Small language models
120 language models with up to 15B parameters: cheap to call, quick to answer, small enough to run on your own hardware.
 Phi 4 MultimodalMicrosoftMITVision | $0.48/1Mvia AIHubMix | 128K | 5.6B | $0.12/1M | $0.48/1M | — | |
| Phi 4 Reasoning PlusMicrosoftMIT | $0.50/1Mvia Azure | — | 15B | $0.125/1M | $0.50/1M | — | |
| Phi-3-medium-4k-instructMicrosoftMIT | $0.68/1Mvia Azure | — | 14B | $0.17/1M | $0.68/1M | — | |
 Qianfan OcrBaiduApache 2.0Vision | $0.248/1Mvia AIHubMix | 32K | 4.7B | $0.062/1M | $0.248/1M | — | |
 Qwen2-VL-7B-InstructAlibabaApache 2.0Vision | $0.70/1Mvia AIHubMix | — | 8.3B | $0.28/1M | $0.70/1M | — | |
| Qwen2.5 Coder 7B fastAlibaba | $0.09/1Mvia Helicone | — | 7B | $0.03/1M | $0.09/1M | — | |
| Qwen2.5 Math 1.5b InstructAlibabaApache 2.0 | $0.20/1Mvia AIHubMix | — | 1.5B | $0.20/1M | $0.20/1M | — | |
| Qwen2.5 Math 7B InstructAlibabaApache 2.0 | $0.40/1Mvia AIHubMix | — | 7.6B | $0.20/1M | $0.40/1M | — | |
| Qwen2p5 14BAlibabaApache 2.0 | $0.097/1Mvia SiliconFlow | — | 15B | $0.097/1M | $0.097/1M | — | |
| Qwen2p5 Coder 1.5BAlibabaApache 2.0 | $0.40/1Mvia AIHubMix | — | 1.5B | $0.20/1M | $0.40/1M | — | |
| Qwen2p5 Coder 3BAlibabaNon-commercial | $0.03/1Mvia Hugging Face | — | 3.1B | $0.01/1M | $0.03/1M | — | |
 Recurrentgemma 2BGoogleGemma | Free/1Mvia Nvidia | — | 2.7B | Free/1M | Free/1M | — | |
| UReMM SLERP 13BUndi95JSON | $0.65/1Mvia OpenRouter | 6K | 13B | $0.35/1M | $0.65/1M | — | |
 Riva Translate 4B InstructNVIDIAOpen weights · with terms | Free/1Mvia Nvidia | — | 4.2B | Free/1M | Free/1M | — | |
| SSao10K Stheno 8bSao10KCC BY-NC 4.0Reasoning | $0.05/1Mvia Hugging Face | 8K | 8B | $0.05/1M | $0.05/1M | — | |
 Sea Lion 7B InstructAI Singapore | Free/1Mvia Nvidia | — | 7B | Free/1M | Free/1M | — | |
 ThudmZhipu AIMIT | $0.10/1Mvia AIHubMix | — | 10B | $0.10/1M | $0.10/1M | — | |
 UI-TARS 7BByteDanceApache 2.0Prompt caching+2 | $0.20/1Mvia OpenRouter | 128K | 8.3B | $0.10/1M | $0.20/1M | — | |
| TUnslopNemo 12BThedrummerJSON | $0.40/1Mvia OpenRouter | 1.02M | 12B | $0.40/1M | $0.40/1M | — | |
 Zamba2 7B InstructzyphraApache 2.0 | Free/1Mvia Nvidia | — | 7.4B | Free/1M | Free/1M | — |
Frequently asked questions
There is no fixed line; this page lists models with up to 15 billion parameters, the size of Qwen3 8B or Mistral 7B. Sizes come from each model's official repository or the size in its name.
Most of these are open-weight: a 7–8B model runs on a recent laptop or a single consumer GPU once quantized, and the smallest (1–4B) run on phones. Larger context windows need more memory.
For classification, extraction, summaries, routing and simple chat, often yes, at a fraction of the cost and latency. For hard reasoning or broad knowledge a larger model still wins; compare their scores before deciding.
Can't find what you're looking for? Contact us