Posts

I trained Qwen2.5 of 0.5B, 1.5B, 3B and 7B with QLoRA and QDoRA, keeping the same recipe within each pair. The question: if the model is already going to be trained in 4 bits with LoRA, is it worth activating DoRA? There were eight official training runs. Then, I evaluated Baseline, QLoRA and QDoRA at the four scales, always with the same 100 GSM8K problems. This resulted in 12 conditions and 1.200 generations. ...

When I started designing how quantization should work in Huyawo, the first question was the most obvious one: how much could I reduce the size of a model without compromising its quality? It’s a good question, that’s true, but I realized it wasn’t enough. Normally, quantization is presented as a compression problem. You start with a model in FP16 or BF16, convert its weights to 8 or 4 bits, and compare the result with the original. If the model starts consuming much less memory, continues scoring well on the selected benchmarks, and does not show a very large increase in perplexity, the tendency is to consider the conversion successful. ...

When I turn on my computer in the morning, one of the first things I do is open whichever Ollama Cloud account I’m currently using and check my consumption. I have three Ollama Cloud accounts, each on the US$20 plan, which gives me considerably more peace of mind regarding usage, even though I still can’t afford to consume resources recklessly. Ollama Cloud’s billing model isn’t based solely on the number of tokens you use, but also on how demanding a model is in terms of GPU resources. In practice, this means I can get significantly more work done with GLM-5.2 (classified as “High” usage) than with DeepSeek V4 Pro (classified as “Extra High” usage). ...

Honestly, it had been a long time since a model surprised me like this — not since the end of 2025. No, I do not mean it surprised me because I liked a response or was impressed by a benchmark. No, and in this article you will understand where I am going with this. I remember the launch of Sonnet 4.5 as if it were yesterday. It was one of those moments when the developer community noticed a considerable leap in quality across all existing models — and, make no mistake, the qualitative leap was undeniable. Deeper reasoning, almost human contextual understanding, truly functional code in complex projects. In my head, I said: “Damn! I do not need anything better.” And, in a way, that was not an exaggeration. Sonnet 4.5 elegantly solved problems that previously required workarounds and patches. ...

June 2026. If you are still paying more than fifty cents per million output tokens in any agentic pipeline, forgive my frankness: you are literally burning money. No, this is not an exaggeration on my part, but simply the relationship between API costs and absurd usage — because, indeed, LLMs should be used absurdly, whether you are a solo researcher like me or a startup. And here, the reason has a name, surname, and version: DeepSeek V4 Flash. And what it represents goes far beyond excellent benchmark results — it represents the consolidation of a trend that demonetizes the already psychologically consolidated commercial models and pushes the center of gravity toward SLMs, the highly specialized Small Language Models. ...

(Freepik/Antonio Franco) The landscape of parameter-efficient fine-tuning has been going through a quiet revolution since 2021, and most people haven’t noticed yet. While LoRA democratized LLM fine-tuning by drastically reducing memory requirements — letting researchers train models on consumer GPUs instead of needing expensive clusters — it always had this nagging problem: it couldn’t quite match the performance of full fine-tuning. The gap wasn’t huge, maybe 2–5% on most benchmarks, but it was there. And in competitive research or production systems, that gap matters. ...