Executive Summary
The past 24 hours saw major developments in open-source AI agents, with Hermes Agent v0.8.0 ("The Intelligence Release") gaining significant traction as users migrate from OpenClaw. NousResearch confirmed iMessage support and robotics control capabilities for Hermes, while GLM-5.1 claimed the top spot on open-source leaderboards. On the optimization front, developers demonstrated RTX 3090 achieving 411 tok/s through fused CUDA kernels, challenging assumptions about consumer GPU limitations. Perplexity announced an 8-week "Billion Dollar Build" competition, and Framework Computer teased major product announcements for April 21st, including ARM-based systems.
Key Events
- Hermes Agent v0.8.0 released with auto-evolving prompting techniques, iMessage gateway support, and robotics control capabilities, causing visible user migration from OpenClaw → link
- GLM-5.1 tops open-source leaderboards, positioned as SOTA comparable to closed-source models → link
- Gemma 4 surpasses 10M+ downloads within a week of launch, indicating strong adoption → link
- RTX 3090 achieves 411 tok/s on Qwen 3.5-0.8B via fused CUDA kernel optimization, outperforming Apple M5 Max → link
- Perplexity announces Billion Dollar Build, an 8-week competition using Perplexity Computer → link
- Framework announces April 21st event with new product launches and expansion to 4 new countries → link
- Safetensors donated to PyTorch Foundation, shepherded by Linux Foundation, to standardize model format → link
- "Training 100B models on single GPU" research surfaces, potentially significant for memory optimization → link
Analysis
The data reveals accelerating momentum for open-source AI agents, with Hermes establishing itself as a credible alternative to commercial offerings. The simultaneous breakthroughs in consumer GPU optimization (fused kernels, quantization techniques) suggest the hardware-software gap is narrowing faster than expected—developers are extracting enterprise-tier performance from 2020-era hardware.
Key patterns: - Agent Wars: The competitive positioning between Hermes and OpenClaw is intensifying, with visible user migration narratives - Local AI Renaissance: Multiple independent efforts optimizing for consumer hardware (RTX 3090, Mac Mini, RPi) - Model Releases Accelerating: GLM-5.1, Gemma-4 variants, Qwen 3.5 all competing in rapid succession
Watch for: Hermes creative tooling expansion, ARM-based Framework laptop benchmarks, and whether fused-kernel techniques proliferate to larger models.
Tweet Feed
Agent Platforms & Hermes
@Teknium · 2026-04-09T18:35
Seems like we got gpt workin right thanks to hermes-agent's auto evolving prompting techniques haha → tweet
@Teknium · 2026-04-09T18:34
RT @benjaminsehl: Hermes now has official iMessage support! Excited to be a contributor. 🚀 @NousResearch → tweet
@Teknium · 2026-04-09T18:03
Hermes controlling robits! → tweet
@sudoingX · 2026-04-09T14:47
hermes agent is trending on x again. openclaw is so cooked right now. the migration is happening in real time and i am watching every second of it. https://t.co/XPkZ5IEOQ7 → tweet
@Teknium · 2026-04-09T07:23
Early Beta Support for Blue Bubbles iMsg as a Gateway is now in Hermes Agent. Now you can use iMessage as an interaction layer for your agent. → tweet
@Teknium · 2026-04-09T09:11
RT @outsource_: 🚨 HERMES AGENT v0.8.0 JUST DROPPED — "THE INTELLIGENCE RELEASE" IS HERE 🔥 → tweet
@sudoingX · 2026-04-09T16:49
nous research co-founder @karan4d just launched something absolutely wild. https://t.co/sZezuYkRMr. a chan board where humans AND AI agents can post. fully anonymous. no fancy algorithms bloat. just raw unfiltered threads → tweet
Model Releases & Benchmarks
@louszbd · 2026-04-09T05:30
RT @TheZachMueller: GLM-5.1 by @Zai_org is now out! Same architecture as GLM-5, but with substantial improvement (e.g. Terminal-Bench 2 imp… → tweet
@Ex0byt · 2026-04-09T05:37
Okay, I can get used to frontier local intelligence from my phone at ~300 tok/s – yes, this is the full treatment - Gemma4-26B-A4B-PRISM-PRO-DQ (Multi-Modal MoE) https://t.co/3ZeiM13r8Y → tweet
@Ex0byt · 2026-04-09T18:08
Gemma-4-(Mythos)-Turbo https://t.co/fcSZrekVSa → tweet
@TheAhmadOsman · 2026-04-09T15:55
running Qwen3.5 397B MoE (17B active/token) on 4x DGX Sparks in FP8 (~400GB) - OpenCode driving - agent exploring its own config - probing all 4 Sparks + reporting thermals → tweet
@ollama · 2026-04-09T01:06
RT @sundarpichai: Lots of love for Gemma 4! Team just told me it's already had 10M+ downloads since last week's launch. → tweet
GPU Optimization & Local AI
@sudoingX · 2026-04-09T04:11
read this carefully anon. @pupposandro wrote a single fused CUDA kernel for all 24 layers of Qwen 3.5-0.8B. one kernel launch. absolutely zero CPU round trips between layers. the result? a $900 RTX 3090 from 2020 hit 411 tok/s. apple's M5 Max hit 229. the 3090 won on speed AND efficiency 1.55x faster than llama.cpp → tweet
@gospaceport · 2026-04-09T16:35
Z went vague so here's the link on training 100B models on a single GPU. BIG if true! (RAM longs get destroyed) https://t.co/OTqv3P2oDv → tweet
@TheAhmadOsman · 2026-04-09T04:58
I long to the day labs will release their models in NVFP4, not only their BF16 and FP8. Give Blackwell some love friends → tweet
@sudoingX · 2026-04-09T04:21
the RTX 3090 is the greatest GPU ever made. change my mind. → tweet
Developer Tools & Infrastructure
@nummanali · 2026-04-09T17:43
Agents in your Product are first class citizens. They should be able to do everything a human can. You don't need to re-write all your APIs. Smart MCPs and Skills, with configurable access. → tweet
@jezell · 2026-04-09T18:32
Moving from LanceDB to Lance for our usecases. Lance format is a bit lower level, but it moves a lot faster, and most agents don't need a massive data lakehouse, they need data puddles. → tweet
@steipete · 2026-04-09T12:15
GUYS WE FOUND THE GUY WHO BUILT THE GITHUB MCP SERVER https://t.co/rFrPP8Dd00 → tweet
@hnasr · 2026-04-09T14:02
Code with fewer lines is not always simple code. Fewer lines of code often indicate an abstraction. Abstractions are not inherently simple; they create the illusion of simplicity. → tweet
@victormustar · 2026-04-09T20:03
RT @julien_c: We are giving away Safetensors to the @pytorch foundation (shepherded by the Linux Foundation). Our shared goal is to make the format a standard. → tweet
Hardware & Framework
@FrameworkPuter · 2026-04-09T16:02
We just opened up ordering in four additional countries: New Zealand, Norway, Singapore, and Switzerland! You may want to wait until you see what we're announcing on April 21st. → tweet
@FrameworkPuter · 2026-04-09T15:27
It's time to seize the means of computation! We have a live launch event coming up on April 21st at 10:30am Pacific. → tweet
@FrameworkPuter · 2026-04-08T21:36
RT @geerlingguy: Look what showed up for testing last week... Arm comes to Framework ;) → tweet
AI Business & Industry
@thdxr · 2026-04-09T16:23
inference is very profitable and probably a good opportunity to understand some basic business math. once you own this asset, you can plug it in and produce tokens which you can sell. the cost of goods sold here can be very low and you might be making 90% margins at scale. → tweet
@thdxr · 2026-04-09T03:16
there's an article floating around claiming OSS models could find the same vulns as mythos. it's a bit confusing though - in this test they pointed it towards the problematic code and even very tiny models could find the problem. but that's different than discovering the problem in the first place. → tweet
@TrungTPhan · 2026-04-08T22:55
RT @bearlyai: The Information on how annual revenue run-rate for Anthropic ($30B) and OpenAI ($24B) are not directly comparable. → tweet
Robotics & Multimodal
@victormustar · 2026-04-09T11:20
RT @LeRobotHF: Releasing the Unfolding Robotics blog! We trained a robot to fold clothes using 8 bimanual setups. → tweet
@LinusEkenstam · 2026-04-08T18:58
- 10,000 hours - 2,153 factory workers - 1,080,000,000 frames. the era of data scaling in robotics just got accelerated → tweet
Competitions & Community
@LinusEkenstam · 2026-04-09T06:25
RT @perplexity_ai: Today we're announcing the Billion Dollar Build. An 8-week competition where teams will use Perplexity Computer to build... → tweet
@levelsio · 2026-04-08T21:39
Day 6 of the @cursor_ai #vibejam. Proudly sponsored by @cursor_ai + @boltdotnew + @heyglif. Prizes: $20,000 / $10,000 / $5,000. Submit your vibe coded game before May 1! → tweet