-core GPU 16 GBQwen3.5 4B Instruct | 27.7 | Q4_K_M | Ollama | - | 3 runs (median) |
| Speed test | Apple M2 Max 30-core GPU 64 GBQwen3.8 27B | 25.3 | Q4_K_M | Ollama | - | 1 run |
| Speed test | Apple M2 10-core GPU 16 GBQwen3.5 4B Instruct | 23.0 | Q4_K_M | Ollama | - | 1 run |
| Speed test | Apple M1 Max 24-core GPU 64 GBQwen3.8 27B | 14.5 | Q4_K_M | Ollama | - | 1 run |
| Speed test | Apple M5 Max 32-core GPU 36 GBQwen3.8 27B | 13.1 | Q4_K_M | Ollama | - | 1 run |
| Speed test | Apple M1 7-core GPU 8 GBQwen3.5 4B Instruct | 10.2 | Q4_K_M | Ollama | - | 1 run |
| Speed test | Apple M2 10-core GPU 16 GBQwen3.5 9B Instruct | 4.9 | Q4_K_M | Ollama | - | 1 run |
Speed test: runs submitted through the ModelFit speed test tool, one median per machine.
Share your config
Frequently asked questions
What is the best Qwen model for 16GB RAM?
Qwen3 8B is the 16GB-machine pick: `ollama run qwen3:8b-q4_K_M` needs 12GB RAM minimum and loads about 6.5GB. With more headroom, Qwen3.8 27B is the local flagship via `ollama run qwen3.8:27b` at 24GB RAM.
Can the newest Qwen flagship run locally?
Qwen3.8 2.4T A95B is API-only in practice. Its BF16 weights are public on Hugging Face under the Qwen3.8-Max license, but no consumer build is registry-verified. The closest local experience is Qwen3.8-Flash-Next, a 125B / 6B active MoE with a 192GB RAM floor.
What Ollama command runs Qwen3.8 27B?
Run `ollama run qwen3.8:27b` on a machine with 24GB RAM minimum, a load of about 16.5GB. The higher-fidelity Q8 build, `ollama run qwen3.8:27b-q8_0`, wants 48GB.
Can I run Qwen on an iPhone?
Yes. Qwen2.5 0.5B runs from 2GB RAM with `ollama run qwen2.5:0.5b-instruct-q4_K_M`, and Qwen3.5 0.8B (`ollama run qwen3.5:0.8b`) also needs just 2GB. Quality is basic but useful for chat and translation.
What is Qwen3.8-Flash-Next?
It is a 125B total / 6B active MoE from the Qwen3.8 line. The unsloth Q4 GGUF is about 111GB, so it needs a 192GB machine, and a ~94GB IQ4_XS build fits 128GB. There is no Ollama tag, so it runs through llama.cpp and similar runners rather than Ollama.
