THX-02: 100,000,000 tokens in our first 4 days Four days ago we open-sourced HAL-X-AI/THX-02, a 4.5B decision model that thinks. Since then, more than 100,000,000 tokens have gone through THX-02. Developers are already using it for: - pull-request review and agent guardrails; - legal and public-sector documents; - bookkeeping, support triage and receipt reading. THX-02 reads text, JSON or images and returns a calibrated probability for every option you define. When a question needs a calculation or multi-step logic, it writes a short thought first. Otherwise it answers directly, in about 80 ms. It scores 65.06 on Decision Index 0.3. The API is free to try. No key is needed to start, and you can integrate it anywhere:
- Model (Apache-2.0): HAL-X-AI/THX-02 - Docs and free API: https://api.hal-x.ai/docs/thx-02 Thank you to everyone who built with THX-02 in its first days. We believe in intelligence for everyone, and with everyone.
Experimental JQK and global target bits‑per‑weight quantization of Nanbeige/Nanbeige4.3-33B.
Unlike standard llama.cpp quantization that relies on fixed type heuristics (e.g., Q4_K_M), the Target BPW approach automatically optimizes per-tensor precision where it matters the most, and produces high quality models that meet a precise global size target.
Key Advantages: - Better model quality: using JQK types can improve KLD between 10% to 50% compared to standard QK types, at the same bpw. - VRAM Maximization: Can generate high quality models sized exactly to fit hardware constraints (e.g., fitting the model into exactly 24GB VRAM). - Data-Driven Precision: Quantization mix is determined by actual weight error sensitivity rather than hardcoded rules, often yielding better PPL/KLD size trade-offs.
“Summarize a Wikipedia article in at least 300 words. Use no commas. Split your answer into three titled sections.”
This is the kind of task used in IFEval, a benchmark of how well models follow multiple requirements at once.
Suppose the model breaks one requirement. Did it lose track of an instruction? Did the no-comma rule make the summary harder to write? Or was summarizing the article difficult to begin with? A program can tell us what failed, but not separate instruction load from task difficulty and interacting constraints.
There is another approach: synthetic tests where a model retrieves a key's value from a long list. Those are easier to control, but they test information retrieval rather than the execution of multiple instructions.
So how can we measure the load itself?
That's the question behind my new article on the IEC instrument. Instead of an essay, the model gets meaningless text made of random words. Instead of complex constraints, it gets simple replacement rules such as lake -> stone. Its job is to rewrite the text while applying every rule.
The simplicity serves a purpose. Each rule applies to its own word, so we can add instructions without changing the operation. We can also set how many different words the text has and how often each repeats. And every application has a known correct outcome.
A single response thus produces many individually verifiable trials, not just one overall score. This lets us study how a model's behavior degrades under different kinds of load.
The article walks through the instrument, an actual prompt and GPT-4o mini response, and how to reproduce the trials. Code and data are public.