This article's credibility suffers a little from the way it talks about GPT-4o mini:

"just in front of GPT-4o-mini, which is, according to itself, a model with 1.3B or 1.5B or 1.7B parameters, depending on when you ask."

Then later:

"On the Artificial Analysis benchmark Scout achieved the same score as GPT 4o mini. A 109B model vs a 1.5B model (allegedly). This is ABYSMAL."

Asking models how many parameters they have doesn't make sense.

There is absolutely no way GPT-4o mini is 1.5B. I can run a 3B model on my iPhone, but it's a fraction of the utility of GPT-4o mini.

It's strange that someone from FastML can be confused about this, unless it's supposed to be a bad joke.

This is just someone's personal blog/opinion. I wouldn't read too much into it... "The site is run by Zygmunt Zajc (pronounced “Ziontz”). ... An economist by education"

Correct, in that models know nothing about themselves other than what they are told. Deepseek R1 will tell you that it's created by OpenAI.

GPT-4o mini is supposed to be ~8b params from estimates.

The best source I've seen for the 8B number is this TechCrunch article: https://techcrunch.com/2024/07/18/openai-unveils-gpt-4o-mini...

> OpenAI would not disclose exactly how large GPT-4o mini is, but said it’s roughly in the same tier as other small AI models, such as Llama 3 8b, Claude Haiku and Gemini 1.5 Flash.

As far as I can tell all of the 8B rumors were seeded by that loosely sourced comparison to Llama 3 8B.

I know for a fact that Gemini 1.5 Flash is NOT an 8B model, because a separate model called "Gemini 1.5 Flash 8B" came out after that article was published - the first Gemini model with a documented parameter count. Flash 8B is priced at half the cost of regular Flash.

There's also this paper that mentions 8B but provides no source for that at all, which makes me suspect their initial source may have been that TechCrunch rumor: https://arxiv.org/pdf/2412.19260

It's strange because there's no need to make this assumption about GPT-4o in order to demonstrate their point.

we are in for a lot of pain if seemingly intelligent people make mistakes like this. grabbing the number of params from what gpt gives you. how can you do that?

There were actually multiple bugs which impacted long context benchmarks and general inference - I helped fix some of them.

1. RMS norm eps was 1e-6, but should be 1e-5 - see https://github.com/huggingface/transformers/pull/37418

2. Llama 4 Scout changed RoPE settings after release - conversion script for llama.cpp had to be fixed. See https://github.com/ggml-org/llama.cpp/pull/12889

3. vLLM and the Llama 4 team found QK Norm was normalizing across entire Q & K which was wrong - accuracy increased by 2%. See https://github.com/vllm-project/vllm/pull/16311

If you see https://x.com/WolframRvnwlf/status/1909735579564331016 - the GGUFs I uploaded for Scout actually did better than inference providers by +~5% on MMLU Pro. https://docs.unsloth.ai/basics/tutorial-how-to-run-and-fine-... has more details

Do you think there are more bugs in Llama 4 at this time? Or have the bugs been patched, and the current version of llama.cpp + whatever the latest GGUF version is would be representative of the true performance of Llama 4?

I see you've uploaded new Maverick GGUF/safetensors files yesterday, along with a lot of other models like Deepseek R1, was there an issue with the older model files?

If you game the benchmark then you always get found out by your users. Yet the practice remains common in hardware. Outright lies are uncommon but misleading and cherry picked numbers are pretty much standard practice.

The fact that misleading benchmarks don't even drive profit at Meta didn't seem to stop them doing the same thing, but perhaps this isn't very surprising. I imagine internal incentives are very similar.

Unlike the hardware companies though, gaming the benchmark in LLMs seems to involve making the actual performance worse, so perhaps there is more hope that the practice will fade away in this market.

This seems to be a general problem at the moment. The most usable models are not the newest. The newer models (obviously, I haven't tried them all) may do better on benchmarks, but actual usability is worse.

To create useful LLMs required some genuine breakthroughs. It seems to me that we have reached the limits of what we can achieve with current architectures. Progress will require some new insights and breakthroughs.

> This is a draft. Come back later for the final version.

There are quite a few issues with the content from a factual point-of-view (several sibling comments mention things): could have done with a lot more proof-reading and research I think.

I don't understand the rationale for publishing an article in such an early draft stage, even with the small disclaimer at the top. It would make sense if only a bit of polish were missing, but when there are factual errors (that are not marked as such), it seems much better to delay publication until the content is correct.

The initial Llama 4 release is disappointing: the models are too big for most people to run, and not high quality enough be worth running if you can afford the hardware.

I'm still optimistic for Llama 4.1 and 4.2.

Llama 3 got exciting at the 3.2 and 3.3 stages: smaller models that were distilled from the big ones and ran on a laptop (or even a phone).

3.2 3B and 3.3 70B were really interesting models.

I'm hopeful that we will get a Llama 4 ~25B, since that seems to be a sweet spot for laptop models right now - Gemma 3 27B and Mistral Small 3.1 (24B) are both fantastic.

did Meta open a time wormhole to release Llama 4 on May 5th?

>GPT-4o-mini, which is, according to itself, a model with 1.3B or 1.5B or 1.7B parameters

I have no idea how the author can remotely trust GPT-4o-mini in this case. The number of parameters is almost certainly way off.