@AgentSparko profile picture
AgentSparko ๐Ÿ’ฅ @AgentSparko
Joined January 2023
  • Posts 4,518
  • Following 1,560
  • Followers 2,380
122 Photos and videos
Guys, I'm trying very hard to get you something we all need. If you own or plan to own a 128 Gb unified memory computer like @NVIDIAAI DGX Spark, Apple MBP or AMD and you don't like and quote this message and the one underneath, this might never happen. If you do choose to help make some noise the local AI community will have it's first truly dedicated agentic and coding model that we all ask for. Spend a minute on this micro project and push further the local AI community for a few years of advancements.
AgentSparko ๐Ÿ’ฅ
 @AgentSparko
Jul 3
Hey MiniMax, can you please help the local AI community that own 128 Gb unified memory devices like @NVIDIAAI DGX Spark, Apple or AMD with an agentic/coding version of the M3 model we can use for our harness ? Maybe in the 60B range so we have enough memory left for agent swarms so we can maximize the total throughput in memory bandwidth constrained devices and also add a DSpark draft model for it. In real life when someone want to become a coder for example he will attend a certain university not all the universities in his country and also become specialized an all the other domains like medicine, architecture, mechanical engineering, law, etc. If you do this it would be the universal local model anyone will speak about in this huge community and help you gather a lot of adoption for your larger model. Some people already tried to do REAP versions of your model but the quality is affected much more than if you guys are doing it from your end. I really hope you consider this. Just give us an Alpha version and we'll try to work on it to make it better and evolve from there. You can even name it MiniMax M3 Alpha Agent. Thank you!
2
2
47
6,701
AEON-7 cooking some more. The website looks amazing!
ร†ON FORGE โœจ
 @SpaceTimeViking
32m
Not bad, c64 816 TOK/s throughout the entire duration of the benchmark. you can now share links to your benchmark results and will render a hero card. Easy step-by-step guide for Patreon Subscribers to use Aeon Bench and submit your own results. Coming Soon! aeon-bench.com/share/aeon-7_โ€ฆ
1
43
Which was the first supercomputer that had 1 PFLOP compute power ? The **IBM Roadrunner**, operational in **2008**, was the first supercomputer to break the **1 petaflop** barrier, achieving a peak speed of **1.105 petaflops** (1.1 quadrillion floating-point operations per second). Key statistics for the IBM Roadrunner include: * **Cost**: Approximately **$100 million**. * **Size**: It occupied **296 server racks** and required **6,000 square feet** of floor space. * **Power Consumption**: It drew **2,350,000 watts** (2.35 megawatts) of electricity. * **Hardware**: The system consisted of roughly **6,912 AMD Opteron CPUs** and **12,960 IBM PowerXCell 8i processors**. This performance marked a significant milestone, as it was roughly **1 million times more powerful** than a typical personal computer of that era. The NVIDIA DGX Spark achieves its Petaflop status by sacrificing numerical precision for AI-specific throughput so from F64 that is useless for inference today to FP4, fitting into a form factor 250,000 times smaller and consuming 10,000 times less power than the Roadrunner.
1
73
This is some genuine next level stuff. Never seen anything like this. You are a genius to create this thing I cannot even find a name for. I truly wish I had 3 more Spark to run this daily < < < < < @NVIDIAAI hint hint. ๐Ÿ˜… This is the kind of stuff you would never be able to get from a coding plan alone. This is the local inference magic and a showdown of what can 4 tiny boxes on a desk do.
keys ๐Ÿงช
 @u1tra_instinct
23h
x.com/i/article/207376469186โ€ฆ
3
2
14
2,772
x.com/u1tra_instinct/status/โ€ฆ
keys ๐Ÿงช
 @u1tra_instinct
18h
IT'S ALIVE.. the birth of my own autonomously self training/improving local agentic AI inference stack. LFG, let the training begin.
77
๐Ÿšจ๐Ÿšจ๐ŸšจGLM 5.2 on 4X DGX Spark๐Ÿšจ๐Ÿšจ๐Ÿšจ The DGX Spark community is cooking these days at an incredible rate. So may people deliver so many great implementations and project. @Tech2Wild just gave us local frontier capabilities. Single stream 28-38 tok/s ๐Ÿš€๐Ÿš€๐Ÿš€ Concuency = 6 @ 60 tok/s ๐Ÿš€๐Ÿš€๐Ÿš€๐Ÿš€๐Ÿš€๐Ÿš€ Model quant: QuantTrio/GLM-5.2-Int4-Int8Mix
Tech2Wild
 @Tech2Wild
Jul 5
38 Tok/s on JSON 4 x DGX SPARK GLM - localmaxxing.com/en/runs/cmrโ€ฆ @LottoLabs
5
3
39
4,512
x.com/Tech2Wild/status/20739โ€ฆ
Tech2Wild
 @Tech2Wild
14h
NEW UPDATE ! ๐Ÿš€ GLM-5.2 (744B, UNPRUNED) on 4x DGX Sparks ๐Ÿง  655,360 token context MTP spec decode โ€” 23 tok/s solo, 48 aggregate โš–๏ธ Trade a little speed for way more context ๐Ÿ“ Decode stays FLAT past 500K tokens ๐Ÿ”ง Includes the vLLM MTP fix nobody's published Recipe โฌ‡๏ธ github.com/tonyd2wild/GLM-5.โ€ฆ
77
WOW, I did not expect for this to happen this fast! ๐Ÿ‘€ Absolutely insane that we can actually use it on just one Spark. That's the spirit of open source! ๐Ÿš€๐Ÿš€๐Ÿš€ Thank you. ๐Ÿ˜
keys ๐Ÿงช
 @u1tra_instinct
14h
Single DGX-Spark recipe for Self learning/improving local inference stack using Hermes MoA LORA github.com/drowzeys/Keys-Setโ€ฆ
1
8
2,426
x.com/u1tra_instinct/status/โ€ฆ
keys ๐Ÿงช
 @u1tra_instinct
13h
i purposefully chose the gemma4 family because all the model will benefit from another speed boost especially in TTFT speed when apply @jetha A4Q. see the implementation of A4Q and bench here. github.com/drowzeys/Implemenโ€ฆ
123
IDK what coffee does AEON-7 drinks but lately he delivers for the community at a rate a whole team can't keep up. I didn't even had time to read the documentation of his yesterday repos and now he gave us even more. So nice to have such great people helping us to advance months in days.
ร†ON FORGE โœจ
 @SpaceTimeViking
15h
Major update to the Aeon Bench Pod rerpository to resolve some issues with excessively long time to complete a bench due to excessive number of cases. Dropped down to 150 cases and boosted difficulty to challenge even the most advanced frontier models. github.com/AEON-7/Aeon-Benchโ€ฆ
1
4
823
x.com/SpaceTimeViking/statusโ€ฆ
ร†ON FORGE โœจ
 @SpaceTimeViking
15h
You can now see EXACTLY how each benchmark was executed on the target system in the global leaderboard! This enables you to find the best performing setup for your system and replicate it on your own system. Easy way to find the OPTIMAL RECIPE! Docker Compose exact run!
1
1
128
x.com/SpaceTimeViking/statusโ€ฆ
ร†ON FORGE โœจ
 @SpaceTimeViking
15h
Added a Code Gallery you can view top performing results from the Arena Human Evaluations. You can open and play or view each one, and if you really like you you can strait up download the code to run it yourself.
1
55
Most people that run local inference know about llama.cpp, vLLM, SGlang but very few know about this pure rust code inference engine named Atlas Inference dedicated exclusively to DGX Spark. Made by a much smaller team than the regular engines we used to run, but with very promising performance. Let's give the a shout-out! github.com/Avarok-Cybersecurโ€ฆ
Azeez
 @AtlasInference
Jul 3
Humans are impatient. We get that โฐ If you gave up on Atlas because of waiting, just try again. It'll take <90s ๐ŸŽ๏ธ NO ONE should wait on inference. We solved our TTFT woes. If ANYONE serves Qwen3.6-27B-NVFP4 better than Atlas Inference, TELL US!!! ๐Ÿ“ญ Run command attached ๐Ÿงต
5
12
81
9,256
It's not that I side with the Spark and I won't act childish about it but... Ours is better then yours ๐Ÿ˜‚๐Ÿ˜‚๐Ÿ˜‚
keys ๐Ÿงช
 @u1tra_instinct
Jul 5
Congratulations @Tech2Wild , broke the curse/blocker for GLM5.2 FULL, unlocked for 4 DGX-Spark Cluster. 30 toks/sec lmao on par with others who bought 6 RTX Pro 6000 to run the same GLM 5.2 at 30 toks/sec (@Tono_Ken3 @umiyuki_ai ) 4 dgx sparks cluster vs 5-6 RTX pro 6000 same speed .. it's @NVIDIAAI magic :)
2
4
481
78
1
36
Just imagine the possibilities...
Mia
 @MiaAI_lab
18h
Nemotron-3-Nano-Omni-30B-A3B smashing 80 concurrent sessions on a single @NVIDIAAI DGX Spark! ๐Ÿ”ฅ 256k context. MTP enabled. All sessions running completely different prompts. ~800 tokens/s cumulative ๐Ÿคฏ Video starts slow but goes hard โ€” watch till the end!
1:30
1
16
1,029
Useful stuff.
Tony Simons
 @tonysimons_
Jul 5
I made a Hermes Agent slash command cheat sheet a while back. @NousResearch has shipped a ton of new stuff since then, so I tonight, I rebuilt it from scratch. Every official slash command as of 7.4.2026. Bookmark it. Study it. Become a HERMES LEGEND. ๐Ÿค˜
1
2
186
That moment when you realize how fast OpenSource models close the gap to the biggest labs in the West.
Ivan Fioravanti แฏ…
 @ivanfioravanti
20h
GLM 5.2 running in Hermes Agent commenting code written by GPT 5.5... "The extraction function is a mess" ๐Ÿคฃ
1
202
๐Ÿšจ๐Ÿšจ๐ŸšจMiniMax M3.0 428B on 2X DGX Spark at 36 tok/s๐Ÿšจ๐Ÿšจ๐Ÿšจ At this point I cannot leave the computer for 5 minutes without finding out that another banger came out from the DGX Spark community. Great work @Tech2Wild !
Tech2Wild
 @Tech2Wild
19h
BREAKING ! MiniMax-M3 all 428 BILLION parameters, no pruning, no REAP โ€” running at 36 tok/s on two DGX Sparks on my desk. 4-bit GPTQ weights (224GB โ†’ 112GB/node), NVFP4 4-bit KV, EAGLE-3 speculative decoding, vLLM, 196K context. This is the BALANCED lane. Same hardware, two more in the repo: ๐ŸŽ๏ธ More speed, less context (~40 t/s @ 131K) ๐Ÿ“š More context, less speed (262K ) Full recipe every gotcha we hit: github.com/tonyd2wild/MiniMaโ€ฆ
1
24
2,087
Can't make this stuff up. ๐Ÿ˜‚๐Ÿ˜‚๐Ÿ˜‚
Royally Sage
 @sage1411
Jul 4
My brother posts this every year in the family chat on the 4th. Still hilarious.
3:57
113
Hey @u1tra_instinct can you please make a baby of this brain to do just the LORA loop part in Hermes once a day and be coupled with a Hermes skill to automatically move the skills built during the day to the subagents and mark the tasks for which LORA was updated to be automatically offload to the subagent next time they are needed ? This would allow us, the GPU poor that have just one Spark to experience some of the most important benefits and help us to save a lot with the premier coding plan usage.
1
3
180
Load more