Nichy retweeted
Hy4 flash is now open-weight, including Hy4 flash-FP8 and Hy4 flash-Base.
224B (12B active), 1M context, natively multimodal.
Fast, efficient and built for everyday productivity.
It's yours now. And stay tuned for Hy4 coming up next!
🤗: huggingface.co/tencent/Hy4-f…
🐙: github.com/Tencent-Hunyuan/H…
📄: hy.tencent.ai/research/hy4-f…
Released under the Tencent Hy License.
See full terms here: huggingface.co/tencent/Hy4-f…
your Hermes Agent can answer as a pirate, a noir detective or Shakespeare, and go back to normal with one command.
> `/personality` lists the built-in styles
> `/personality pirate` switches this session to Captain Hermes
> `/personality teacher` when you want patient answers with clear examples
> `/personality none` clears it, from your next message
Nichy retweeted
Introducing Qwen3.8-27B-pi-adapted 🤗
We adapted Qwen's model specifically for the pi coding agent using reinforcement learning and a LoRA adapter.
The results, measured as pass@1:
🚀 SWE-bench Pro (Hard 51): 29.4% → 62.7%
💻 Terminal-Bench 2.1: 39.5% → 55.8%
⚡ 55–62% fewer tool calls per task
This is the power of open models: you can adapt them to YOUR tools and YOUR use cases. Let’s build! 💛
Open weights: huggingface.co/harness-hub/Q…
Methodology: huggingface.co/blog/harness-…
Nichy retweeted
You've installed 𝗛𝗲𝗿𝗺𝗲𝘀 𝗔𝗴𝗲𝗻𝘁. You're probably using 3 of its features.
Here's what to turn on first, ranked by hours saved. 🏆
S · turn on today
▸ Memory: stop re-explaining yourself every chat
▸ Scheduled jobs: "every morning at 8, send me the news"
▸ Helpers in parallel: 3 jobs at once, not one by one
A · set up this week
▸ Saved skills: teach it once, reuse it forever
▸ Login vault: it signs in, your password never hits the chat
▸ Telegram: message it from your phone
▸ Search old chats: "what did we decide last Tuesday?"
B · worth the setup
▸ /goal: it keeps going until the job is done
▸ Backup model: if one model fails, another takes over
▸ Browser control: it clicks through websites for you
▸ MCP tools: plug in apps you already use
C · nice, but niche
▸ Mixture of Agents, Kanban board, voice mode
D · stop doing these
▸ Pasting passwords into chat → use the login vault
▸ One giant chat for everything → /new for each task
▸ Auto-approve (YOLO) on everything → approve risky steps
▸ Skipping updates → run hermes update
Start with the S row. One a week.
Save this. Which one did I rank wrong? 👇
Nichy retweeted
I built my own desk pager for $24
It shows me:
1. agent notifications
2. track usage limits
3. texts from my gf
No phones. No pings.
Just a tiny configurable e-ink display.
Learn electronics by playing with it
1. Circuits, with the current drawn live
falstad.com/circuit
2. Arduino & ESP32 without buying a board
wokwi.com
3. Breadboards in the browser
tinkercad.com/circuits
4. Logic gates, up to your own CPU
circuitverse.org
5. A real 1975 CPU, transistor by transistor
visual6502.org
6. Build a computer from NAND gates (free course)
nand2tetris.org
7. Signals & Fourier, visually
jackschaedler.github.io/circ…
8. Ohm's law & Kirchhoff, by hand
phet.colorado.edu/en/simulat…
9. An 8-bit computer on breadboards, every wire explained
eater.net/8bit
10. Code a microcontroller in the browser
makecode.microbit.org
P.S. #5 is the actual chip inside the Apple II, the NES and the Commodore 64, photographed layer by layer and rebuilt so you can watch every transistor switch
I replaced MIT “Robotics Engineering” course with Fable 5.2 in 3-hours..
full course in one site - no need to spend $3,750+ to learn what a GPIO pin does
• 8 real boards in 3D: Arduino Uno, Raspberry Pi 4…
• hyper-realistic graphics with real-life actions
• build part by part: place chips, connectors and headers, then solder them
• wire real circuits: blink an LED, read a sensor, drive a motor
Fable 5.2 generated every 3D model in code - no ready-made 3D files or photos.
try it free, no sign-up ↓
robotics-lab-app.vercel.app
Nichy retweeted
testing d1-Omni on my vision harness side project, ran 100 receipts for parsing misc info
got dates with 98% acc, full address with 52%, company names with 85%
I'd still opt for tiny OCR models by lighton etc for now, but this is promising direction
you can see how cheap this is. and on top of that it's actually useful
an ESP32, a tiny screen, a mic, a speaker and a bulb on a breadboard. you talk, it answers, and even follow commands
how it works:
1. a small mic module listens, the ESP32 records your voice
2. the screen shows what it's doing: listening, thinking, speaking
3. the voice goes over wi-fi to an AI model, the answer comes back as speech
4. if the answer is a command, the ESP32 switches the light itself
the chip is the ears, the face, the mouth and the hands. the brain is a model it calls. all of the hardware is probably under $20
such a great and deep explanation on how AI Agents work.
if you're interested in Hermes Agent and how it works under the hood, watch this video (it's only 2 days old - can't be fresher).
youtu.be/nU2a2lm6vDE?si=VSMg…
Nichy retweeted
Hy4 flash, the new model from @TencentHunyuan, is free in Nous Portal for two weeks.
It is built to be extremely cost-effective at everyday work tasks like documents, slides, and computer use. Try it today at portal.nousresearch.com
⚡Meet Hy4 flash, the cheapest model for everyday productivity. Matches Claude Haiku 5.5 (high) on daily work (human blind tests), yet 18x task cost reduction. API pricing (per 1M tokens): $0.003 cached · $0.070 input · $0.281 output. 🔗:hy.tencent.ai/research/hy4-f…
Nichy retweeted
We shipped Chandra 2.7!
Nothing massive, just steady OCR improvement:
- Better tables: balanced 91.6 → 94.2 on OmniParseBench
- Better on non-English and rotated pages
Nichy retweeted
QVAC's RAG is 8.2x faster, and more accurate
QVAC SDK 0.24 is out, with vector search for RAG up to 8.2x faster and new image generation models.
What's new:
- TurboVec 1.1.2: RAG search in the SDK is now 2x to 8.2x faster, and more accurate
- MegaLoc: a new model that recognizes places from a photo
- Image generation: adds Flux 3 Klein, Qwen-Image-2.1 Turbo and MOSS-Image with layout control
- qvac serve: better integration with tools such as Codex from @OpenAIDevs and @t3dotcodes
- Sessions: one SDK runtime can now serve several apps at once, each with its own requests
- Pocket TTS: now runs directly through the SDK, with streaming and cancellation
- Cache sharing: conversations reuse each other's cache, so the start of a prompt they share is processed only once
- AMD Ryzen AI NPU: Qwen3.5 can now use the NPU in recent AMD Ryzen AI chips, so your GPU stays free for another model
- And several other developer experience improvements
Learn more in the thread below.
Complete release doc here: github.com/tetherto/qvac/rel…
Star ⭐️ the repo to stay updated.
Nichy retweeted
Sonntagsbeschäftigung:
In der Früh einen post von Nazeem gesehen, der ein kleines Sprachmodell mit nur 4 Millionen Parametern trainiert hat.
Meine 3090 war gerade frei, also: Warum nicht eine Variante auf Deutsch erstellen.
Ca 12 Stunden später:
Replying to @n4ze3m @Alibaba_Qwen
A note on languages: this version of tiny-voca is English only.
That's on purpose. It's an experiment with synthetic data, and I don't speak other languages, so if a voice sounded wrong I would have no way to know. So I stuck with English.
Anyone can add more languages with the open recipe. And let's see, I may add more soon ;)
someone's Hermes Agent checks his inbox every five minutes, and it costs almost nothing:
a small script runs first. if no mail is new, "the run stops there: no model call, no prompt, no notification, no bill."
two more things it does for him every day:
> a briefing before 05:00 with today's and tomorrow's events, the emails that need him, and the weather for wherever he actually is
> a language drill each morning in French, Italian or Portuguese, with audio and a slowed-down copy. "Generating the drills costs one model call a week"
he runs it on DeepSeek V4.1 Flash at medium reasoning.
more about his setup:
reddit.com/r/hermesagent/s/t…
🚀 MiniCPM-OCR 🤝 vLLM & SGLang
Native support for two leading open-source inference frameworks — @vllm_project and @sgl_project — making it easier for developers to integrate MiniCPM-OCR into their existing workflows.
⚡ Powerful OCR, flexible deployment, and a growing open-source ecosystem.
🤖 Made with AI
Your OCR says it’s done—but did it get everything right? 🤔 In high-stakes workflows, even a small error can be costly.
🚀 Meet the MiniCPM-OCR suite: an open-source solution to parse fast and guard the results.
Two new models:
🔹 MiniCPM-OCR 1.0 — SOTA OCR performance and speed at 1B scale. Compact, accurate, and fast.
🔹 MiniCPM-OCR Jev — A 2B decision model that helps assess parsing accuracy, with performance approaching that of Qwen3.5-397B-A17B on JevOCRBench.
⚡ What’s new?
Small model, SOTA results: MiniCPM-OCR 1.0 scores 93.68 on Real5-OmniDocBench and 54.7 on OCRBench v2.
A sharp goalkeeper 🛡️: MiniCPM-OCR Jev is model-agnostic, checking text, formulas, and tables from different OCR parsers. It achieves 93.6 on JevOCRBench, outperforming JEV-35B-A3B (90.1) and approaching Qwen3.5-397B-A17B (93.8).
Up to 128 pages per minute 🤯: With DSpark, MiniCPM-OCR runs at 1.49× the speed of PaddleOCR-VL-1.6 and 3.62× that of MinerU2.5 Pro. The SDK’s concurrency optimizations further boost throughput to as much as 128 pages per minute.
Deploy where you work: Day-0 support for Arm, Houmo AI, Intel, and NVIDIA, plus 10 more chip vendors through FlagOS.
Ready for your stack: vLLM, SGLang, llama.cpp, and Ollama, with quantized models, fine-tuning support, and Agent Skills.
🔗 GitHub: github.com/OpenBMB/MiniCPM-O…
🤗 Hugging Face:
MiniCPM-OCR 1.0:huggingface.co/openbmb/MiniC…
MiniCPM-OCR Jev:huggingface.co/openbmb/MiniC…
👉 Try the Demo:platform.modelbest.ai/consol…
Use the API:platform.modelbest.ai/consol…





































