The numbers: local vs cloud, side by side

Every week we run identical prompts through our AI Arena benchmark rig — an NVIDIA RTX 5060 Ti with 16 GB VRAM for local models (Ollama), plus cloud APIs for hosted models. This week's test: generate an English article from a short prompt. Here is what we measured on 23 July 2026:

At first glance this looks like a blowout for DeepSeek: 4.6× the raw tokens-per-second. But there is a second number in that table that deserves more attention — the time-to-first-token.

TTFT: the number that actually matters to your workflow

Time-to-first-token (TTFT) measures how long you wait between hitting Enter and seeing the first character appear. For DeepSeek V4 Pro, that wait was 65,699 milliseconds — over a full minute. For gemma4:12b running locally on our RTX 5060 Ti, it was 137 milliseconds. That is a 480× difference in favour of the local model.

Why the massive gap? DeepSeek V4 Pro is a reasoning model. Before it writes a single word, it runs an internal chain-of-thought — analysing the prompt, planning its response structure, considering alternatives. The "thinking" tokens are invisible to you, but they are counted and billed. Our benchmark captured 99 input prompt tokens, but DeepSeek almost certainly generated thousands of reasoning tokens internally before the visible output began. That is where those 66 seconds went.

Gemma 4, by contrast, is a standard autoregressive model — it starts predicting the next token immediately. No prelude, no internal monologue. Just output, from millisecond one.

What this means in practice

If you are writing an article, coding, or chatting interactively, the user experience difference is enormous:

• With DeepSeek V4 Pro: you paste your prompt, press Enter, and then… nothing happens. For over a minute. Eventually, the model starts dumping text at a blistering 196 tok/s — the entire 4,096-token response appears in about 21 seconds of actual generation. Total wall-clock time: ~87 seconds. The result will likely be well-structured because the model thought hard about it before writing.
• With gemma4:12b locally: you press Enter and text starts flowing immediately — 137 ms delay is imperceptible. But the generation pace is a relaxed 42.5 tok/s, so the 3,567-token response takes ~84 seconds to finish. The total wall-clock time is nearly identical: ~89 seconds.

Both models finished in roughly the same time — about a minute and a half. But they feel completely different. DeepSeek's experience is "wait… wait… flood." Gemma's is "start right now, read along as it writes." For brainstorming, iterative prompting, or any back-and-forth workflow, that 137 ms TTFT is the difference between flow and frustration.

Speed ranking: English article generation

Cost: what you actually pay for each approach

DeepSeek V4 Pro pricing (as of July 2026): $0.435 per 1M input tokens (cache miss) and $0.87 per 1M output tokens. At current exchange rates, that is roughly €0.40 and €0.80 per 1M tokens. For our benchmark run — 99 input tokens + 4,096 output tokens — the cost was approximately €0.0033. That is cheap per run, but it scales: 100 article-generation requests per day over a year add up to about €120.

Gemma 4 costs nothing per token because it runs on your own hardware. The upfront investment is the GPU: an RTX 5060 Ti 16 GB currently sells for €450–500 in the EU. Electricity at the European average of roughly €0.30/kWh adds about €0.001 per benchmark run (150 W for 89 seconds). At 100 runs per day, the local setup pays for itself in roughly 18 months compared to cloud API costs — and that is before factoring in that you can run it for anything else: code generation, summarisation, translation, and private data that should never leave your machine.

There is also the privacy dimension. When you send text to DeepSeek's cloud, it leaves the EU — DeepSeek's servers are outside EU jurisdiction. Gemma 4 runs entirely on your desk, no network calls, zero data leakage risk. For companies concerned about GDPR or handling proprietary data, that alone can justify the GPU purchase.

VRAM: gemma4 fits comfortably

The q4_K_M quantisation of Gemma 4 12B peaked at 7,642 MB on our RTX 5060 Ti — well within the 16 GB budget. You have room for a 16k context window, or even running a second small model alongside it. For European developers and small companies, a single mid-range gaming GPU is all you need to get a capable 12B-parameter model running at interactive speeds. Compare that to the old days of needing a €3,000+ workstation card just to get started.

What to choose: a practical decision matrix

If your work involves batch processing — summarise 50 documents, generate structured reports, translate long texts — the cloud model's raw throughput wins. Queue your requests, collect results, and the TTFT penalty matters less. DeepSeek V4 Pro's 196 tok/s chews through content fast once it gets going, and the output quality from its reasoning process may be higher for complex analytical tasks.

If your work is interactive — chatting, pair-programming, iterating on drafts — the local model's instant TTFT is transformative. You are not staring at a spinner. You are reading, reacting, and refining in real time. For many European users, a local model on a €450 GPU plus a cloud API subscription for heavy lifting gives you the best of both worlds.

We are testing more models each week — keep an eye on ai-jarvis.eu/ai-arena for fresh numbers, and try the model comparison tool to see side-by-side results across all our benchmarks.