Executive Summary
The past 24 hours saw a significant breakthrough in open-source AI with the release of Kimi K2.6, claiming state-of-the-art performance on multiple coding benchmarks and outperforming GPT-5.4 Extra High and Claude Opus 4.6 Max. The Hermes Agent ecosystem from Nous Research reached a major milestone of 100K GitHub stars, overtaking OpenClaw in weekly stars and demonstrating rapid community growth. Developer tooling continued advancing with Codex's new "Chronicle" memory feature and native Ollama integration with Copilot CLI. Meanwhile, the trend toward local AI inference on consumer hardware is accelerating, with users reporting successful autonomous agent operation on laptops with 24GB VRAM.
Key Events
-
Kimi K2.6 achieves open-source SOTA on coding benchmarks (HLE 54.0, SWE-Bench Pro 58.6), beating proprietary frontier models. → tweet
-
Hermes Agent hits 100K GitHub stars, overtaking OpenClaw in weekly new stars two months after launch. → tweet
-
Codex introduces "Chronicle" experimental feature giving AI persistent memory of user's work context. → tweet
-
Ollama adds native Copilot CLI support, enabling offline AI workflows with custom models. → tweet
-
MCPorter v0.9.0 released, adding per-server tool filtering and Windows OAuth fixes for MCP infrastructure. → tweet
-
Gemma 4 testing begins on consumer hardware, with comparisons against Qwen 3.5 27B for autonomous agent tasks. → tweet
Analysis
Open-Source AI Acceleration: The gap between open-source and closed frontier models continues to narrow. Kimi K2.6's release demonstrates that open-weight models can now match proprietary systems on agentic coding tasks. The ~2 month gap mentioned by analysts suggests open-source is approaching parity for practical developer use cases.
Agent Ecosystem Wars: Hermes Agent's rapid 100K star milestone and native Ollama support indicate strong community preference for self-improving, locally-run agents over cloud-dependent alternatives. The competition between OpenClaw and Hermes is driving rapid feature development and ecosystem expansion.
Local AI Inference Maturation: Reports of sustained 15 tok/s on 24GB VRAM laptops running 31B parameter models mark a significant democratization milestone. Combined with mentions of Gemma 4 31B and Qwen 3.5 27B achieving performance previously requiring much larger models, local AI deployment is becoming viable for professional workflows.
Security Concerns Rising: Multiple posts highlight Chrome extension malware risks, reinforcing the trend toward vibe-coding personal tools and verifying open-source dependencies—areas where AI coding assistants can help but also introduce new trust considerations.
What to Watch: Kimi K2.6 real-world agentic performance vs benchmark claims; Hermes Agent ecosystem growth trajectory; Gemma 4 release timing and capability leap; potential Codex service disruptions (reported down twice); DeepSeek V4 launch rumored in coming weeks.
Tweet Feed
AI Model Releases & Benchmarks
@sudoingX · 2026-04-20T17:49
holy shit. kimi just shipped K2.6 and the open source chart flipped the frontier. SOTA on humanity's last exam (54.0), deepsearchqa (92.5), swe-bench pro (58.6). beating gpt-5.4 xhigh, claude opus 4.6 max effort, gemini 3.1 pro thinking high on multiple benchmarks. → tweet link
@crystalsssup · 2026-04-20T15:30
Kimi K2.6, let's goooooooo🥳 Open-source SOTA on HLE w/ tools (54.0), SWE-Bench Pro (58.6), SWE-bench Multilingual (76.7), BrowseComp (83.2), Toolathlon (50.0), Charxiv w/ python(86.7), Math Vision w/ python (93.2) → tweet link
@Teknium · 2026-04-20T18:21
Kimi 2.6 looks like it could be the new contender for best open agentic model for Hermes Agent, stacking up strongly against even opus 4.6 in these agentic benchmarks 😲 → tweet link
@victormustar · 2026-04-20T15:37
RT @Kimi_Moonshot: Meet Kimi K2.6: Advancing Open-Source Coding 🔹Open-source SOTA on HLE w/ tools (54.0), SWE-Bench Pro (58.6), SWE-bench… → tweet link
@victormustar · 2026-04-20T15:29
moonshotai/Kimi-K2.6 · Hugging Face https://t.co/ySL9cmrJiz → tweet link
@nummanali · 2026-04-20T15:31
Kudos to them for benching against GPT 5.4 Extra High and Opus 4.6 Max. So we're like 2 months gap between closed source and open source now? → tweet link
@sudoingX · 2026-04-20T15:48
before i unleash gemma 4 on octopus invaders tonight i will run 3 smaller agentic tests to calibrate it. gemma will build gpu marketplace ui, small visuals and tools. this is gemma 4 q4 on 24gb vram of 5090 mobile, sustained 15 tok/s on hermes agent as autonomous agent. → tweet link
@sudoingX · 2026-04-20T09:32
i am not being able to recover from what grok 4.3 is doing. been pushing it through autonomous agent work overnight and its operating at a level other grok versions was not capable of. the way it handles multi step reasoning, the way it doesnt bail halfway, different model energy entirely. → tweet link
@sudoingX · 2026-04-20T17:22
after testing nemotron from nvidia, gemma from google, hermes from nous research, the future of open source is more flourished than people realize. → tweet link
Hermes Agent Ecosystem
@Teknium · 2026-04-20T18:22
RT @RoundtableSpace: HERMES AGENT IS NOW GAINING MORE INTEREST THAN OPENCLAW > open source > self-improving > runs local > remembers ever… → tweet link
@Teknium · 2026-04-20T07:19
RT @Delphi_Digital: Hermes Agent recently overtook OpenClaw in weekly new GitHub stars. Two months after launch, the agent from @NousResearch… → tweet link
@Teknium · 2026-04-20T03:18
RT @minchoi: Holy smokes... Hermes Agent just hit 100K GitHub stars. It's an AI Agent like OpenClaw but with real self-improving skills an… → tweet link
@Teknium · 2026-04-20T05:12
RT @NFTCPS: 我靠！Ollama 现在原生支持 Hermes Agent 了！一行命令直接起飞：ollama launch hermes → tweet link
@sudoingX · 2026-04-20T16:53
this is a laptop running a 31b parameter model at 99% gpu autonomously through hermes agent, 15 tok/s sustained, 22.8 of 24gb vram gone, 94 watts at 50c. no api keys. no rate limits. no monthly subscription. → tweet link
@Teknium · 2026-04-20T10:27
RT @0xJeff: https://t.co/vujtA3V8Wi → tweet link
@steipete · 2026-04-20T16:08
Kudos to the folks from Tencent for working with us and providing evals to improve OpenClaw's harness performance! We're also working with them to bring fixes/improvements back to the open source repo. → tweet link
Developer Tools & IDEs
@gdb · 2026-04-20T18:24
Chronicle is an experimental feature giving Codex the ability to see and have recent memory over what you see, automatically giving it full context on what you're doing. Feels surprisingly magical to use. → tweet link
@gdb · 2026-04-19T21:16
codex is becoming the universal app for developers: → tweet link
@nummanali · 2026-04-20T15:54
Codex is very down https://t.co/LXLSuP3qmS → tweet link
@nummanali · 2026-04-20T14:48
Oh no Codex is down https://t.co/KWsi3i7Yj2 → tweet link
@steipete · 2026-04-20T11:42
MCPorter 🧳 0.9.0 is out. Call MCPs from TypeScript or as CLI - per-server tool filtering - sturdier stdio shutdowns - Windows OAuth URL quoting fix - OAuth config docs - schema-declared string coercion for tool calls → tweet link
@ollama · 2026-04-19T19:02
RT @_Evan_Boyle: Ollama now natively supports Copilot CLI! Bring your own models, and even work completely offline! → tweet link
@ollama · 2026-04-20T03:55
RT @osanseviero: What a week in SF for Gemma! 💎 - Gemma 4 SF Meetup, with Unsloth, MLX, VLLM, NousResearch, and Cactus Compute - Hackathon… → tweet link
@nummanali · 2026-04-19T20:29
Codex <---> Claude Code comms with cmux. New way of working is letting Codex and Claude communicate between themselves while I co-ordinate as needed. → tweet link
@badlogicgames · 2026-04-19T23:18
pi tip: in capable terminals like ghostty, use CMD + up arrow key to jump to the top of final assistant messages in a turn, and user messages. → tweet link
@TheAhmadOsman · 2026-04-19T23:30
There's so much alpha in opening Codex Cli and telling it to teach you how to do X. X can be shell basics, optimizing local ai inference, hacking your pet feeder camera, setting up a self-hosted search engine → tweet link
@badlogicgames · 2026-04-19T21:47
i'm now convinced they switched opus 4.6 to be haiku. only half joking. → tweet link
Vibe Coding & Indie Development
@levelsio · 2026-04-20T16:56
Now just 1.5 hours into this mini project. I have replaced ALL my Chrome extensions with my own vibe coded ones for safety. My own Chrome extension is called 🚀 SuperLevels and now has: Tab Cleaner, Cookie Editor, Redirect Tracer, Dark Mode, X Dim, Music Finder, and more. → tweet link
@levelsio · 2026-04-20T15:24
Okay that was WAY too easy 😂 I created 🗑️ Tab Cleaner: My own Chrome extension that has no malware because I coded it myself and I can read the source code. → tweet link
@levelsio · 2026-04-20T15:12
Another Chrome extension allegedly sold to mine your personal information. Your best bet is to use no Chrome extensions at all, except uBlock Origin for ad block. Or better: just vibe code your own so you know what is running in your browser. → tweet link
@levelsio · 2026-04-19T23:15
🧑🚀 Day 17 of the @cursor_ai #vibejam. Prizes to win (submit your vibe coded game before May 1!) $25,000 / $10,000 / $5,000. YOU HAVE 11 DAYS LEFT! → tweet link
@levelsio · 2026-04-20T14:57
💾 I renamed it to https://t.co/6tRObkZbZa because AltaVista is still owned by Yahoo so I didn't wanna infringe. 🔎 is a fully functional search engine and proxy that lets ANY old browser browse most of the web with just HTTP. → tweet link
@crystalsssup · 2026-04-20T16:17
One-shotted WebGL website by Kimi K2.6 😭 https://t.co/9ju1JOSnki → tweet link
Open Source & Research
@victormustar · 2026-04-20T13:42
RT @LysandreJik: We're opening a Hugging Face office in Tokyo! Our goal: help open-source AI develop in Japan and grow the local community… → tweet link
@victormustar · 2026-04-20T13:38
RT @AdinaYakup: MOSS-VL 🔥 Vision model from @Open_MOSS Model: https://t.co/bKopYaALra Demo: https://t.co/ulx8TiqmQX ✨ 11B - Apache 2.0 → tweet link
@victormustar · 2026-04-19T20:17
RT @CardilloSamuel: releasing qwen3.6 35b a3b carnice edition (inspired by @kaiostephens) to enjoy an hermes agent tailored finetuned on un… → tweet link
@victormustar · 2026-04-19T20:12
RT @0xSero: I interviewed Huggingface's head of product Victor. Huggingface is AI's core. I think this is the most important platform if… → tweet link
@jsuarez · 2026-04-20T14:45
Reinforcement Learning dev with Joseph Suarez https://t.co/c20OeZZ7Ut → tweet link
@jsuarez · 2026-04-20T14:36
Can't cast fireball yet, but it's pretty good. Narrow intelligence you can quickly train on a ton of tasks for pennies is general intelligence → tweet link
@nummanali · 2026-04-20T15:52
The best post on RLMs that I have read. It covers the full history of CoT and Tool use and how RLMs finally aim to collapse the disconnect. → tweet link
@tinygrad · 2026-04-20T02:15
When we succeed at making the most performant full stack from Tensor to hardware for both major GPUs in 25k lines, a lot of people are going to have to ask what the other 10M lines were doing. → tweet link
AI Agents & Automation
@kunchenguid · 2026-04-20T18:13
alright - finally ready to share another secret weapon in my agentic engineering setup. git push no-mistakes. that's not a joke - it's the real command i run when i push my changes nowadays, and it's what allows me to ship 50-100 good PRs a day → tweet link
@TrungTPhan · 2026-04-20T17:40
RT @bearlyai: CEO Rene Haas on why AI agents are spiking CPU demand: ▫️use >15x more tokens than humans ▫️a 1GW data centres now need 4x C… → tweet link
@carlvellotti · 2026-04-20T15:33
There's a generational career opportunity RIGHT NOW. Become your team's AI https://t.co/sxBtZXD5n2 → tweet link
@sudoingX · 2026-04-20T09:37
i keep getting DMs asking what to run on 8gb and 16gb vram. lets just solve this for everyone. what you running, or want me to dig and come back with a worth it list for your tier ? → tweet link
@sudoingX · 2026-04-20T11:01
everyone chasing new model releases when half of you havent maxed what you already have. run what you got harder before you upgrade → tweet link
@sudoingX · 2026-04-20T17:05
2 years ago this laptop couldnt run a 31b model. in 3 years every dev machine runs 200b locally. in 5 years your phone will. novel architectures already cooking, we are at the acceleration point. → tweet link
Industry & Enterprise
@gdb · 2026-04-20T16:59
Hyatt is deploying AI broadly to its employees: → tweet link
@jezell · 2026-04-20T17:09
Companies like Google / Amazon likely struggle to use LLMs with their own codebases more than you do because their own codebases aren't in the training set and they don't just throw together vibe coded wrappers on top of open source projects together that are in the training set. → tweet link
@RydMike · 2026-04-20T00:09
What has happened with @OpenAI #ChatGPT 5.4, feels like it has gotten a lot smarter at code. Used to be the diff between 5.4 and Codex 5.3 was marginal. Maybe 5.4 was a bit nicer to talk with, but now, man 5.4 is insanely good and keeps nailing stuff no model has done before. → tweet link
@steipete · 2026-04-20T12:08
Interesting shift. These highly subsidized subs are out there to get your code to improve their models. If you use AI for things useful to you, but not code, you are not valuable to them. → tweet link
@jezell · 2026-04-20T16:59
There's a misalignment between how APIs currently work... People just throw crap in Cursor / Codex / Claude and let it call the APIs. Probably low request rate APIs need LLM powered guardrails on the server side. → tweet link
@jezell · 2026-04-20T17:13
So who's gonna build a public CI/CD platform on top of Temporal? → tweet link
@TrungTPhan · 2026-04-20T03:34
Fencing at LA 2028 Olympics looking to have this cool sword visualization technology... trained a deep neural net system for object detection... built a miniature fencing set for evaluating 3D sword tip estimation algorithm. → tweet link
@TheAhmadOsman · 2026-04-20T05:59
LLM Inference Engine Stack Breakdown and Workload/Bottlenecks Cheatsheet. From the upcoming Inference Engine Comprehensive Article I am writing. → tweet link
@badlogicgames · 2026-04-20T12:03
RT @SIGKITTEN: i cant even get my kids to give 2 shits about ai no matter what i do. most ive gotten is them asking (new) al… → tweet link
@badlogicgames · 2026-04-20T11:01
hi, my name is mario, i'm building coding agents and other LLM bullshittery, and i think the parent is directionally right. kids + AI is a spectrum. handing them a sycophantic LLM unsupervised and uninstructed is on the "terrible fucking idea" end of that spectrum. → tweet link