Casper Hansen
@casper_hansen_
Followers
8K
Following
2K
Media
398
Statuses
4K
NLP Scientist | AutoAWQ Creator | Open-Source Contributor
Joined August 2019
2.1k stars, 2+ million downloads, and 7000+ models on Huggingface later, and I am officially ready to retire my long-time project AutoAWQ ⚡️. Proud to say that AutoAWQ has been adopted by the @vllm_project and will now be maintained by 55+ contributors 🥳
8
12
140
Kimi K2 did what we know works. Muon and WSD, DeepSeek V3 arch, and a bunch of tokens. It’s that simple. No need to change your data mixture or the expert routing mechanism while training (*hrhr* Llama 4).
6
5
117
Prior to Kimi K2 release, I beta tested with Claude Code. Here is what it cost me:.- $3.43 for 816 requests. Noticably, output tokens is low, suggesting they are not breaking the bank. What if you used Claude 4 Opus? My bet is $100-200.
8
8
187
Can someone do a dense Kimi 2 32B by finding the optimal experts?.
1
0
9
too much control over training invites constant tweaking. lock in your best guess for data + config, then commit. cancel the run if there are signs of failure.
0
0
1
It’s beautiful how the Huggingface thesis has been playing out for a while. They own the distribution center of the world of AI. Your model does not exist to most people if not on Huggingface.
1
1
21
Predicted this a bit of time ago. The signs were there. Reasoning model, DeepResearch model, and now the most powerful open-source model.
S-tier AI Labs (Chinese):.- DeepSeek (R1).- Bytedance Seed (veRL, Seedance).- Moonshot AI (Kimi 1.5 ≈ old R1, Kimi-Researcher).
0
0
3
The only reason Kimi K2 is not on OpenRouter is that Chutes has not hosted it yet. Actually amazing lol.
7
2
51
Meta offers being sent out as we speak😭.
BTW, the founder used to work at Meta.
0
0
4
imagine when they add o3 type of reasoning to this. it will be so over, instant sota.
🚀 Hello, Kimi K2! Open-Source Agentic Model!.🔹 1T total / 32B active MoE model.🔹 SOTA on SWE Bench Verified, Tau2 & AceBench among open models.🔹Strong in coding and agentic tasks.🐤 Multimodal & thought-mode not supported for now. With Kimi K2, advanced agentic intelligence
4
0
25
The AMD comeback is insane, partially carried by SGLang, vLLM, and even xAI. It will be interesting to see what happens with MI400.
1
0
15
I have been trying Kimi K2 for the last few days. It’s trained to be a Claude Sonnet competitor and a drop-in replacement in Claude Code. It’s a DeepSeek V3 architecture with 1T parameters 32B active!.
1
0
18
All you need to try Grok 4 is LMArena. Available for free and in direct chat. I wish this was available for free in the X or Grok app
0
0
5
"error happened, factory reset your phone to continue". guess it will be more than 15 minutes. .
0
0
1
i need uv but for iphone. ain't no way it should take 15 minutes to setup.
2
0
5
I have an arcane iPhone XS and just bought an iPhone 16. I wonder if I will feel a difference when I pick it up.
0
0
4
Releasing data, models, and code openly and FIRST is the only right way. Then you can always write a paper with more details on what you did. This is the open way.
0
0
5
>Meet Dan Hendrycks.>Creates MMLU.>Labs adopt it.>Labs ace it.>Creates Humanity's Last Exam.>The final closed-ended academic benchmark. That lasted about until Grok 4 was released.
0
0
7