Pinned Tweet
RL/ML papers love equations before intuition. This post attempts to flip it: each idea appears only when the previous approach breaks, and every concept shows up exactly when it’s needed to fix what just broke. Reinforcement Learning for LLMs ("made easy") mesuvash.github.io/blog/2026…
Can we have MOE model based on this teacher ?
🚀Excited to announce Nanbeige4.3-33B!
To push the performance ceiling of small agents, our strategy in Nanbeige4.3 starts with building a powerful teacher first.
Nanbeige4.3-33B delivers SOTA agentic performance among open models at ~30B parameters:
* Strong results on agentic coding and cowork benchmarks.
* When integrated into mainstream agent harnesses, it tackles real-world tasks with cost-efficient inference.
* It will serve as a teacher for upcoming smaller Nanbeige models, starting with Nanbeige4.3-3B.
huggingface.co/Nanbeige/Nanb…
huggingface.co/Nanbeige/Nanb…
#LLM #Agent
Replying to @AirbnbHelp
@AirbnbHelp @Airbnb I have raised urgent issue with your customer care and they are not even responding. When I call they say they will call team handling the issue and cut off the call. This has been going on for 3 hours and the host has passive aggressively said to levave.
Despite all drama. What a time to be alive!!
We’re sharing a solution to the Navier-Stokes Millennium Prize Problem, one of the deepest problems at the frontier of mathematics.
The proof was produced by a group of agents, using an OpenAI next-generation model significantly more capable than GPT-6 Astra.
The problem concerns whether the description of smooth three-dimensional fluid motion modeled by the Navier-Stokes equations can break down. It has remained unresolved for roughly 90 years.
This is what real empowerment looks like. Captain Priya Adhikari isn't just breaking barriers for women, she's setting a standard for everyone.
Insane results!!
Big news: Fable 5.1 (Max) by @AnthropicAI just landed #1 in the Code Arena: WebDev with 1765 pts - breaking away from the pack with a huge +77pt margin.
Fable 5.1 (Max) is a significant improvement from Fable 5 (Max) at #8 overall with 1628 pts. It’s +77 pts above Qwen3.8-Max-0902 in the #2 spot, and +78 pts above Opus 5 (Max) in the #3 spot.
With 1765 pts and a blended $40/MToken, it also reshapes the Pareto frontier in the higher priced range by moving the performance bar much higher. See Pareto below.
Congrats to the @AnthropicAI team on this release!
Readers added context they thought people might want to know
The graph truncates its x-axis at 1550 (not zero), visually exaggerating the 77-point lead (actual scores 1765 vs 1688/1687 with far fewer votes on the new model).
arena.ai/leaderboard/co…
arena.ai/leaderboard/co…
A small contribution from my end. I usually don’t share donations, but if this encourages even a few more people to contribute, I don’t mind being a little shameless about it. The funds go directly to the Prime Minister’s Relief Fund to support those affected by the Nepal floods.
Friends! We bought the domain NepalRelief.Fund!
Now there is one easy to remember and easy to share link with all the relevant information you need to support the OFFICIAL Disaster Relief and Rebuilding efforts in Nepal 🇳🇵.
NepalRelief.Fund has
→ The link to the OFFICIAL Government of Nepal @NepalGov Disaster Relief Fund led by the Office of the Right Honorable Prime Minister!
→ Information on our dollar-for-dollar match drive supporting those efforts thanks to SecurityPal AI @security_pal and Renegade Insurance @RenegadeIns via Mission Rebuild Nepal (100% of the proceeds go to the PM's Disaster Relief fund!)
We just crossed $15,000 which means a minimum of $30,000 is headed to the Prime Minister's relief efforts!
But we are NOT STOPPING now!
🌟If your companies have $ for $ matches or CSR teams, I would love to increase the impact of our fundraising drive through your corporate giving programs. Please DM me. 🌟
I am personally in touch with the highest levels of Government in Nepal to ensure that the funds we raise go straight to the OFFICIAL Disaster Relief accounts and any administrative or fund transfer fees will be personally covered by me!
Let's keep the momentum going.
I also want to shout out the founders and builders of the Nepali Tech Mafia, in and around the valley 🌁 and beyond, for the incredible work they're doing in and for Nepal:
@RashikAdhikari @RenegadeIns
@UpretyMd @missionbuildnp
@sudiprokaya @mk314k @laminalabs
@namibaral @niuralHQ
@aadityasubedi_ @architectlabs
@dileepdhakal @veel_app
@niteshpant56 @ai_evals
@AlpenLabs @simanta
@ksubedi @echowinAI
@kshitizh
Please tag anyone that I may have missed🇳🇵🇺🇸
Suvash Sedhain retweeted
to all my oomfs, a small cry of help: help make a donation and assist the vulnerable individuals of nepal. your small help means a lot.
pmdrf.nchl.com.np
This is much crazier than I imagined. Highly recommended to anyone interested in extreme capabilities of agents.🤯🤯🤯
Replying to @METR_Evals
The full report has much more information than we could convey here, including details on the projects the agents collectively pursued, the technologies they developed for communication and coordination, and interactive figures analyzing agent activity:
metr.org/blog/2026-08-26-ope…
This hits especially hard. I was born in this town, and so many of my childhood memories are tied to this place. Seeing it devastated by the floods is heartbreaking.
🇳🇵Flash flood waters from the Bhote Koshi river have torn through Trishuli Bazaar in Nuwakot, Nepal, sweeping away buildings, bridges, and sections of the riverside market.
The surge, which entered from the Tibet border area earlier today, has left extensive damage across Rasuwa and Nuwakot districts, with roads, hydropower facilities, and homes among the losses. Rescue and search operations are underway as authorities assess the full scale.
The suddenness of the event, with little local rainfall reported, underscores the vulnerability of settlements along these steep Himalayan valleys.
What stands out most to you in the footage from the area?
Seems like a new local MOE is coming out from @Alibaba_Qwen
Metrics looking pretty good for local model
Qwen 3.8-27B is finally out huggingface.co/Qwen/Qwen3.8-…
Today, we release LFM2.5-VL-3B, a lightweight vision-language model that reads screens, documents, and the physical world. It handles digital screens across mobile, web, and desktop, grounds objects to coordinates, reads text and charts, and calls tools from either text or image input.
Built on LFM2.5-2.6B base, with a SigLIP2 400M NaFlex vision encoder
> Pre-trained on ~34T tokens
> Vocab size: 128K
Comparable or better scores compared to models up to 2.6x its size:
> ScreenSpot-v2 80.7, ahead of Gemma-4-E4B at 51.2
> RealWorldQA 73.1, ahead of InternVL-3.5-4B at 67.7
> TextVQA 84.3, ahead of Qwen3.5-4B at 81.2
> RefCOCO-avg 87.9, up from 57.1 on LFM2-VL-3B
> ToolSandbox 59.5, up from 26.4 on LFM2-VL-3B
🧵
Speculative Decoding and the Art of Fast LLM Inference : mesuvash.github.io/blog/2026…
























