Artificial Intelligence thread

S

shinanhua

Just Hatched
Registered Member
It is worth anticipating the true performance of official DeepSeek V4. While the lack of multimodal capabilities results in subpar front-end presentation, it has demonstrated near state-of-the-art competitiveness against leading mainstream models in my series of task benchmark tests.
Judging by current performance, both Kimi K3 and DeepSeek V4 deliver highly competitive capabilities, and GLM-5.3 is also well worth looking forward to.
 
H

HighGround

Senior Member
Registered Member
Please, Log in or Register to view URLs content!


1784223319322.png
I did say this whole Fable 5/Mythos thing was just hype and a marketing exercise. Not that Anthropic isn't a legit lab. Anthropic and OpenAI are probably on the cutting edge...

But it's obvious that China is just as capable, right on their heels, and can do it for way cheaper.

EDIT: Just to make it clear. Zephyr said these were leaked rather than officially released. Something to keep in mind. Im sure we'll get the official numbers soon enough.
 
bsdnf

bsdnf

Senior Member
Registered Member
We built a kernel optimization arena to see if Kimi K3 could truly optimize the GPU kernel from start to finish. Each model ran independently in its own GPU sandbox, with completely consistent tasks, testing frameworks, and production loads. Within a maximum of 24 hours, models could analyze existing implementations, rewrite kernels, and repeatedly run benchmarks to verify performance.

The arena included two hardware platforms, three types of kernels, and four tasks: Attention Residuals and KDA linear attention on an NVIDIA H200 GPU, a 512-head-dimensional MLA kernel implemented from scratch, and a KDA task on a domestically produced GPU. Under maximum thinking intensity, Kimi K3's performance approached Fable-5 (including fallback mechanism) and significantly outperformed Opus 4.8, GPT-5.6 Sol, and GPT 5.5.
Click to expand...
 
M

meedicx

Junior Member
Registered Member
I've been testing out Kimi K3 for the past few hours. Very strong model. It is much slower than K2.7 (especially the K2.7 high speed I've been spoiled by), but is very rigorous. It doubles/triple checks everything trying to find issues from all angles; I feel I can trust it analyzing data or make changes without introducing random regressions somewhere else in the code base.

I've only been running it for data anlaysis tasks so far and haven't tried other areas. But it other testers and benchmarks shows that it's very strong on front-end coding

1784228443313.png
 
bsdnf

bsdnf

Senior Member
Registered Member
meedicx said:
I've been testing out Kimi K3 for the past few hours. Very strong model. It is much slower than K2.7 (especially the K2.7 high speed I've been spoiled by), but is very rigorous. It doubles/triple checks everything trying to find issues from all angles; I feel I can trust it analyzing data or make changes without introducing random regressions somewhere else in the code base.

I've only been running it for data anlaysis tasks so far and haven't tried other areas. But it other testers and benchmarks shows that it's very strong on front-end coding

View attachment 178298
Click to expand...
GPT-5.6 Sol Ultra also has the same problem; it will be much better when the K3 is released with lower inference levels.
 
M

magmunta

Junior Member
Registered Member
Just my opinion; I have been assigned to translate old Delphi, it's a programing language based on Pascal, codes into modern python codes. For a few days, GPT-5.6 Sol, that costs 200$/month, can't do sh!t about it. Honestly, it's not trivial though. GPT-5.6 Sol is , in my opinion, no better than a good and intelligent programmer. Probably better than majority of programmers, but it's nowhere near the top.
 
You must log in or register to reply here.
Top