Lollipop

[Lollipop @lollipop_stat]

@stevibe have you all running on gtx 5090? i tried and it is all garbage as 2bits, which is the only one fitting fully in gtx VRAM

[Lollipop @lollipop_stat]

@guitaripod @IterIntellectus same impression, flash introduces many bugs pro wouldn't

[Lollipop @lollipop_stat]

@guitaripod @IterIntellectus pro or flash?

[goodalexander @goodalexander]

> get super stoked
> 328 tps
> wow that's sick
> sign up for account
> run quick benchmark bc I'm not a trusting person
> it's 58 tps the same as everyone else

thanks for playing

[Lollipop @lollipop_stat]

@DotDotJames @UnslothAI which inference engine and command line you using? i tried days ago but quality was garbage and slower than my current setting "Gemma-4-26B-A4B-it UD-Q4_K_XL (187 t/s, 530 t/s aggregate in 4 slots parallel)

[Lollipop @lollipop_stat]

@AdhiOffX deepseek is one of the most perfomance/token/cheap tradeoff, really great flash v4

[Lollipop @lollipop_stat]

@MiaAI_lab Thanks!

[Lollipop @lollipop_stat]

@MiaAI_lab Which inference engine and command line optimization are you using for deepseek v4 flash? I am barely getting 40 tok/sec

[Shannon Sands @max_paperclips]

one of the saddest things that happens as you age is watching your heros fall

[Diego Rodriguez @asciidiego]

remember the guy who used to write tiny c files and train gpt-2 without a gpu?

he's shilling slack integrations now

[Lollipop @lollipop_stat]

@0x0SojalSec Tried, forse than gemma on triaging..on detection slightly Better but loses on all other metrics. DPO done Is over all bad

[Przemek Chojecki | PC @prz_chojecki]

Kimi 2.7 ranked 2nd after Fable 5 and before GPT-5 xhigh

We have re-run our ErdosBench smoke test on 14 problems with Kimi 2.7, Qwen 3.7 Max, Grok 4.3 and compared it with the top performers from previous runs.

Kimi 2.7 is amazingly good. More below.

[Lollipop @lollipop_stat]

@0xSero you sure Diffusion Gemma26B really works? tried it on 5090 and it is slower compared to autoregression for normal prompts (variable input, variable output)

[Lollipop @lollipop_stat]

@antirez @dantecpp i guess because SSD does not guarantee the ECC that some GPUs do?

[Lollipop @lollipop_stat]

@atomic_chat_hq i am curious which quantization for the 26b a4b? i have 22gb in my vram for Gemma-4-26B Q4_K_M, not 15 gb

[Lollipop @lollipop_stat]

@bendee983 I am same token processing tier with opus 4.6. cache saves 95% context otherwise we would be talking about milions/month.

But inference costs will most likely and hopefully go down in next quarters, this Is the way, let s just wait for it

[Lollipop @lollipop_stat]

@MemoryReboot_ inference costs will go down, i argue

[Lollipop @lollipop_stat]

@ItsmeAjayKV what is the performance on 9900x alone (no gpu used)?

[Lollipop @lollipop_stat]

@leftcurvedev_ Thanks for this service!
Ryzen 9950x, 2x patriot viper ddr5 6000mhz cl32
Gpu AMD W7800 ai top 48gb ECC RDNA3