Reading
An open-source project based on Qwen3-0.6B demonstrates how constrained decoding technology allows small language models with only 600 million parameters to achieve 100% valid JSON function call outputs, completely changing our understanding of the relationship between model size and reliability.
Section 01
Call Me Maybe is an open-source project based on Qwen3-0.6B (600 million parameters). It achieves 100% valid JSON function call outputs via constrained decoding technology, subverting the inherent perception of the relationship between model size and reliability and demonstrating the strong potential of small models under proper engineering guidance.
Section 02
Call Me Maybe was developed by rogard-antoine as part of the curriculum at 42 Programming School, with the goal of converting natural language prompts into structured executable function calls. Traditional small models without processing have only about 30% validity in JSON outputs, which cannot be used in production environments. This project uses constrained decoding to increase this figure to 100%.
Section 03
Constrained decoding ensures structural validity by restricting token selection during generation: 1. An 11-step state machine builds JSON step by step (complete process from array start to end); 2. Token-by-token syntax validation (generate logits → validate valid tokens → mask invalid tokens → safe sampling); 3. Semantic constraints (predefined function selection, parameter type checking).
Section 04
Project performance data: Function selection accuracy >95%, parameter extraction accuracy >90%, JSON validity 100%; Single prompt processing time: 2-3 seconds, batch processing of 100 prompts: 4-6 minutes; Model memory usage: ~2.5GB (fp16), vocabulary size: 100MB, per-prompt state memory <1MB; No crashes in edge cases, graceful handling of malformed inputs.
Section 05
Challenges encountered and solutions: 1. Token mapping issue: Use format_text() to convert special tokens; 2. Function boundary detection: Step5 filters functions that only start with the current token; 3. Parameter type enforcement: Step8 maintains independent allowed token sets for different types; 4. State reset: Create a new JSONState for each prompt; 5. Stop condition error: Adjust the state machine to ensure parameters are processed before entering the end step.
Section 06
Application example: Input prompt "What is the sum of 40 and 2?" outputs the correct function call JSON (fn_add_numbers with parameters a=40, b=2). Insights: Structural guidance is more important than model size; deterministic output is more critical in production; high interpretability; small models + constrained decoding are suitable for resource-constrained scenarios.
Section 07
Call Me Maybe is not just a technical demonstration but also a reflection on LLM application development: an elegant constraint mechanism can surpass the parameter stacking of large models. Project link: https://github.com/ROGARD18/call-me-maybe, which is a valuable resource for developers working on LLM function calls and structured outputs.
Keep going with more reads from the same topic.
Nornir MCP Server is an enterprise-level server based on the Model Context Protocol (MCP). It seamlessly integrates large language models (such as Claude) with the Nornir network automation framework, supporting natural language orchestration for multi-vendor network devices (Cisco, Arista, Juniper, etc.), and providing production-grade features like a dual-engine architecture (NAPALM + Netmiko), intelligent filtering, and a secure sandbox.
Bibliothèque Française LLM is a structured indexing and annotation project for French public domain literature designed specifically for large language models (LLMs). It integrates multiple authoritative sources such as DraCor, Common Corpus, and Wikisource, providing metadata indexing categorized by genre, author, and era, as well as in-depth annotations for dramatic texts (including characters, lines, stage directions, etc.). Its aim is to enable LLMs to efficiently read and understand classic French literary works.
Splinter is a minimalist, high-performance key-value (KV) and vector storage system enabling zero-latency inter-process communication via shared memory and atomic operations. With only 766 lines of core code, it supports millions of operations per second and 768-dimensional vector storage, offering a new architectural approach for local LLM inference and data-intensive applications.
Folkering OS is the world's first AI-native bare-metal operating system, entirely written in Rust no_std without relying on Linux, POSIX, or libc. It can generate commands from scratch, compile them into WASM, and run them in 10 seconds, achieving true self-evolution.