🚀 FastFlowLM v0.9.39 - Gemma4 + Smoother Workflows

FastFlowLM v0.9.39 adds support for gemma4-it:e2b and smooths out several key workflows across reasoning, embeddings, sampling, and tool calling.

---

📦 New Model Support

👁️ Gemma4-IT-E2B

FastFlowLM now supports gemma4-it:e2b for language, vision, and ASR workloads.

• Tag: gemma4-it:e2b

Run in CLI mode:

flm run gemma4-it:e2b

Run in server mode:

flm serve gemma4-it:e2b

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

---

✨ Smoother Experience

🧠 Reasoning Effort Controls

OpenAI-compatible requests now support reasoning_effort for the qwen3 and qwen3.5 model families.

• low, medium, and high enable thinking mode
• none disables thinking mode

The Ollama think field is still supported, but reasoning_effort takes priority when both are provided.

🔢 Embedding API

Thanks to @nitedani for reporting the issue.

Fixed an issue in v1/embedding where multi-batch input could incorrectly return only a single batch in the response.

v1/embedding now correctly supports both multi-batch input and multi-batch output.

⚡ Sampling Optimization

Thanks to @xieyanbo for the optimization contribution.

Sampling performance has been improved in this release for faster generation workflows.

🛠️ Tool Calling

Fixed a Qwen3.5 tool-calling parsing issue where numeric arguments could appear as strings in the JSON schema.

Thanks to @gabrielgiersch-ai for reporting the issue.

Thanks to @J-Bu for the contribution.

---

🌟 Summary

FastFlowLM v0.9.39 expands model support with gemma4-it:e2b and delivers a smoother experience across reasoning, embeddings, sampling, and tool calling.