Which do you think will be better: Qwen-3 or Llama-4

[-]

FriskyFennecFox@reddit

I'm mostly in the Qwen team because they release (some) of their models under Apache-2.0 and don't bother all that much about censoring the data pile for the base model, making them more suitable for making cool finetunes.

I wish Mistral would just drop something already though, can we poke them with a stick?

[-]

somesortapsychonaut@reddit

Ligma

[-]

Sellitus@reddit

Llama 4, Qwen is just a cheap copy going for benchmarks instead of real world performance

[-]

pigeon57434@reddit (OP)

no Qwen consistently performs better than llama on a lot of real world tasks too

[-]

Sellitus@reddit

Not according to technical users, it's just a benchmark chaser if you're doing real work

[-]

Secure_Reflection409@reddit

I'm a technical user.

Nothing comes close to Qwen at 32b Q4.

[-]

Majestical-psyche@reddit

Gemma
Llama
Qwen

[-]

Secure_Reflection409@reddit

Gemma was the OG Qwen, to be fair.

Might well be Gemma again.

[-]

silenceimpaired@reddit

I’m only interested in in local models and Gemma has never released a large enough model to compete with Llama or Qwen in what I do.

[-]

PavelPivovarov@reddit

Llama4 for me.

I heard a lot of good things about Qwen 2.5 but it never impressed me in my day-to-day tasks (apart from qwen2.5-coder which is impressive) while Llama3.1 is my workhorse since it was released.

[-]

silenceimpaired@reddit

I think Qwen was right up against Llama 3.3 performance. Often Llama 3.3 excelled other times Qwen excelled. I found it superior at taking a large chunk of text and transforming it based on complicated instructions. Llama 3.3 seemed superior at longer contexts.

[-]

PavelPivovarov@reddit

For me personally llama3 is absolute winner for sticking to the prompt regardless of the context size, and even with exceeding context window. Nothing else I tried was even remotely close to llama3. Gemma2, Phi3\4 or Qwen start generating rubbish when you push too much context but llama3 somehow still sticks to what the task.

[-]

silenceimpaired@reddit

At lower context lengths Qwen seems stronger for me.

[-]

AppearanceHeavy6724@reddit

Absoltely. Llama 3.1 is a very good balance of coding, acceptable spoken/written language and world knowledge. The only thing I've noticed, not good for rag, as it tends to deviate from the text (not quite hallucinations, but not strict following either). 3.2 in that respect is better, but way too small.

[-]

Healthy-Nebula-3603@reddit

Current state :

Llama even 3.3 for coding is nothing compared to qwen 2.5 32b coder But to everything else llama 3.3 70b would be my choice.. except reasoning ..QwQ here would be better as is using deep reasoning.

[-]

AppearanceHeavy6724@reddit

No one argues about this, it is just as generalist llama is better. Now for coding true qwen 7b is better than most 14b non-coding models, but did you try to use qwen coding models for fiction writing?

[-]

Healthy-Nebula-3603@reddit

For writing..lama 70b 3.3 without the question.

[-]

AppearanceHeavy6724@reddit

depends on your tastes and your hardware I guess. Cannot afford 2x3090.

[-]

MarceloTT@reddit

Truly, the meta LLAMA ecosystem has provided a revolution in research with numerous known use cases. But for the SAAS vertical you need an even more powerful model. I believe that Meta, Alibaba and Deepseek can create spectacular models to improve agent performance, I don't believe they will be useful for critical activities yet, but I hope that the knowledge developed last year and the possible developments and innovations this year will create the necessary technology we will need in 2026 and hope they reduce inference costs to less than 1 cents per million tokens. Only with extremely low costs, high accuracy and performance can we operate with reasoning models on scales unimaginable now.

[-]

2deep2steep@reddit

Qwen obviously, they do better all around work.

[-]

Hoodfu@reddit

Really depends on what you're looking for. Llama is better than qwen at creative writing and handling concepts for text to image prompt expansion with long instructions with lots of examples created by Claude sonnet. With qwen 2.5 a lot of the time, it would really understand what I meant. Mistral 22b small edges llama out a bit in this way but not by a lot.

[-]