FreeLLMAll models › google
Arena score1303#287 of 419 in Chat, paid models included
Size4.3 Bparameters
4-bit download2.6 GBweights only, Q4_K_M
Max context128Ktokens, from the model config
Where to use it for free
- The lab's own chat: aistudio.google.com (free with daily limits)
- Run it in one command with whichllm (free, open source):
uvx whichllm@latest plan "gemma-3-4b-it"shows the GPU you need,uvx whichllm@latest run "gemma-3-4b-it gguf"downloads it and starts a chat - Download and run it at home: official weights · GGUF files for Ollama and LM Studio
Arena scores
| Arena | Score | Rank | Votes |
|---|---|---|---|
| Chat | 1303 | #287 of 419 | 4K |
Rank among all models, paid ones included. Source: Arena leaderboard dataset.
Speed on popular hardware
| Hardware | GPU memory | Fits? | Speed |
|---|---|---|---|
| RTX 3060 12 GB | 12 GB | Fits in GPU | 74 tok/s |
| RTX 4060 Ti 16 GB | 16 GB | Fits in GPU | 60 tok/s |
| RTX 3090 | 24 GB | Fits in GPU | 171 tok/s |
| RTX 4090 | 24 GB | Fits in GPU | 182 tok/s |
| RTX 5090 | 32 GB | Fits in GPU | 284 tok/s |
| RX 7900 XTX | 24 GB | Fits in GPU | 175 tok/s |
| 2 × RTX 3090 | 48 GB | Fits in GPU | 166 tok/s |
| 4 × RTX 3090 | 96 GB | Fits in GPU | 157 tok/s |
| Mac M4 Max 128 GB | 96 GB | Fits in GPU | 85 tok/s |
| Ryzen AI Max+ 395 128 GB | 96 GB | Fits in GPU | 54 tok/s |
| 8 × H100 SXM | 640 GB | Fits in GPU | 352 tok/s |
Estimate at 4-bit with 32K context, 64 GB of DDR5 for offload. Speed ≈ memory bandwidth × 78% ÷ data read per token, plus a fixed cost per token, calibrated on 44 real llama.cpp tests; typical error about ±10%. Compare all GPUs.
GPU memory needed
| Level | 8K context | 32K context | 128K context | Quality vs full |
|---|---|---|---|---|
| Q3_K_M | 3.1 GB | 3.7 GB | 6.1 GB | +10.5% noticeable loss |
| Q4_K_M | 3.5 GB | 4.1 GB | 6.5 GB | +2.8% small loss |
| Q5_K_M | 4.0 GB | 4.6 GB | 7.0 GB | +0.9% very small loss |
| Q6_K | 4.4 GB | 5.0 GB | 7.5 GB | +0.3% near-lossless |
| Q8_0 | 5.5 GB | 6.1 GB | 8.5 GB | ≈ 0% same as full |
| F16 | 9.5 GB | 10 GB | 13 GB | ≈ 0% same as full |
Weights + context cache (from the model's architecture) + runtime. Quality: increase in perplexity against the full model, from llama.cpp tests, scaled by model size. Simulate your own setup.
Chat score over time
130014001500
MayJunJulAugSeptOct
Gemma 3 4B IT 1303Best paid model
Arena score from 29 Apr 2026 to 28 Oct 2026.
Similar free models in Chat
- Mistral Small 3.1 24B Instruct (2503) 1303
- Qwen2.5-72B-Instruct 1303 · 72.7 B
- Granite 4.1 8B 1304 · 8.8 B
- Olmo 3 32B Think 1306
- Mistral Large (2411) 1306
- Llama 3.1 Nemotron 70B Instruct 1299
Share this page:XLinkedInRedditWhatsApp
Spotted a wrong number or a broken link? Report a mistake