( ✦﹏✦）

idea collector

[SiliconFlow @SiliconFlowAI]

🎁 Free Access for First 2 Weeks
Meet @TencentHunyuan Hy3
Now with T+0 support on SiliconFlow 🎉

Hy3: 295B MoE / 21B active / 262K context
✅ Refined iteratively through 50+ real business scenarios
✅ Cuts hallucinations & knowledge errors in half
✅ Stays on-intent over long-horizon tasks
✅ Stable agent execution with more tool-call wins, no infinite loops

Put it to real work through SiliconFlow ⬇️

[0xMarioNawfal @RoundtableSpace]

Zero is a terminal coding agent you actually own.

* Bring any model you want - OpenAI, Anthropic, Gemini, DeepSeek, Ollama, 25+ providers
* File writes, shell commands, and network access all go through its own permission and sandbox layer
* Sessions stored locally, resumable, never uploaded as telemetry

Most coding agents pick your model for you, keep your data, and quietly call home.

Zero flips the default: you choose the model, you set the permissions, you keep the context.

Github: https://t.co/zdi6Yl2YCj

[Victor M @victormustar]

bottlecapai/ThinkingCap-Qwen3.6-27B (definitely trying this one)

they say it matches Qwen3.6-27B -> BUT with 50% less thinking tokens on average, and over 90% less in best cases 👀👀

[fal @fal]

⚡️ Ideogram V4 Fast and Instant are now available on fal

-The fastest Ideogram V4 variants, built in-house by fal
-Full images in 0.4 seconds on Instant and one second on Fast
-Same prompts and parameters as the original at up to 8x the speed

[Victor M @victormustar]

@wildmindai looks great - btw idk if you have seen https://t.co/o3qBg1e5gG

[Jun Song @jun_song]

Finally an open weight model that can replace Deepseek V4 Flash.

Testing it today.

Still need to check the license because they banned Korea and EU for their last world gen model.

[Teortaxes▶️ (DeepSeek 推特🐋铁粉 2023 – ∞) @teortaxesTex]

profoundly disgusting.

[0xSero @0xSero]

zcode is an excellent desktop app for local models, the best in the game outside of local studio.

It's free too! I setup my local provider in settings and I'm good to go.

So easy

[ModelScope @ModelScope2022]

🎨 Boogu-Image-0.1-Edit-Turbo is now on ModelScope, a 4-step distilled image-to-image model for fast visual editing. ⚡️Apache-2.0
🔗 https://t.co/8jRQQE0SE6

✏️ Edit tasks: object changes, style transfer, scene/background edits, and text-aware visual transformation
🧩 Boogu-Image-0.1 family: Base, Turbo, Edit, and Edit-Turbo cover text-to-image, fast generation, and image editing workflows

[HARUKI｜AI美女アーティスト @ai_haruki3]

CIVITAIに公開しました。（CIVITAIで公開できるのわすれてたよね）

https://t.co/Y1DhEv3PU0

[Georgia Channing @cgeorgiaw]

Seems like no one's noticed the 80TB of astrophysics data from 30+ sources that just dropped on @huggingface.

...and you only need ~4GB of RAM to load it.

We're talking over 80TB of galaxy imagery taken across the spectrum, spectra of galaxies and stars, time series of variable stars, and a whole zoo of assorted measurements and physical data.

And all of it can now be wrangled on your laptop, thanks to Multimodal Universe's just released cross-matching. SDSS x Gaia means you can match 800k objects against 122M objects, and it never climbs above ~4GB of RAM.

Huge congrats to @smith42mike for leading this and making the world of astro accessible to probably 10,000x more people. Let's discover some shit

[Wildminder @wildmindai]

ASASR is now open-sourced.

performs x4 image SR with a FLUX.1-dev and dual-LoRA inference: upscaling prior, OminiControl + DPO LoRA

https://t.co/K5JFyXQgpr

[ModelScope @ModelScope2022]

Meet SenseNova-U1-8B-MoT-Infographic-V2! This model is designed exactly for dense content, complex layouts, text-heavy visuals.
🤖 https://t.co/ytyUq5j2aH

🏆 Open-source SOTA performance
Strong results on infographic benchmarks, including BizGenEval and IGenBench.

📊 Improves chart accuracy, text rendering, size appropriateness, layout understanding, and arXiv-style page quality, while reducing unintended black backgrounds.

🧠 Native unified multimodal architecture
Built on SenseNova-U1, unifying visual understanding, reasoning, and generation in one model.

[Zephyr @zephyr_z9]

Rather than routing data thru PCB, they are using lots of dedicated flyover cables to achieve ultra fast memory access
Manufacturing this in volume will be an extremely tough job

[Paul Couvert @itsPaulAi]

Best local models you can run on a laptop today

(And without a $10k Mac Studio)

1. Qwen3.6-27B: by far the best for coding agents
or Qwen3.6-35B-A3B for a faster option

2. Gemma 4 12B: everyday tasks and questions

3. Parakeet 0.6B v3: basically the best voice to text model you can find

4. Gemma 4 E4B: surprisingly good for the size and can run offline on a phone

5. Gemma 4 26B diffusion: the most tok/s you can have with a local model

You should use a quantized version from Unsloth when you're running them on a laptop: the precision/size/speed are fantastic.

LM Studio or the 'raw' llama.cpp are your best options.

Or Google AI Edge Gallery on ios/android.

[Nous Research @NousResearch]

Hermes Agent now reads the web up to 60x faster and 49x cheaper.

Scraping backends pass clean content straight to the agent without redundant processing steps; large pages are saved locally and paged on demand so you get the same quality at a fraction of the time and cost.

[DailyPapers @HuggingPapers]

Microsoft just released a new GUI agent on Hugging Face

Sico-Evolution jumps from 39.8% to 82.9% Task Success Rate

Outperforming GPT-5.4, Claude Opus 4.6, and Claude Opus 4.7

All from a 4B parameter model

[Wildminder @wildmindai]

KVAE-Audio by Kandinsky Lab - continuous 48kHz audio tokenizer for SOTA text-to-audio.
- 166.9M params
- full spectrum
- mimics natural oscillations (no "robot Voice")
- tops MMAudio in T2A quality
-only 64 latent channels
- MIT

https://t.co/5riEIoEZiY

[JP (Zenoware) @Zenoware]

I've published a guide on how to make nearly ANY OpenAI-compatible provider + model combination work with Codex using the very popular codex-shim tool (translates Chat Completions API requests to Responses API requests).

Use DeepSeek, MiniMax, GLM, Kimi, Mitsubishi, Panasonic, Honda, Sanryo, etc. models on the provider of your choice without having to proxy your calls thru OpenRouter.

Hope it helps some people out.

https://t.co/ZAVSFlD04K

[Will @hampsonw]

I think i understand why @cursor_ai doesn't release the performance of Composer 2.5 on the common public benches. At least in the case of @datacurve deepswe, it's because they train on it. it's a good model but nothing ACES it with ~125x fewer tokens than GPT-5.5 - medium without having learned the test.