xevyo-01
a fast action planner for AI assistants
3 min readupdated
This model acts as a pre-router (turn planner) for chatbots: instead of writing a reply to the user, it analyzes their message and outputs a compact JSON action plan for the main model. Built by the Xevyo team on top of Qwen/Qwen3.5-9B using LoRA SFT (supervised fine-tuning via adapters that were later merged into the weights), it produces valid JSON far more reliably and identifies the right tools an order of magnitude faster than both the base model and the larger teacher model, Qwen3.6-27B. It is a great fit for agent developers who need low-latency routing for web search, calculators, memory operations, or file generation.
Key features
- Comprehensive routing: assesses safety levels, determines whether web search is needed (along with ready-to-use queries), and handles image workflows, documents, memory, and specific tool calls.
- High speed: hits a median latency of 0.40 seconds per turn when served via vLLM on an H100-class GPU.
- Accurate planning: delivers 100% valid plans on a 1,610-dialogue test set and reaches an 89.0% F1 score for tool selection, compared to 38.6% for the original Qwen3.5-9B.
- Multilingual support: trained on data in English, Spanish, Hindi, and Hinglish (a blend of Hindi and English).
- Open license: released under Apache 2.0, with the standalone original LoRA adapter also included in the repository.
What you can do
- Place it in front of your chatbot's main LLM: run every incoming message through xevyo-01 via vLLM to execute web searches, run calculations, or pull facts from memory before generating the final response.
- Route multimodal and file tasks: automatically detect when a user wants to generate or edit an image versus assemble a document (PDF, DOCX, XLSX, or PPTX).
- Adapt it to your own tool stack: pass your custom list of available functions in the system prompt, or fine-tune the included LoRA adapter for specific business logic.
Example
pip install "vllm>=0.17.0"
vllm serve meet4150/xevyo-01 \
--served-model-name xevyo-01 --dtype bfloat16 \
--max-model-len 8192 --language-model-only \
--enable-prefix-caching --mamba-cache-mode align \
--reasoning-parser qwen3 --generation-config vllm
Read more
DetailsBenchmark (1,610 synthetic turns): 62.8% exact match across the entire plan (vs. 14.9% for Qwen3.5-9B and 26.8% for Qwen3.6-27B), with p50 / p90 latency at 0.40 / 0.62 s.
Hardware: a single GPU with 24 GB VRAM or more is sufficient (bfloat16 weights in the repo take up 19.3 GB; 9.04B out of 9.65B parameters are active, with the base model's vision components unused).
Training: 3 rounds on 66,666 synthetic examples (generated by Qwen3.6-27B), LoRA rank 32, alpha 64.
Limitations: tested on a context window of up to the last 8 messages (~2,600 tokens); it is not a standalone safety system and requires deterministic filtering on top of its safety scores.










Comments