deepseek-v4.1IMAGE

Resin inference scheduler author's own test: V4.1 Flash 5–10% slower than DeepInfer

Developer Erik Zhang open-sourced the Resin inference scheduler, claiming he built standalone engines for DeepSeek-V4.1-Flash and GLM-5.3-Flash on Blackwell in 12 hours; his own test puts V4.1 Flash 5–10% slower than DeepInfer with no throughput numbers, and the project is not production-grade.

Colby

@dsoch8cArchived 2 days agoOriginally posted 2026.10.27

Resin inference scheduler author's own test: V4.1 Flash 5–10% slower than DeepInfer (image 1 of 3)

Resin inference scheduler author's own test: V4.1 Flash 5–10% slower than DeepInfer (image 2 of 3)

Resin inference scheduler author's own test: V4.1 Flash 5–10% slower than DeepInfer (image 3 of 3)

Colby @dsoch8cOriginally posted 2026.10.27

Open original on X

开发者 Erik Zhang 开源推理调度工具 Resin，自称 12 小时适配两款模型且可匹敌主流引擎。他在文档中称，自己在 Blackwell 显卡上为其配齐 DeepSeek-V4.1-Flash 与 GLM-5.3-Flash 独立引擎。项目开源且可商用，最新更新在北京时间 10 月 27 日。成绩均出自作者自测且无第三方复测，自测称 DeepSeek V4.1 Flash 比 DeepInfer 慢 5%–10%，未公布吞吐数值。作者说明该项目非生产级软件，需要真正推理引擎请用 SGLang 或 vLLM。 https://github.com/ekzhang/resin

x.com/dsoch8c/status/2114840934592761959Archived 2026.10.27 12:09