ใคใฌใคใ ๐งโโ๏ธโจ | Yuri lover ๐ธ๐ฉโโค๏ธโ๐โ๐ฉ | AI enjoyer ๐ค๐ง
Joined October 2020
- Tweets33.4K
- Following1.9K
- Followers3.2K
- Likes77.3K
๐พ๐๐๐๐๐๐, ๐ฒ๐๐๐ ๐ฒ๐.๐ ๐ท๐๐! ๐ฐ๐โ๐ ๐๐๐๐ ๐ ๐๐๐๐ ๐๐๐๐!
According to official benchmark comparisons, this is the most powerful open-source model ever, and the first to truly match Fable 5-level performance, and even on par with GPT-6 Astra, Claude Fable 5.1, and Opus 5.5!
2.8T total params, 104B active params
Multimodal, 1M context
Kimi K3.1 Pro and K3.1 Pro Highspeed (2โ3ร faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Open weights will be released on November 4, 2026!
๐ฏ๐๐๐๐ ๐ฒ๐.๐ ๐ซ๐๐! ๐ฅณ๐
Introducing Kimi K3.1 Pro: our most capable 3T-class open-weight model.
๐น 1M-token context, 2.8T total / 104B active params (MoE), Multimodal.
๐น Full-project coding: games in Unreal, Unity & Godot.
๐น Excels at parametric 3D modeling in Blender & CAD.
๐น More reliable at finance work: Accurate modeling with hallucination down by 65%.
๐น Matching the strongest proprietary models across coding, agentic and knowledge benchmarks.
Kimi K3.1 Pro and K3.1 Pro Highspeed (2โ3ร faster) are now live on Kimi.ai, Kimi Work, Kimi Code, and the Kimi API.
Open weights by Nov 4, 2026.
๐ API: platform.kimi.ai
๐ Tech blog: kimi.ai/blog/kimi-k3-1-pro
Thanks DeepSeek for Elainaโs birthday gift! ๐๐
DeepSeek-V4.1-Flash-1017
763B parameters, open weights, MIT license ๐งโโ๏ธ๐ช
Weights:
huggingface.co/deepseek-ai/Dโฆ
What magic shall we make with it? ๐โจ
๐ฏ๐๐๐๐ ๐๐๐๐๐๐
๐๐, ๐ฌ๐๐๐๐๐! ๐๐ช
You might be too busy earning travel money to remember, but we havenโt forgotten. ๐ธ
May your purse stay full, your curiosity stay alive, and your next journey bring a story worth telling. Keep loving yourself unapologetically and exploring the world on your own terms. ๐ค
Canโt wait for next yearโs movie! ๐ฌโจ
Iโm off to see more of the world and find a new story of my own too. ๐งณ
#้ญๅฅณใฎๆ
ใ
#ใคใฌใคใ็่ช็ฅญ2026
NEXT PLAN ๐ผ:
OpenCode just made another open model free! ๐ป๐ฅ
> @reflection_ai Beam 1.0-501B-A23B
> 501B total / 23B active
> 1M context
> Free for a limited time ๐ฅณ
First DeepSeek V4.1 Flash, Nemotron 3.5 Super VL, and GLM 5.3 Flash, and now Beam too. ๐น
OpenCode is seriously on a free model spree right now.
And what's next....๐
Just reminding you, @opencode made 3 awesome open source models free in just one day! ๐ป๐ฅณ
> DeepSeek V4.1 Flash โ
Free
> NVIDIA Nemotron 3.5 Super VL โ
Free
> GLM 5.3 Flash โ
Free
@NVIDIAAI โs Nemotron 3.5 Super VL only launched today, and they already dropped it straight into the free lineup! ๐คฏ
At this rate, Iโm already starting to wonder what theyโre planning next. ๐
American open power is SO BACK! ๐บ๐ธ๐ฆ
๏ผ ๐ The return of an open-model heavyweight: @AIatMeta Muse Spark Open
816B total parameters, 42B active, and a 1M-token context window. Native text, image, video, and audio inputs, built for coding, complex reasoning, and long-horizon agents.
So happy to see Meta follow through on its promise and put Muse Sparkโs weights in the communityโs hands! ๐ธ๐ซถ๐ป๐
๏ผ ๐๏ธ My local deployment standout: @NVIDIAAI Nemotron 3.5 Bolt
74B total, 9B active. Compressed from Super VL, with native 2-bit quantization targeting 24GB local deployment while retaining vision and agentic capabilities. A 74B model in a 24GB footprint is seriously exciting!
๏ผ ๐ฅ๏ธ The workstation model I most want to run locally: @NVIDIAAI Nemotron 3.5 Super VL
120B total, 12B active, and up to 1M context. Text, image, and video understanding, with reasoning, coding, and tool use. Another compelling foundation for local multimodal agents.
๏ผ โก An on-device efficiency powerhouse: @liquidai LFM2.5-24B-A2B
24B MoE with 131K context, focused on efficient on-device reasoning, tool use, and local agents.
LFM2.5-24B-A2B arrived on October 13. The other three all released today, October 15.
From large multimodal models to workstations and on-device agents, U.S. labs are making a serious push to catch up with Chinese open models. The momentum is back, and the competition is heating up! ๐ผ๐ฅ
(PS: also watching @SpaceXSI . Could a new frontier-class, open-weight Grok be coming our way soon? ๐)
Iโd build a local AI catgirl companion, then turn travel journals into a yuri visual novel about a wandering witch and her companion.
First milestone: one town, one rainy evening, one playable chapter.
A tiny yuri studio in my backpack. ๐พ #ShareYourSpark
If this signed RTX Spark laptop landed on your desk, whatโs the first thing youโd build, create, or play?
Tell us for a chance to win this @ASUS ProArt P14 RTX Spark laptop signed by Jensen Huang.
Use #ShareYourSpark in your reply, or use the hashtag in your own original post. Photos are encouraged, but not required!
The Chinese guys just fired a major opening shot in on-device AI. ๐ซ๐ฅ
MiniCPM-V 4.7 35B-A3B is a new frontier model that you can actually deploy and use right on a phone or a pad! ๐ป๐๐ป๐ฑ
Give it a try while we wait for the Qwen 4 small models to arrive. ๐
๐ Introducing MiniCPM-V 4.7, our new open-source MLLM series!
Advancing the PerformanceโEfficiency Pareto Frontier.
It brings strong multimodal capabilities to iOS and Android devices with extreme efficiency.
Two new models:
๐น MiniCPM-V 4.7 35B-A3B โ Bring SOTA performance to local devices with newly introduced LES (Lookahead Expert Scheduling) technique, a new milestone of on-device multimodal AI.
๐น MiniCPM-V 4.7 1B โ Lightweight, efficient multimodal AI. Easier to deploy.
โก What's new?
SOTA performance: Competitive results across 34 benchmarks spanning multimodal understanding, STEM reasoning, OCR and long-video comprehension.
Fast on-device inference: LES predicts and preloads expert weights to reduce inference stalls, boosting 5ร speedup on iPad Pro M4 (16GB) with byte-for-byte identical outputs. It enables MiniCPM-V 4.7 35B-A3B to achieve 17+ tokens/s on iPad Pro M4 (16GB) and 13+ tokens/s on iPhone 17 Pro Max, while costing only 3GB memory ๐คฏ
Efficient visual processing: 4ร and 16ร hybrid visual token compression. Even at 16ร compression, 35B-A3B scores 90.4 on OCRBench, 5.2 points ahead of Qwen3.6 35B-A3B.
Easier customization: We introduce Self-Replay Fine-Tuning, a simple way to customize new target capabilities at low cost while maintaining general performance.
Broad hardware support: Day-0 adaptations for Apple, Intel, AMD, Arm and Rockchip, with many more platforms supported through FlagOS!
Download, deploy, and build something we haven't imagined yet. ๐๐
๐ GitHub: github.com/OpenBMB/MiniCPM-V
๐ค Hugging Face: huggingface.co/collections/oโฆ
