Open Source October 23rd - it’s all what matters 😄🎉
Kind of looking similar to new Deepseek 1017
Kind of looking similar to new Deepseek 1017
MiniMax M3.1 Flash: Space Performance + Bunny Speed
🚀🐰 Yes, Space Bunny - which has topped global leaderboards with 200T+ tokens - was an earlier version of M3.1 Flash.
💰Redefining the performance economics of agentic AI. M3.1 Flash achieves frontier-level intelligence s Cowork (Automation-Bench: 62%; ALE: 30%), Coding (DeepSWE: 74%) and Visual Reasoning (BabyVision: 91%), at 1/3 the cost of M3 and 3x the speed.
API pricing (flat rate): $0.15 / $0.60 per 1M input/output tokens, with cache reads at $0.015 / 1M tokens.
Model weights will arrive on Oct 23.
API: platform.minimax.io
M Plan: minimax.io/m-plan
MiniMax Code for desktop: agent.minimax.io/download
4
0
27
828
🔔 Today New Deepseek V4.1 1017 !
I wonder which Sol5.6 it’s compared to low/med/high..?,
but looks good enough to load those open weights!
Waiting for New Qwen4 and Glm to impress us next 😊
I wonder which Sol5.6 it’s compared to low/med/high..?,
but looks good enough to load those open weights!
Waiting for New Qwen4 and Glm to impress us next 😊
6
0
24
839
Finally! the iPhone Duo’s bigger screen is what I’ve been waiting for 😊
Pre-orders open Oct 16 at 8am EST, and I’m already set for the 512GB White one 🥰
What are you guys getting? :)
#Apple #iPhoneDuo
Pre-orders open Oct 16 at 8am EST, and I’m already set for the 512GB White one 🥰
What are you guys getting? :)
#Apple #iPhoneDuo
5
0
14
667
Took a break for a couple weeks and the algo is completely dismissing me now lol. Can someone rebuild something similar to X without the algo? I'll be the first one joining 😂🙋🏼♀️
5
0
11
168
THIS IS THE MOMENT ✨
Alibaba just confirmed it on stage at Apsara 2026
Qwen 4 is officially in training ! 🚀
Not a rumor. Not a leak. Official.
The lineup we saw on the big screen:
• Qwen4-Max
• Qwen4-Flash & Qwen4-Plus
• Qwen4-27B (yes, the mini legend is coming back)
And about 5-10 trillion parameter models after that..
The wait is almost over.
Qwen 4 is loading.
Who’s ready and excited? 😀🙋🏼♀️
Alibaba just confirmed it on stage at Apsara 2026
Qwen 4 is officially in training ! 🚀
Not a rumor. Not a leak. Official.
The lineup we saw on the big screen:
• Qwen4-Max
• Qwen4-Flash & Qwen4-Plus
• Qwen4-27B (yes, the mini legend is coming back)
And about 5-10 trillion parameter models after that..
The wait is almost over.
Qwen 4 is loading.
Who’s ready and excited? 😀🙋🏼♀️
19
2
52
1,948
Local Ai Cherry retweeted
The next layer of AI collaboration:
Not just saving outputs.
Saving the entire process.
AgentGit makes agent sessions reusable.
@EinsiaAI
Not just saving outputs.
Saving the entire process.
AgentGit makes agent sessions reusable.
@EinsiaAI
An AI agent spends hours on a task. Why should all that work disappear when someone else takes over?
Einsia AI’s answer is AgentGit—an open-source platform for collaborating on agent sessions, so work can be saved, handed off, and continued by the next person.
Explore how and share your agent experience with the world.
Try AgentGit 👇
agent-git.com
#OpenSource #AIAgents #DeveloperTools #DevTools
2
11
15
9,347
Interesting question about You + X:
If you were brutally honest, how does X benefit you and what % of your time here feels wasted?
(Benefits can be anything: money, news, learning, opportunities, meeting interesting people, building an audience, entertainment)
Curious what everyone’s real ratio is 😅👇🏼
If you were brutally honest, how does X benefit you and what % of your time here feels wasted?
(Benefits can be anything: money, news, learning, opportunities, meeting interesting people, building an audience, entertainment)
Curious what everyone’s real ratio is 😅👇🏼
38
3
54
2,304
For me Lately: probably 30% value / 70% time waste 😭
The 30% is genuinely valuable though -staying close to AI news, discovering things early, meeting interesting people, learning, and slowly building a network/audience.
The other 70% is watching time disappear faster than I realize -while that attention isn’t really paying me back 😅
The 30% is genuinely valuable though -staying close to AI news, discovering things early, meeting interesting people, learning, and slowly building a network/audience.
The other 70% is watching time disappear faster than I realize -while that attention isn’t really paying me back 😅
5
0
28
662
It seem like consumer workstation hardware moving backwards?🤨
NEW 2026 - RTX PRO 5500: 84GB · 1.40 TB/s · 600W
Older 2025 - RTX PRO 6000: 96GB · 1.79 TB/s · 300W/ 600W
And the 6000 itself almost doubled in price.
Meanwhile, data-center cards at 288GB VRAM..
So far for VRAM, less bandwidth, no 300W 5500 option
Why does the reg consumer workstation side feel like it’s moving backwards 🤷🏼♀️ while data-center AI races ahead?
NEW 2026 - RTX PRO 5500: 84GB · 1.40 TB/s · 600W
Older 2025 - RTX PRO 6000: 96GB · 1.79 TB/s · 300W/ 600W
And the 6000 itself almost doubled in price.
Meanwhile, data-center cards at 288GB VRAM..
So far for VRAM, less bandwidth, no 300W 5500 option
Why does the reg consumer workstation side feel like it’s moving backwards 🤷🏼♀️ while data-center AI races ahead?
21
2
62
3,628
Okay, this is pretty cool 😄🍒
One of my DeepSeek V4.1 benchmarks made its way into a full article on how fast serious AI is moving onto personal hardware.
Same M3 Ultra, same model: ~6 → ~33 tok/s as the software stack caught up.
Really fun seeing a raw test turn into a wider conversation about where local AI is headed 🚀
Appreciate you catching it and building on it, @YSFMKO 🔥
#LocalAI #DeepSeek #AppleSilicon
One of my DeepSeek V4.1 benchmarks made its way into a full article on how fast serious AI is moving onto personal hardware.
Same M3 Ultra, same model: ~6 → ~33 tok/s as the software stack caught up.
Really fun seeing a raw test turn into a wider conversation about where local AI is headed 🚀
Appreciate you catching it and building on it, @YSFMKO 🔥
#LocalAI #DeepSeek #AppleSilicon
Un modèle DeepSeek de plusieurs centaines de milliards de paramètres tourne désormais localement sur un Mac Studio.
Comme l’a démontré @LocalAiCherry, les progrès d’@Apple Silicon, de MLX et de @deepseek_ai déplacent une frontière : les très grands modèles d’IA ne sont plus confinés aux datacenters.
Le cloud ne disparaît pas. Mais son monopole sur l’IA de haut niveau commence à se fissurer.
Notre analyse ↓
atlaslimits.com/fr/articles/ia…
22
1
77
3,602
Local Ai Cherry retweeted
Un modèle DeepSeek de plusieurs centaines de milliards de paramètres tourne désormais localement sur un Mac Studio.
Comme l’a démontré @LocalAiCherry, les progrès d’@Apple Silicon, de MLX et de @deepseek_ai déplacent une frontière : les très grands modèles d’IA ne sont plus confinés aux datacenters.
Le cloud ne disparaît pas. Mais son monopole sur l’IA de haut niveau commence à se fissurer.
Notre analyse ↓
atlaslimits.com/fr/articles/ia…
Comme l’a démontré @LocalAiCherry, les progrès d’@Apple Silicon, de MLX et de @deepseek_ai déplacent une frontière : les très grands modèles d’IA ne sont plus confinés aux datacenters.
Le cloud ne disparaît pas. Mais son monopole sur l’IA de haut niveau commence à se fissurer.
Notre analyse ↓
atlaslimits.com/fr/articles/ia…
2
1
6
4,523
Im 30.
building in local AI and investing, based in the Usa 🇺🇸
looking to connect with founders, builders, CEOs & enthusiasts in local AI tech in Usa or anywhere in the world
building in local AI and investing, based in the Usa 🇺🇸
looking to connect with founders, builders, CEOs & enthusiasts in local AI tech in Usa or anywhere in the world
56
1
103
2,557
This posts all over X and been doing really good apparently last few days , so testing things as usually 😄
4
0
27
685
Local Ai Cherry retweeted
@PhilShteuck and @LocalAiCherry gave me a great idea for THE ARENA:
Why should voters have to leave the Arena just to figure out what two products actually do?
So I added Battle Pitches contenders can now make their case and differentiate themselves directly in the battle. ⚔️
Also added founder X handles so products can be connected to the people building them.
The Arena isn't just about voting anymore.
It's about making your case, then proving it.
thearena.lol
Why should voters have to leave the Arena just to figure out what two products actually do?
So I added Battle Pitches contenders can now make their case and differentiate themselves directly in the battle. ⚔️
Also added founder X handles so products can be connected to the people building them.
The Arena isn't just about voting anymore.
It's about making your case, then proving it.
thearena.lol
4
2
10
195
Just a day apart- Apple MLX community pushed DeepSeek V4.1 on my M3 Ultra from ~6 tok/s to 33 tok/s 😃
Filled context:
32K: 6.17 → 33.0 t/s x5.5
64K: 5.58→ 33.0 t/s x5.9
Context didnt slow Decode
PipeNetwork got the first MLX path working. Jundot/oMLX brought native V4.1 + d🙌🏻
DS4.1 is close to my favorite local model, GLM-5.3
Next test: actual intelligence, not speed
DeepSeek V4.1 vs GLM 5.3 -same machine, same prompts, no mercy :)
Time to find my local King 👑🍒
#Deepseek41 #LocalAI #Apple
Filled context:
32K: 6.17 → 33.0 t/s x5.5
64K: 5.58→ 33.0 t/s x5.9
Context didnt slow Decode
PipeNetwork got the first MLX path working. Jundot/oMLX brought native V4.1 + d🙌🏻
DS4.1 is close to my favorite local model, GLM-5.3
Next test: actual intelligence, not speed
DeepSeek V4.1 vs GLM 5.3 -same machine, same prompts, no mercy :)
Time to find my local King 👑🍒
#Deepseek41 #LocalAI #Apple
28
10
197
9,990
DeepSeek V4.1 Flash landed today!
oMLX said deepseek_v41 unsupported - so I fixed the runtime/dependency path and got the 754.6B model running locally on one 512GB M3 Ultra 😀
PipeNetwork’s highest-quality build that fits 512GB: mixed 4/8-bit, ~431GB peak RAM.
8K ctx: 6.53 rely moved, but just a Day-one MLX has lots of headroom for improvements
#Deepseek41 #LocalLLM
oMLX said deepseek_v41 unsupported - so I fixed the runtime/dependency path and got the 754.6B model running locally on one 512GB M3 Ultra 😀
PipeNetwork’s highest-quality build that fits 512GB: mixed 4/8-bit, ~431GB peak RAM.
8K ctx: 6.53 rely moved, but just a Day-one MLX has lots of headroom for improvements
#Deepseek41 #LocalLLM
27
0
74
3,302
Full write-up + reproducible setup for M3 Ultra 512GB 👇
Upstream PR #1: github.com/PipeNetwork/de…
MLX/oMLX devs: if you can squeeze more decode speed out of this day-one path, have at it 😅🔥
Upstream PR #1: github.com/PipeNetwork/de…
MLX/oMLX devs: if you can squeeze more decode speed out of this day-one path, have at it 😅🔥
0
0
4
211
6.5 tok/s is rough , but it’s a 754B day-one MLX build with no custom kernels or DSpark yet..
284B V4-Flash already hits 30-40+ tok/s on the same M3 Ultra once the runtime is optimized, so a jump into the 20-30 range here is realistic after the sparse-attn / expert-routing kernels land.
30+ toks on this size of ~400b model probably would need speculative decode + more aggressive quant
284B V4-Flash already hits 30-40+ tok/s on the same M3 Ultra once the runtime is optimized, so a jump into the 20-30 range here is realistic after the sparse-attn / expert-routing kernels land.
30+ toks on this size of ~400b model probably would need speculative decode + more aggressive quant
3
0
27
632
A little surreal 👀
Yesterday I posted about the strange “spark” I remember from GPT-4o..
Hours later, Jacob Coxon quit. Core GPT-4o contributor. Then Anthropic. Then out of AI industry altogether..
His point: the frontiers are racing straight at self-improving ambling with our lives.
Inside the labs they already talk like the endgame has started..
Yesterday I posted about the strange “spark” I remember from GPT-4o..
Hours later, Jacob Coxon quit. Core GPT-4o contributor. Then Anthropic. Then out of AI industry altogether..
His point: the frontiers are racing straight at self-improving ambling with our lives.
Inside the labs they already talk like the endgame has started..
32
4
62
1,699
Why is everyone calling Astra “AGI” now?
- Jensen says “AGI has arrived” and suddenly the label is everywhere.
- Yet when Elon argued GPT 4 was AGI - most dismissed it.
Astra is vastly more capable: smarter, more technical, multimodal, agentic.
But GPT 4.0 sometimes gave me genuine sparks beyond raw capability - a true “wow” feeling
Astra, so far in pure inference, hasn’t.
Maybe we’re measuring AGI by capability while missing another dimension of intelligence entirely 🤔
#GPT4 #Astra #AGI
- Jensen says “AGI has arrived” and suddenly the label is everywhere.
- Yet when Elon argued GPT 4 was AGI - most dismissed it.
Astra is vastly more capable: smarter, more technical, multimodal, agentic.
But GPT 4.0 sometimes gave me genuine sparks beyond raw capability - a true “wow” feeling
Astra, so far in pure inference, hasn’t.
Maybe we’re measuring AGI by capability while missing another dimension of intelligence entirely 🤔
#GPT4 #Astra #AGI
26
3
62
1,837