Models in job postings
Who a company hires is the clearest public sign of what it runs. These are the model families, serving frameworks and fine-tuning tools that open AI postings name, and the employers behind them.
Model families
| Family | Weights | Employers | Postings |
|---|---|---|---|
| Llama | Open weights | 88 | 147 |
| Mistral | Open weights | 67 | 148 |
| Qwen | Open weights | 45 | 72 |
| DeepSeek | Open weights | 29 | 69 |
| Stable Diffusion | Open weights | 17 | 42 |
| Gemma | Open weights | 15 | 27 |
| Whisper | Open weights | 9 | 11 |
| Nemotron | Open weights | 6 | 11 |
| Kimi | Open weights | 5 | 7 |
| Phi | Open weights | 5 | 7 |
| GLM | Open weights | 4 | 11 |
| gpt-oss | Open weights | 4 | 8 |
| Falcon | Open weights | 3 | 4 |
| FLUX | Open weights | 1 | 1 |
| Claude | API | 694 | 1347 |
| GPT | API | 320 | 1279 |
| Gemini | API | 303 | 579 |
Serving and fine-tuning
| Tool | Use | Employers | Postings |
|---|---|---|---|
| vLLM | Serving | 268 | 660 |
| LoRA | Fine-tuning | 162 | 256 |
| SGLang | Serving | 133 | 333 |
| TensorRT-LLM | Serving | 102 | 268 |
| PEFT | Fine-tuning | 81 | 132 |
| Triton Inference Server | Serving | 60 | 99 |
| QLoRA | Fine-tuning | 56 | 81 |
| TGI | Serving | 34 | 50 |
| Ollama | Serving | 26 | 38 |
| llama.cpp | Serving | 25 | 41 |
| NVIDIA NIM | Serving | 9 | 15 |
| Unsloth | Fine-tuning | 3 | 5 |
| Axolotl | Fine-tuning | 3 | 4 |
| LMDeploy | Serving | 2 | 2 |
Employers naming llama.cpp
25 employers whose open postings name llama.cpp. Every employer
Read from the words of each open posting, title and description, and counted by the employers whose postings name a term. A serving framework (vLLM, SGLang, TensorRT-LLM) runs open weights on the employer's own GPUs; a fine-tuning tool (LoRA, PEFT) means it adapts them in-house. A family isn't counted in its maker's own postings. Names that mean other things count only in forms that can only mean the model: “Phi-3” or “Microsoft Phi”, not PHI; “Falcon-40B”, not CrowdStrike Falcon. API models are often named as tools people use at work rather than what a product runs on. 52,255 postings read, updated every few hours. How models are matched.