DeepSeek Models: Cloud V4, Local R1 Reasoning
DeepSeek's V4 generation is a cloud story. V4.1 Flash (released 2026-09-10, MIT) is a multimodal image + text MoE with a 552B backbone and a 196B Engram conditional memory that activates 8B on prefill and 16B on decode, with 1M context. V4 Flash 0731 (284B total / 13B active) and V4 Pro (1.6T total / 49B active) are API models, and none of the three has a registry-verified Ollama tag. The local DeepSeek line is R1: the 7B, 14B and Llama 70B distills run in Q4_K_M via Ollama starting at 8GB RAM, and the full R1 671B needs a 512GB Mac Studio. V4 Flash 0731 does get community 2-bit dynamic quants that fit 128GB Macs through MLX, llama.cpp or the ds4 engine.
- developer
- DeepSeek AI
- local models
- 8
- API-only models
- 6
- size range
- 7B-671B
- RAM range
- 8-512 GB
All DeepSeek models
8 local models| Model | Size | Quant | VRAM | Min RAM | Best For | Quality | Ollama |
|---|---|---|---|---|---|---|---|
| DeepSeek-R1 Distill Qwen 7B | 7B | Q4_K_M | 5.5 GB | 8 GB | Reasoning, Coding | 80 | |
| DeepSeek-R1 Distill Qwen 7B (Q8) | 7B | Q8_0 | 7.5 GB | 16 GB | Reasoning, Coding | 82 | |
| DeepSeek-R1 Distill Qwen 14B | 14B | Q4_K_M | 11 GB | 16 GB | Reasoning, Quality | 86 | |
| DeepSeek-R1 Distill Qwen 14B (Q8) | 14B | Q8_0 | 14.6 GB | 24 GB | Reasoning, Quality | 88 | |
| DeepSeek-R1 Distill Qwen 32B | 32B | Q4_K_M | 19.9 GB | 32 GB | Reasoning, Quality | 87 | |
| DeepSeek-R1 Distill Qwen 32B (Q8) | 32B | Q8_0 | 34.8 GB | 64 GB | Reasoning, Quality | 89 | |
| DeepSeek-R1 Distill Llama 70B | 70B | Q4_K_M | 42 GB | 64 GB | Reasoning, Quality | 85 | |
| DeepSeek-R1 671B | 671B | Q4_K_M | 380 GB | 512 GB | Reasoning, Coding | 91 |
Key features
- V4.1 Flash: multimodal 552B MoE, API only
- R1 distills 7B-70B run locally in Ollama
- Full R1 671B needs a 512GB Mac Studio
- V4 Flash 0731 quants fit 128GB Macs
- Reasoning-first: strong for coding and debugging
Cloud / API-only models
6 APIWeights are open but far beyond consumer hardware. These run as APIs. Each page lists the local alternatives.
| Model | Size | Access | Best For |
|---|---|---|---|
| DeepSeek V4 Flash 0731 | 284B | API only | Reasoning, Coding, Long context |
| DeepSeek-V3 | 671B | API only | Quality, Coding |
| DeepSeek-V3-0324 | 671B | API only | Quality, Coding |
| DeepSeek-R1 | 671B | API only | Reasoning, Quality |
| DeepSeek V4.1 Flash | 763.21B | API only | Agentic coding, Reasoning, Long context, Vision |
| DeepSeek V4 Pro | 1600B | API only | Frontier reasoning, Coding |
Device compatibility
Which DeepSeek models can run on each device class, based on minimum RAM requirements.
| Model | iPhone | Air | Pro | Studio | Mini | |
|---|---|---|---|---|---|---|
| DeepSeek-R1 Distill Qwen 7B (7B) | Possible | Possible | Excellent | Excellent | Excellent | Possible |
| DeepSeek-R1 Distill Qwen 7B (Q8) (7B) | No | Possible | Possible | Excellent | Possible | Possible |
| DeepSeek-R1 Distill Qwen 14B (14B) | No | Possible | Possible | Excellent | Possible | Possible |
| DeepSeek-R1 Distill Qwen 14B (Q8) (14B) | No | Possible | Possible | Possible | Possible | Possible |
| DeepSeek-R1 Distill Qwen 32B (32B) | No | Possible | Possible | Possible | Possible | Possible |
| DeepSeek-R1 Distill Qwen 32B (Q8) (32B) | No | No | Possible | Possible | Possible | No |
| DeepSeek-R1 Distill Llama 70B (70B) | No | No | Possible | Possible | Possible | No |
| DeepSeek-R1 671B (671B) | No | No | No | Possible | No | No |
RAM requirements
Largest dense Q4 model that fits each memory tier
Get 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
No speed test run for this family yet. Run the speed test: npx @modelfit/cli@latest bench
Share your config
Frequently asked questions
Can I run DeepSeek V4.1 Flash locally?
No. V4.1 Flash is a cloud-only multimodal MoE with no registry-verified Ollama tag, and the same holds for V4 Pro. The local DeepSeek line is the R1 reasoning distills; V4 Flash 0731 is the only current V4 member that runs locally, and only through community 2-bit quants on 128GB Macs.
What is the difference between V4.1 Flash, V4 Flash 0731 and V4 Pro?
V4.1 Flash (2026-09-10) is MIT-licensed and multimodal: a 552B backbone plus 196B Engram conditional memory, with 8B active on prefill and 16B on decode, and 1M context. V4 Flash 0731 is a 284B total / 13B active MoE with 1M context. V4 Pro is the 1.6T total / 49B active cloud model. All three are API-only from DeepSeek.
What Ollama commands run DeepSeek R1 locally?
Run `ollama run deepseek-r1:7b` on 8GB RAM (about 5.5GB load), `ollama run deepseek-r1:14b` on 16GB (about 11GB), or `ollama run deepseek-r1:70b` on 64GB (about 42GB). The full `ollama run deepseek-r1:671b-q4_K_M` needs a 512GB machine and loads around 380GB.
Can the newest DeepSeek flagship run on consumer hardware?
No. V4.1 Flash is cloud-only with no registry-verified build, and V4 Pro targets the API. The closest thing to local is V4 Flash 0731: 128GB-Mac owners can run it through community 2-bit dynamic quants via MLX, llama.cpp or the ds4 engine, but that is not an official Ollama tag.
Is DeepSeek R1 good for coding?
Yes. The R1 distills are reasoning-first, which helps with debugging, architecture and multi-step coding tasks. The 14B and 70B versions are the strongest local options in the family; the 7B stays fast enough for laptops.
