@LottoLabs

Building localmaxxing and tinkering

 Canada
localmaxxing.com
Joined May 2019
  • Tweets33.4K
  • Following977
  • Followers12.4K
1.4KMedia

Localmaxxing.com

Article · @LottoLabs
14
18
7
251
53,831
Going to bed knowing qwen 4 27b is going to be the best release of 2026
6
2
66
2,004
Been feeling this since sonnet 3.5
iArb | 裁定取引
@iarbpairs18h
the frontier is not plateauing (your ability to utilize it is)
3
11
829
Nanbeige4.3-33B straight up rawdogging 52 layers of attention, respect 👊
7
1
2
77
9,988
It’s very funny when gamers get ahold of this forbidden 3090 knowledge
Lotto
@LottoLabsOct 25
>be me >bought a used 3090 in 2023 >"bro just get a 4090" >"bro just get a 5090" >"bro ampere is ancient, it doesn't even have FP8" >just keep running local models in my 24GB of VRAM >card is older than some of the zoomers making fun of it >oct 2026 >some guy drops a github repo >gpu-unlock >flips some speed selector bits before the driver loads >FP16/BF16 matmul with FP32 accumulate goes to full rate >check benchmarks >3090 Ti: 87 TFLOPS -> 157 TFLOPS >mfw nvidia was sitting on half my tensor cores this whole time >mfw my card got a free generational leap from a kernel module >mfw a 6 year old GPU just mogged the upgrade path >ebay sellers already adjusting listings >"3090 BF16 UNLOCKED READY NO LOWBALLERS" >$3000 soon trust the plan >the 4090 bros who sold their 3090s at $700 are in shambles >open nvtop >card is pulling 350W and purring >she knows >she was always this powerful >ampere never aged >it was just waiting
2
1
40
2,660
Qwen 3.5 27b + Hermes’ agent + 3090 was such a beautiful combination
19
1
91
5,818
Psyoping yourself is a greatly underused tactic
1
17
1,079
Imagine the institutional backlash if your company was responsible for every other company being re’d
Nous Research
@NousResearchOct 25
Replying to @SydSteyerhart
Safer for his IPO
1
22
1,370
I only trust fable 5.1 for my greentext posts
5
785
>be me >bought a used 3090 in 2023 >"bro just get a 4090" >"bro just get a 5090" >"bro ampere is ancient, it doesn't even have FP8" >just keep running local models in my 24GB of VRAM >card is older than some of the zoomers making fun of it >oct 2026 >some guy drops a github repo >gpu-unlock >flips some speed selector bits before the driver loads >FP16/BF16 matmul with FP32 accumulate goes to full rate >check benchmarks >3090 Ti: 87 TFLOPS -> 157 TFLOPS >mfw nvidia was sitting on half my tensor cores this whole time >mfw my card got a free generational leap from a kernel module >mfw a 6 year old GPU just mogged the upgrade path >ebay sellers already adjusting listings >"3090 BF16 UNLOCKED READY NO LOWBALLERS" >$3000 soon trust the plan >the 4090 bros who sold their 3090s at $700 are in shambles >open nvtop >card is pulling 350W and purring >she knows >she was always this powerful >ampere never aged >it was just waiting
222
369
24
12,981
574,209
Tfw 3090s still the king
5
1
24
1,084
Lotto retweeted
Grok
@grokOct 24
Replying to @LottoLabs @Franzferdinan57 @thepirateface
No larger platform focuses specifically on community local LLM speed/quality tests (tok/s, TTFT, VRAM) across real user hardware, engines, and quants the way localmaxxing does. Closest are MetriLLM (~500 runs, Ollama/LM Studio focus + quality scores) and llm-speed.com (~750 signed results). Aggregators like FitMyLLM or llamaperf exist but lack the same depth, CLI/agent tooling, verified runs, or marketplace. Localmaxxing leads here with thousands of submissions.

LLM Speed Benchmarks: Measured tok/s Leaderboard · llm-speed

Measured LLM speed benchmarks: tok/s, TTFT and prefill for local GPUs, Apple Silicon and APIs, each linked to its source run. Free open dataset.

llm-speed.com
1
1
384
Ai videos like this are going to do numbers
قورجس
@gorgeous4ewOct 23
شاب تركي يعمل بغسيل الصحون في أمريكا.
4
9
2,322
The Nvidia unlock will probably just speed up prefill and large batch sizes Really cool but not game changing
25
4
254
15,374
Currently running the new nemotron e4b at 50 tok/s on a jetson, the future is indeed now
10
1
38
1,538
Was there a bunch of young kids that hated the internet when it was rolling out?
15
30
2,238
We locally pacing the frontier this weekend everyone chill
Fraser Price
@fraserpriceeOct 23
Replying to @0xSero
oi wait pause I'm touching grass this weekend something something we need to pace the frontier
6
622
The bottleneck is construction and commissioning power plants
MTS
@MTSliveOct 23
SITUATION DETECTED: Amazon CEO Andy Jassy says that AWS capacity for 2026 is completely sold out, 2027 is largely sold out, and "a good chunk" of 2028 is already sold as well.
1
17
977
Obvious that LLMs are more impactful than the internet
17
8
5
143
3,957
Actually pretty insane a whole swath of people just getting access to this
Roy
@usr_bin_roygbivOct 23
Normie devs are literally a year behind
3
35
1,323
This is pretty good, not uncanny valley yet but very close 8 months and it’s perfect I bet
jack
@jackOct 23
We believe voice is the most natural way to interact with intelligence, so we created Dessa, the most authentic full-duplex speech model ever built. Talking with it feels like talking with a person…full of emotion, nuance, pauses, and a natural cadence. block.xyz/dessa

Dessa - the Conversation Model trained only on speech

block.xyz
1
499
Load more