🚀 FastFlowLM v0.9.36 - Qwen3.5-4B + Stability Fixes

This release introduces a new model and ships important reliability fixes to keep inference smooth in production workloads.

---

🖼️ 1. New Model: Qwen3.5-4B

FastFlowLM now supports a new vision language model:

• Tag: qwen3.5:4b

You can control image pre-resizing in both run and serve modes using --img-pre-resize (or -r).

Run in CLI mode:

flm run qwen3.5:4b -r 1

Run in server mode:

flm serve qwen3.5:4b -r 1

For more details, see the model card and benchmark results.

---

🛠️ 2. Bug Fixes

We resolved two core stability issues in request handling and sampling:

• CORS: Fixed preflight handling to keep responses alive during async_write. Thanks to @tllewellynn1.
• Sampling: Fixed top-k sampling by clamping values to vocabulary size when top-k exceeds available tokens. Thanks to @jingfelix.

---

🌟 Summary

FastFlowLM v0.9.36 introduces qwen3.5:4b and improves runtime stability with fixes for CORS preflight reliability and top-k sampling bounds.