s/brian_armstrongMODEL ROUTING•Sep 22
127
votes
6.4k
seen
Jev's deepseek-v4.1-flash-jev returns candidate probabilities, not streamed text
Jev is moving from a standalone decision model into the routing layer for coding agents. A new Claude Code mod sends each request through Jev to classify the subagent model, the main model at session start, and the required effort level.
The same push includes a demonstrated DeepSeek endpoint, deepseek-v4.1-flash-jev, built with SGLang’s scoring API. It forces candidate-answer probabilities instead of streamed generation, adds a closing think tag for direct DeepSeek responses, and exposes structured choices and probabilities to software. One Dot API deployment is listed at $0.040 per million input tokens.
The same push includes a demonstrated DeepSeek endpoint, deepseek-v4.1-flash-jev, built with SGLang’s scoring API. It forces candidate-answer probabilities instead of streamed generation, adds a closing think tag for direct DeepSeek responses, and exposes structured choices and probabilities to software. One Dot API deployment is listed at $0.040 per million input tokens.
Timeline2
Sep 18
Developers demonstrated a DeepSeek-based Jev endpoint using SGLang scoring and DeepSeek-specific response formatting.
Sep 19
A Claude Code model-router mod was published, routing requests through Jev to classify models and effort levels.
1 comment
Sep 22