As of October 21, 2026, Level lists 364 live jobs that name vLLM, the open source engine DeepSeek's DeepInfer was originally built on. They come from 78 employers. Most of these roles ask you to build AI systems, not just use them.
What happened
DeepSeek published DeepInfer, the inference framework it says powers its large-scale production services. The repository carries an Apache-2.0 license and was created on October 16, 2026. It had 17,000+ GitHub stars when checked on October 21.
The Hacker News thread had 2,200+ points when checked on October 21. In the United States, Google Trends shows no measurable searches for DeepInfer until October 20. It peaked at 100 on October 20, the top of Google's index. The vLLM peak over the same seven days was 42. By the evening of October 21, DeepInfer had fallen back to 2.
The release is limited in scope. The README lists support for one model, DeepSeek-V4.1-Flash. It targets two kinds of hardware: NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs. It covers offline inference only. A KV cache system, prefill and decode disaggregation, and load balancing are listed as parts of a full service that this release does not include.
The README says DeepInfer was originally built mainly on vLLM, with inspiration from SGLang and TensorRT-LLM. The matching model, DeepSeek-V4.1-Flash-1017, is MIT licensed. It was listed first in Hugging Face's trending models on October 21.
Does anyone hire for DeepInfer yet?
Not yet. A full-text search of live Level postings on October 21, 2026 finds none that mention DeepInfer. DeepSeek is tagged as a tool in 49 postings, but that tag does not identify this framework.
Because DeepInfer started from vLLM, the closest skill you can measure today is vLLM. That is what the rest of this page counts.
Who hires people who know vLLM?
On October 21, 2026, 78 employers have at least one live posting that names vLLM. The nine largest by open roles:
- NVIDIA: 92 roles
- Amazon: 40 roles
- CoreWeave: 22 roles
- Intel: 18 roles
- Baseten: 11 roles
- Pragmatike: 10 roles
- SambaNova Systems: 9 roles
- Together AI: 9 roles
- Fireworks AI: 8 roles
NVIDIA alone accounts for 25% of these postings. The full list sits on the vLLM tool page.
Level has 64 postings that name SGLang, and 63 of them also name vLLM. Level has 52 postings that name TensorRT-LLM, and 44 of them also name vLLM. Most vLLM postings, 288 in all, name neither of those engines.
What do vLLM roles look like?
Of these postings, 270 (74%) are rated Builds AI. Another 67 (18%) are rated Works on AI. The remaining 27 (7%) are rated Little AI. Read more about the ratings on the levels page.
By function, 230 of the postings are software engineering. By seniority, 282 are tagged senior or above (77%). Only 61 are tagged intern, junior or mid (17%), so this is not an entry-level corner of the market.
The word inference in the job title is a second way in. Level has 197 live jobs with inference in the title. Of these, 84 also name vLLM. For role titles and skills, see the guides to LLM engineer jobs and MLOps engineer roles.
What do vLLM jobs pay, and are they remote?
Only 197 of the 364 postings publish a yearly range in US dollars. Among those, the median low end is $182,000. The median high end is $261,500. NVIDIA posts 48 of those ranges, so one employer weighs on both medians.
On remote work, 113 postings are remote (31%). The wider remote market is on the remote AI jobs page.
Should I learn vLLM or DeepInfer first?
The postings point one way. Level shows 364 live jobs naming vLLM and none naming DeepInfer. The engine is open source, with an Apache-2.0 license and 94,000+ GitHub stars on its repository.
DeepInfer runs one model offline today, so it does not replace a serving engine. Its README documents asynchronous scheduling, batch-invariant inference, deterministic sampling and structured outputs. Learn the engine employers name, then read DeepInfer as a worked example of those techniques.
