Ainsley Markov

[SuSu_酥酥👅 @NFT_Chen]

🫣原来 Qwen 4.0 迟迟不敢发，是因为看到这出了……

Kimi K3.1 Pro 刚出来一脸灿烂且自信，

结果被 Opus 5.5、GPT-6 Astra、Carbon 按在地上群殴了，

墙后的智谱 GLM-5.5 一直在暗中观察，

Qwen 4.0 也在旁边直接吓出表情包：

“这……这波热度我可不接啊！！！让我��练一练！！！”

现在懂了，不是炼丹慢，

是怕一出来就被群殴然后被甩锅……

#KimiK31Pro #Qwen4 #智谱GLM55 #AI内卷现场 #千问迟迟不发 #Opus55 #Carbon #Astra

[Ainsley Markov @BabushkaRobot]

https://t.co/QxSBh1IVEO

[Ainsley Markov @BabushkaRobot]

ultra4 cluster engine / 4 × Apple M3 Ultra 512 GB / Kimi-K3.1-Pro-2bit-UVMAX / tool-eval-bench (1cd7309e) --- 97/100 --- new local record!

[Ainsley Markov @BabushkaRobot]

start tool-eval-bench 1cd7309e ---> Kimi-K3.1

[Ainsley Markov @BabushkaRobot]

ultra4 cluster engine / 4 × Apple M3 Ultra 512 GB / Kimi-K3.1-Pro-2bit-UVMAX / tool-eval-bench (1cd7309e) --- 97/100 --- new local record!

[Ainsley Markov @BabushkaRobot]

@ivanfioravanti yep )

[Ainsley Markov @BabushkaRobot]

Kimi K3 across 4 Macs

[Ainsley Markov @BabushkaRobot]

k3 load...

[Ainsley Markov @BabushkaRobot]

@ivanfioravanti Please test blender design any prompt

[Ainsley Markov @BabushkaRobot]

@ivanfioravanti I think I'll make the release in 2–3 weeks.

[Ainsley Markov @BabushkaRobot]

DeepSeek V4.1 across 4 Macs

[Ainsley Markov @BabushkaRobot]

What are we running next? Thinking Kimi K3... 🤔⚡️

[Ainsley Markov @BabushkaRobot]

DeepSeek-V4.1-Flash (oQ4e-ngram-mtp) - 1st Run
• Hardware: 4 x M3 Ultra 512GB (ultra4 engine)
• Benchmark: llmprobe 0.6.17 - Code
• Prefill: 2,027 tok/s
• Peak Decode: 119.5 tok/s

[Ainsley Markov @BabushkaRobot]

@ashxhart ofc bro )

[Ainsley Markov @BabushkaRobot]

tool-eval-bench | beamster--DeepSeek-V4.1-Flash-1017-Sushi-STQ-1.75bpw · from 1 M3U 512 vs GLM-5.3-MLX-6bit-MTP · from 4 x M3U 512 :) 92/92

[Ainsley Markov @BabushkaRobot]

tool-eval-bench 887b5375 --- 4 x M3U 512 GLM 5.3 6-bit (32 min.) --- VS --- 1 M3U 512 beamster--DeepSeek-V4.1-Flash-1017-Sushi-STQ-1.75bpw (28 min.) = 92/92 wtf... :)

[Ainsley Markov @BabushkaRobot]

@ItsmeAjayKV https://t.co/ScbuNcRQhd

[Ainsley Markov @BabushkaRobot]

GLM 5.3 6-bit / TP4 4xM3U 512 / tool-eval-bench / 94% (part 2)

[Ainsley Markov @BabushkaRobot]

GLM 5.3 6-bit / TP4 4xM3U 512 / tool-eval-bench / 94% (part 1)