Andy Davies @TopCoder_Andy

Another exciting announcement “we plan to open up the model weights in the coming days”

MiniMax (official) @MiniMax_AI

a day ago

MiniMax H3: Omni-Reference, Commercial-Grade Generation, Unbeatable Cost Efficiency, Open Weights x.com/i/article/2082…

112 298 2K 285K 797

what actually keeps firms on the cloud isn't cost or capability, it's the belief that running your own model is some enterprise scale infra project. it isn't anymore. one box on a desk, one open model that reads documents, images, and video, benched and ready today. your files never leave the room, and setup is closer to installing an app than standing up a datacenter. the hard part was never the tech. it's deciding your clients' data is worth keeping in the building.

Sudo su @sudoingX

a day ago

an open message to every business and firm running on claude or gpt. every client contract you paste into that chat box, every id scan, every confidential file, leaves your building the second you hit enter. it lands on a server you don't own, under terms you didn't write and

21 26 158 14K 59

This is likely the best post training DeepSeek has ever done

xjdr @_xjdr

23 hours ago

if this is accurate ... oh boy

67 72 2K 191K 252

Ok so both Flux 3 and Minimax H3 (Hailuo-03) will be open weights. Those of you who tried both, what's your take so far when it comes to video generation?

Anthropic engineer: “90% of our engineers were using self‑improving loops. Now everyone shifted to building agentic Graphs" "No more prompting” In 10 minutes she shows her full Claude Code setup and workflow live, from a blank terminal. Worth more than a $500 agentic course. Watch this video, then save the article below on how to become a graph architect👇

rody @0x_rody

5 days ago

x.com/i/article/2081…

21 39 362 968K 2K

Qwen3.6-35b-NVFP4 on your @NVIDIAAI DGX Spark has been upgraded✨ 95 tok/s single session 317 tok/s at 8 sessions 256K context • 24 concurrencies • vLLM 0.26 This is one of the best models for a single spark! Get it here: github.com/MiaAI-Lab/Unsl…

NEW on my AI Coding Daily website: AI Coding TOOLS database! I started my "hand-picked collection" with 23 tools: aicodingdaily.com/tools?mtm_camp… The plan is to shoot video reviews for many of them, and fill the list with many more. Any popular tool you would add to the list?

Our lab just released our AI Behavioral Observatory open source. It lets you run statistically valid tests on how AI behavior changes under various types of prompts. We have been using it for our own studies & I think it could help others do similar work. gail.wharton.upenn.edu/research-and-i…

Extract text, layout, tables, and formulas from documents into structured Markdown with TurboOCR, using a single GPU pipeline powered by PP-OCRv6 and TensorRT. github.com/aiptimizer/Tur…

Cursor is now available on iPad for all paid users. The redesigned interface uses the larger screen to keep agent chats pinned in the sidebar, display complete file changes and place pull-request reviews beside active conversations. Users can monitor several agents simultaneously, review comments and checks, change reviewers, approve or merge pull requests and ask an agent to resolve feedback. The app also supports screenshot annotations and freehand instructions using Apple Pencil.

Cursor @cursor_ai

2 days ago

Cursor is now on iPad. All the power of Cursor on iPhone, with more room to work with agents.

221 275 4K 731K 615

Datacenters in space? He has one already there. Introducing @PhilipJohnston, founder/CEO of @Starcloud_ which has plans to put 88,000 datacenters up in space. We talk about how it all works and how real it all is.

Shares 12 chapters of Jupyter notebooks implementing Transformers for time series, computer vision, audio, video generation, and reinforcement learning. github.com/Nicolepcx/tran…

Anthropic engineer watched his agents burn through $2.3M in compute before finding the one fix that stopped it prompt engineering taught him to talk to one model and graph engineering taught him to run a whole team of them instead Linear AI → Fake Edges → Diamond Pattern → Verification → Human Gate → System Design most agents still run like one guy wearing every hat, researching then writing then verifying then publishing, all in one head and waiting on itself the whole way through the pattern winning in production right now is plan then split into parallel workers then force a separate reviewer to actually challenge what they found then merge, because trusting the first answer is exactly what gets agents in trouble never let the model grade its own homework, a separate reviewer checking accuracy and freshness and logic catches what self review just doesn't one missing approval step is genuinely how you burn through millions overnight, so publish and send and approve stay behind a human while everything else runs on its own bookmark this and rebuild your next agent as a graph instead of a chat, because the agents earning money instead of burning it are the ones built this way

wandermist @wandermist

7 days ago

x.com/i/article/2080…

11 27 173 252K 443

also Interpretability work keeps finding internal reasoning structure that are not language based...

I spent a few minutes diving into this philosophy in my "Dependency Cultures" talk from Software Should Work a couple weeks ago. Now the recording is up! youtu.be/E82ly38YEEQ?si…

Mitchell Hashimoto @mitchellh

2 months ago

Fork your dependencies, trim them to only your use case, never update unless it breaks for your users. I’ve been vocal about this for 10+ years. I’ve always said that updating is way riskier than latent bugs (which can be tracked and CVEs monitored). If you are updating a

290 770 9K 1.2M 3K

Excited to launch MLX.fast with @eigenlabs today. It's an open autoresearch competition to make Laguna XS 2.1 inference as fast as humanly (and agentically) possible on consumer Macs. Eigen's agents already found 36.8% faster inference, and that's before the competition even started. The best part of open weights is that the community takes a model further than any of us could. 
Can't wait to see what everyone does on the leaderboard!  mlx.fast

Gajesh @gajesh

4 days ago

At 6% the size of DeepSeek V4 Pro, Laguna S 2.1 delivers significantly higher performance. Today, we’re teaming up with @poolsideai to make Laguna XS 2.1 (same architecture as Laguna S 2.1) even faster on consumer Macs. We’re launching MLX(.)fast: a public competition where

15 37 214 95K 65

Kimi K3 has taken first place on Design Arena’s Slides leaderboard. The open-weight model earned an Elo rating of 1,379 and won by the largest margin Design Arena says it has recorded in this category. Design Arena ranks models through continuous competitions where users compare the actual outputs and vote for the one they prefer. So this is not simply Kimi K3 scoring well on another academic benchmark. People preferred the finished presentations it produced.

Design Arena @DesignArena

4 days ago

BREAKING: Kimi K3 lands in 1st on Slides Arena (Python-PPTX) by DesignArena This open-weight model has an Elo of 1379 and wins 1st by the largest margin we've seen on Slides Arena! We'll be following up with an example and deep dive shortly. Congratulations to the

24 76 941 87K 114

Open-source AI financial advisor that simulates scenarios and provides direct guidance based on your actual bank balance and debts. github.com/cdinnison/ray-…

This is epic

Anshu @anshuc

4 days ago

I ran Claude Opus 5 in one marathon session for 24 hours straight, and ended up with this game. No external assets or code. Every pixel and sound was built by Claude. Prompts, workflow, code, and link below. (sound on!)

190 242 4K 521K 3K

look at this. 50 tok/s on qwen 3.6 27b, mobile 5090, with mtp on. for reference i get about 35 on that same card without it. mtp is multi token prediction, the model drafts several tokens per forward pass instead of one, so decode speeds up when the drafts land. 35 to 50 on a 27b would be a serious free jump. but i don't take numbers on faith, i verify them. so two things. if you're running qwen 3.6 27b with mtp, drop your card and your tok/s below, let's build a real picture. and i'm taking my own 5090 laptop for a spin, posting exactly what i get, mtp on and off, same model, same conditions. if it holds, that's free speed almost nobody's talking about. let's find out.

🐜t0nil0🏭🔥🌒🐜 @t0nil0

5 days ago

@sudoingX qwen 3.6 27B with MTP (50t/s on mobile RTX5090)

2 0 31 11K 16