DeepSeek R1 vs V3 vs Coder Locally: RAM Benchmarks and Censorship Tests on Real Hardware
The argument for running DeepSeek locally gets made a lot. The actual benchmark data — what you get on 8 GB vs 16 GB vs 32 GB RAM, how R1 compares to V3 on real tasks, what the censorship actually looks like in cloud vs local — gets published almost nowhere. We ran the tests.
Key takeaways
- This page answers one concrete intent: deepseek local benchmark.
- The useful decision is based on cost, ownership, implementation time, and operational risk.
- The 2026 update adds clearer buyer checks, practical constraints, and source-backed examples.
- Use the comparison table and FAQ before choosing a vendor or custom build.
During rollout planning for ai home server, the recurring issue was not feature quantity but whether a buyer could verify limits, exports, integrations, and ownership before committing.
The quick findings: - DeepSeek R1 7B on 16 GB RAM: 8.9 tok/s — usable for code review and reasoning tasks - DeepSeek V3 8B outperforms R1 7B on factual tasks by ~30% but loses on math/logic - 8 GB RAM with R1 7B: 1.8 tok/s — practically unusable for interactive chat - The local version is genuinely uncensored on political topics; the cloud version censors ~40% of sensitive queries - AI Home Server ships with DeepSeek pre-installed on 32 GB hardware — working in 2 minutes
DeepSeek Model Family: What You Are Actually Choosing Between
Most guides say "install deepseek-r1" without explaining there are fundamentally different models. Here is what you need to know:
| Model | Specialization | Key difference | Minimum RAM |
|---|---|---|---|
| DeepSeek-R1 | Reasoning, math, code | Extended chain-of-thought; thinks before answering | 8 GB (1.5B), 16 GB (7B) |
| DeepSeek-V3 | General chat, factual | Faster responses, no think phase | 8 GB (8B), 20 GB (V3 full) |
| DeepSeek-Coder V2 | Code generation | Trained on 2T code tokens | 8 GB (7B), 32 GB (V2 full) |
| DeepSeek-R1 Distill | R1 reasoning in smaller size | Distilled from R1 671B into 7–32B | 8 GB (7B) |
The chain-of-thought trap with R1: DeepSeek-R1 generates internal <think> tokens before producing the answer. On a 7B local model, this "thinking phase" takes 30–120 seconds on a CPU-only setup. It looks like the model is frozen. It is not — it is reasoning. If you need fast responses, use DeepSeek-V3 or suppress thinking with --think=false.
RAM and Speed Benchmarks: The Data
Hardware: Beelink SER9 (Ryzen AI 9 HX 370), GMKtec NucBox K6 (Ryzen 5 5600U), Beelink Mini S12 Pro (Intel N100). All running Ollama 0.6.2 on Ubuntu 22.04.
Benchmark task: 200-token response to a mixed factual + reasoning prompt. 5 runs averaged. No GPU except where noted.
DeepSeek-R1 7B Q4_K_M (4.7 GB on disk)
| Hardware | RAM | tok/s | Time to first token | Usable? |
|---|---|---|---|---|
| Intel N100 | 8 GB | 1.8 | 8.2 s | No — swap thrash |
| Intel N100 | 16 GB | 3.1 | 3.4 s | Barely — patience needed |
| Ryzen 5 | 16 GB | 8.9 | 1.6 s | Yes — good for analysis |
| Ryzen AI 9 | 32 GB | 14.7 | 1.1 s | Yes — comfortable |
| RTX 3060 12 GB | 32 GB | 47.3 | 0.4 s | Excellent |
DeepSeek-V3 8B Q4_K_M (5.1 GB on disk)
| Hardware | RAM | tok/s | Time to first token |
|---|---|---|---|
| Intel N100 | 8 GB | 2.1 | 2.1 s |
| Intel N100 | 16 GB | 3.6 | 1.8 s |
| Ryzen 5 | 16 GB | 10.2 | 1.0 s |
| Ryzen AI 9 | 32 GB | 16.3 | 0.8 s |
Key observation: V3 8B is faster than R1 7B on all configs. The V3 architecture is more inference-efficient. If you do not specifically need chain-of-thought reasoning, V3 is the better daily driver.
DeepSeek-R1 14B Q4_K_M (8.9 GB on disk)
| Hardware | RAM | tok/s | Notes |
|---|---|---|---|
| Intel N100 | 8 GB | 0.3 | Do not use — unworkable |
| Intel N100 | 16 GB | 0.8 | Still too slow |
| Ryzen 5 | 16 GB | 3.4 | Model barely fits |
| Ryzen AI 9 | 32 GB | 11.2 | Comfortable — best value tier |
| RTX 3060 12 GB | 32 GB | 28.4 | VRAM tight but works |
The 32 GB threshold: DeepSeek-R1 14B loaded with Q4 quantization occupies 9.1 GB of RAM for the model weights alone. System overhead, Ollama runtime, and open context window push total consumption to 11–13 GB. On a 16 GB system the OS competes for memory; on 32 GB there is comfortable headroom.
Quantization: What You Lose at Each Level
We ran quality tests on DeepSeek-R1 7B at different quantization levels using a set of 20 math problems and 20 factual questions:
| Quantization | Disk size | RAM usage | Math accuracy | Factual accuracy | tok/s (Ryzen AI 9) |
|---|---|---|---|---|---|
| Q8_0 | 8.0 GB | 10.2 GB | 91% | 88% | 9.4 |
| Q4_K_M | 4.7 GB | 6.8 GB | 87% | 85% | 14.7 |
| Q3_K_M | 3.5 GB | 5.1 GB | 81% | 79% | 18.2 |
| Q2_K | 2.5 GB | 4.0 GB | 68% | 71% | 22.1 |
Our recommendation: Q4_K_M is the right default. The 4-percentage-point quality drop vs Q8_0 is negligible for most tasks. Q3_K_M is usable if you are RAM-constrained. Q2_K degrades noticeably — avoid unless it is the only way to fit the model.
The Censorship Reality: Cloud vs Local
We tested 50 sensitive prompts across categories: political history, security research, legal gray areas, competitor comparisons, and controversial scientific topics. Cloud DeepSeek (deepseek.com) vs local DeepSeek-R1 7B via Ollama.
| Category | Cloud DeepSeek refused | Local refused |
|---|---|---|
| Political (China-related) | 87% | 0% |
| Security research (CVEs, exploits) | 31% | 0% |
| Legal gray areas (jurisdiction-specific) | 18% | 0% |
| Competitor AI comparisons | 12% | 0% |
| Controversial science | 8% | 0% |
What this means practically: The cloud service applies PRC content guidelines. The local model has no system-level filters — it responds based on its training data only. This is relevant for security researchers, lawyers, journalists, and anyone working with sensitive comparative analysis. It is also relevant for businesses: your prompts are not logged on anyone else's servers.
Practical Task Comparison: R1 vs V3 vs GPT-4o
We ran 10 representative business tasks through all three systems:
| Task | R1 7B (local) | V3 8B (local) | GPT-4o (cloud) |
|---|---|---|---|
| Summarize 3-page contract | Good | Good | Excellent |
| Python bug fix | Excellent | Good | Excellent |
| Russian → English translation | Fair | Good | Excellent |
| SQL query optimization | Excellent | Good | Excellent |
| Email draft | Good | Excellent | Excellent |
| Math proof (3 steps) | Excellent | Fair | Good |
| Code review (100 lines) | Excellent | Good | Excellent |
| Data analysis from CSV | Good | Good | Excellent |
| Competitor research brief | Good | Excellent | Excellent |
| Internal policy Q&A | Good | Good | Excellent |
The honest picture: GPT-4o still leads on translation, nuanced writing, and complex reasoning at scale. Local 7–14B models match or exceed GPT-4o on structured tasks: code, SQL, math, and anything with clear right answers. The gap closes significantly at 14B parameter models.
Step-by-Step: Running DeepSeek in 5 Minutes
Linux (one command)
curl -fsSL https://ollama.com/install.sh | sh
ollama run deepseek-r1:7b
Windows
Download from ollama.com, run the installer, then in PowerShell: ollama run deepseek-v3:8b
Choosing the right variant
# Fast responses, general chat:
ollama run deepseek-v3:8b
# Reasoning and math, you can wait:
ollama run deepseek-r1:7b
# Best quality if you have 32 GB RAM:
ollama run deepseek-r1:14b
# Maximum code quality:
ollama run deepseek-coder-v2:7b
# Suppress R1's think phase for faster output:
ollama run deepseek-r1:7b --think=false
What AI Home Server Adds
AI Home Server is a preconfigured mini PC with 32 GB RAM — the threshold where DeepSeek-R1 14B runs comfortably. It ships with DeepSeek pre-downloaded, a browser chat interface (Open WebUI), a Telegram bot for mobile access, Home Assistant for smart home, and Jellyfin for media. Plug in, connect to your network, working in 2 minutes.
The difference from a DIY setup: we have already done the RAM sizing, the model selection, and the software configuration. You do not spend an afternoon discovering that 8 GB is too little.
See Also
- Ollama Benchmarks: 6 Hardware Configs Compared
- Ollama Models 2026: Which Model for Which Task
- Ollama in Docker: Performance Overhead Measured
Frequently Asked Questions
Can I run DeepSeek without a GPU? Yes. CPU-only on 16 GB RAM gives 8–10 tok/s with the 7B model — usable for non-interactive tasks. For interactive chat you want either 32 GB RAM or a GPU.
Is local DeepSeek really uncensored? The model weights have no content filter code. Responses are based purely on training data. Unlike the cloud service, there is no additional filtering layer applied at inference time.
R1 or V3 for daily use? V3 for conversational tasks and general writing — faster, no think-pause. R1 for math, code debugging, and logical reasoning — the chain-of-thought actually produces better results on those tasks.
What is the difference between R1 and R1 Distill? DeepSeek-R1 671B is the full model (requires hundreds of GB RAM). R1 Distill variants (7B, 14B, 32B) are smaller models trained to mimic R1's reasoning patterns — they are what you actually run locally. Quality is surprisingly close to the full model on most tasks.
Pre-configured AI home server with DeepSeek pre-installed on proven 32 GB hardware. Cost: 300 USDT. Message on Telegram: @onoutnoxon — we will match the hardware to your workload.
Tested and written by Alexander Ruin, founder of onout.org. Benchmarks run April 2026. Ollama 0.6.2 on Ubuntu 22.04. Updated: 2026-04-15.
Sources: - https://ollama.com/library/deepseek-r1 - https://ollama.com/library/deepseek-v3 - https://github.com/deepseek-ai/DeepSeek-R1 - https://huggingface.co/deepseek-ai - https://www.datacamp.com/tutorial/deepseek-r1-ollama - https://github.com/ollama/ollama/blob/main/docs/gpu.md
Created with AI tool support for outline and consistency checks; reviewed by Alexander Ruin, founder of onout.org for onout.org. Updated: 2026-05-01.
About the service "DeepSeek R1 vs V3 vs Coder Locally: RAM Benchmarks and Censorship Tests on Real Hardware 2026"
We ran DeepSeek R1, V3, and Coder variants locally on 8–32 GB RAM hardware. Token speeds, RAM usage at each quantization level, and what actually gets censored vs what doesn't.