I tried V3.1 but it was driving me crazy by ignoring parts of user input, which R1 never did. I had many such instances when e.g. asking about running DeepSeek 671B it instead picked DeepSeek 67B because 671B is too large to exist so I must have made a mistake etc. I concluded that despite being better in benchmarks than R1, it was essentially useless due to this characteristics and I instead started using R1 at OpenRouter. Not sure why deepseek.com removed R1 and left only V3.1 without any ability to switch back, I guess it's cheaper to run.
Matches my experience in general as well. I find benchmarks largly useless for comparing current models. Many, despite improved metrics, are strictly worse than predecessors. What little gains they show in some areas, like agentic use here, are often set by far broader and often catastrophic losses.
I wish there was some easy resource to keep up with the latest models. The best I have come up with so far is asking one model to research the others. Realistically I want to know latest versions, best use case, performance (in terms of speed) relative to some baseline, and hardware requirements to run it.
Dumb collation of benchmarks that the big labs are essentially training on. Livebench.ai is the industry standard - non contaminated, new questions every few months.
Thanks! Are the scores in some way linear here? As in, if model A is rated at 25 and model B at 50, does that mean I will have half the mistakes with model B? Get answers that are 2x more accurate? Or is it subjective?
It says the best "coding index" is held by Grok 4 and Gemini 2.5 Pro. Give me a break. Nobody uses those models for serious coding. It's dominated by Sonnet 4/Opus 4.1 and GPT-5.
I use Aider heavily and find their benchmark to be pretty good. It is updated relatively frequently (a month ago, which may be an eternity in AI time).
The fast Cerebras thing got me to try the Qwen3 models. I couldn't get them working all that well: they had trouble using the required output format and following instructions. On the other hand, benchmarks say they should be great, and it sounds like maybe some people use them OK via different tools.
I'm curious if my experience was unusual (it very much could be!) and I'd be interested to hear from anyone who's used both.
Interesting--I'd seen Chinese characters surprise inserted when it was just repeating back input with one provider, but not others. (I'd also occasionally seen tokens surprise-translated to Chinese.)
The language mixup thing seems to be an issue across all LLMs, as soon as you put some Chinese in the prompt they will often randomly respond in Chinese.
Also, given a partly Chinese prompt, Qwen will sometimes run its whole thinking trace in Chinese, which anecdotally seems to perform slightly worse for the same prompt versus an English thinking trace.
I usually use GPT-oss-120B with CPU MoE offloading. It writes at about 10tps, which is useful enough for the limited things I use it for. But I’m curious how Q3 Next will work (or whether I’ll be able to offload and run it with GPU acceleration at all.)
So there are two ways to look at this - both hinge on how your define "consumer":
1) We haven't managed to distill models enough to get good enough performance to fit in the typical gaming desktop (say, 7B-24b class models). Even then though - most consumers don't have high end desktops, so even a 3060 class GPU requirement would exclude a lot of people.
2) Nothing is stopping you/anyone from buying 24ish 5090s (a consumer hardware product) to get the required ~600GB-1TB of VRAM to run unquantized deepseek except time/money/know how. Sure, it's unreasonably expensive but it's not like labs are conspiring to prevent people from running these models, it's just expensive for everyone and the common person doesn't have the funding to get into it.
> 1) We haven't managed to distill models enough to get good enough performance to fit in the typical gaming desktop (say, 7B-24b class models).
That really depends on what "good enough" means. Qwen3-30b runs absolutely fine at q4 on a 24GB card, although that's also stretching "typical gaming desktop". It's competent as a code completion or aider-type coding agent model in that scenario.
But really we need both. Yes it would be nice to have things targeted to our own particular niche, but there are only so many labs cranking these things out. Small models will only get better from here.
The Deepseek provider may train on your prompts: https://openrouter.ai/deepseek/deepseek-v3.1-terminus