Video
CommunityJev Repos & SDKs
Open Jev-Compatible API (SGLang)
An independent Jev-compatible public API backed by Qwen3.6-35B-A3B and SGLang radix caching, reporting 64 parallel typed-decision tasks in under a second.
- Format
- Video
- Reading time
- 2 min
- Published by
- Community
- Author
- Eric Zhang
- Last verified
- Sep 2026
Works forJev for Classification
Highlights
- Serves a Jev-shaped API over Qwen3.6-35B-A3B.
- Uses SGLang radix cache for prefill reuse; 64 tasks in under a second.
- Open-sourced as openjev-sglang for anyone to try.
Watch out
An independent reproduction, not hosted Jev; answers and calibration differ from TypeSafe's model.
Questions about this entry
Short answers drawn from the linked project or page.
A different model. It serves Qwen3.6-35B-A3B through a Jev-compatible interface, and the author frames it as an experiment to see how well an off-the-shelf model works. Answer distributions and calibration should not be assumed to match TypeSafe's hosted model.
The tasks share a common prefill, so the radix cache preserves that prefix reuse and one batch amortizes it across parallel System One generations. The under-a-second figure is the author's own result on their Modal endpoint, not a throughput guarantee for your workloads.
Yes, openjev-sglang is open on GitHub and the post's endpoint is a Modal URL you can aim a frontend at for a quick trial. Self-hosting means bringing your own accelerator for Qwen3.6-35B-A3B, which is the real cost question the post does not address.
Treat it as a playground. The post invites people to point a frontend at it as a quick experiment, and a reply in the thread notes that open Jev claims and reproductions are proliferating with no way to tell which ones actually work. Run it yourself if it matters.
Eric Zhang, who describes himself as a computer systems person and interaction designer at thinkymachines and previously a founding engineer at Modal. The public endpoint is a Modal URL, so the demo runs on infrastructure he knows well, which fits his framing of the project as a quick experiment rather than a product.
Yes. A reply from apolinario points to a Hugging Face Space that categorizes thirty-eight artefacts, split between GitHub repositories and Hugging Face models, specifically to separate signal from noise among the flood of open Jev claims and to note which ones run on a laptop. That is a useful index for navigating this entry.
How well an off-the-shelf model works in a Jev-shaped harness. He says explicitly that it is just an experiment, and the value is that the request shape is identical to Jev's while the model behind it is an ordinary open Qwen, so it isolates how much of the experience comes from the interface rather than the model.
Reactions & coverage
Posts, threads, and videos about this entry from around the web.
X: 19 open-source Jev projects
X: Jev repositories worth a look, in Japanese
From the Jev guides
Original write-ups on what this entry is for.
From the Jev AI model community
Posts from builders shipping with Jev right now.