Orhan Yılmaz

developer ¦ cto @paycell

[Bilgin Ibryam @bibryam]

19 latency patterns for faster applications

The model is only one segment of the critical path. The rest of the system still has distance, repeated work, serial waits, and cold setup.

A practical map for reducing non-model latency:
https://t.co/bTXW6yXd5L

[Youssof Al Toukhi @Youssofal_]

20 hours later and I’m done!

Qwen 3.8 27B @ 73 TPS (peak) on a Macbook pro M5 max.

MTPLX V2.7 out now with 3 new models.

- Bare Speed: short burst decode, good for chat.

- Optimized Speed: higher quality and faster on long coding.

- Optimized Quality: 30% slower but super!

[Stagehand @Stagehanddev]

Introducing Stagehand v4: the SDK for browser agents.

Playwright was built for testing, we built Stagehand for your agent: with improved context management, self-healing actions, and iframe support.

[Cactus Compute @cactuscompute]

We release Needle 2: A 14MB agentic LLM for phones, wearables, smart home, robots and microcontroller. The whole model is a single 14MB binary that runs a full session in 28MB of RAM. It is built on our Simple Attention Network findings, compressed to CQ2-bit with Cactus Quants, and baked into its own engine.

Needle 2 has 45m parameters trained from the ground up on 140B tool call, device use structured generation tokens. On mobile device use benchmarks, Needle 2 trades wins with frontier small LLMs like LFM2.5 230M, Apple FM Gemma-270m, at 5× to 70× smaller, and 2 bits against their f16.

Needle hits 500 tokens/sec decode speed on a Raspberry Pi 5, between 400–1,500 tokens/sec on VR devices like Meta Quest 3S and Apple Vision Pro, and ranges 300–700 on sub-$200 phones such as the Samsung A-Series. Needle also runs on newer microcontrollers like ESP32.

A conventional transformer of Needle's width and depth spends 164 MFLOPs per token, and even one squeezed down to Needle's parameter count spends 87, Needle spends 70. Even on a high-end phone, an always-on assistant lives inside a power budget; every MFLOP is milliwatt-hours, and Needle spends 7x to 85x fewer of them per token than the smallest performant LLMs.

Read more: https://t.co/dLK2tIKXQu

[Markus Eisele @myfear]

Quarkus Multitenancy Without Tenant Plumbing in Every Service

What the Quarkiverse Multitenancy extension gives you: one resolution point, clean domain code, and three tested resolution modes for database-per-tenant isolation.

https://t.co/TsOb6aCgSZ

#java #quarkus

[Jun Kim @jundotkim]

oMLX 0.4.4rc2 is out with early MiniMax M3 support, made possible by the awesome mlx-vlm work from @Prince_Canuma and @ivanfioravanti, tracking the upstream mlx-vlm PR. Stable 0.4.4 is planned after a short final RC test pass!

MiniMax M3 is supported with oMLX features including SSD cache, prefix cache, continuous batching, and the OpenAI-compatible API.

This release also adds stronger macOS 27 compatibility, safer native MTP batching, more robust Gemma 4 / Harmony tool-call handling, and additional cache / Memory Guard hardening.

MiniMax M3 single-request results (M3U 512G, ssd-cache on)
pp1024 332.6 tok/s, tg128 28.9 tok/s
pp4096 359.8 tok/s, tg128 20.8 tok/s
pp8192 340.2 tok/s, tg128 20.2 tok/s
pp32768 243.5 tok/s, tg128 18.9 tok/s

Continuous batching at pp1024/tg128:
1x 28.9 tok/s
2x 40.1 tok/s
4x 51.3 tok/s
8x 57.8 tok/s
https://t.co/cIvWwzZ9SS

[Jake @JakeKAllDay]

a very exciting release: qwopus 3.6 27b + MTP is my daily driver on a 3090 ti (3 K L performs well with 98K+ ctx). Qwopus is the highest performance FT of 27b, expecting the code FT to be 🔥

[Jun Kim @jundotkim]

oMLX v0.4.3 is now available with macOS 27 compatibility!
https://t.co/cIvWwzZ9SS

This release also fixes a Memory Guard performance regression affecting Qwopus Series, DFlash Gemma, and all other models, where generation/decode could slow significantly while Memory Guard was enabled.
In a Qwen3.6-35B-A3B cache-on single-run check, tg512 improved from 77.5 -> 79.0 tok/s (+1.9%) compared with 0.4.2. (M3U 512GB)

It also improves per-model MTP eligibility handling, and DFlash/Qwen compatibility.

[InfoQ @InfoQ]

Java devs: you might be missing a key tool in your toolkit - #DataFrames.

Why consider them?
✅ Transform, organize, and analyze data directly in Java
✅ Improve efficiency, flexibility, readability, and developer productivity

🔗 https://t.co/x2zyk33N3I

#Java

[露西的百宝箱 @Lucy_love_AI]

10 个本不该免费存在的 GitHub 仓库。

它们正在一点点掏空价值 500 亿美元的商业软件市场。

保存它

1. yt-dlp
从YouTube、X、TikTok、Instagram等任何地方下载视频。YouTube Premium每月收费14美元但功能不如它。100%免费。
仓库 → https://t.co/VcuOdGkEJf

2. Ollama
在笔记本上运行GPT-4级AI。无API成本。开发者每月花500美元用OpenAI，而Ollama离线运行免费。
仓库 → https://t.co/SjLv5kHi79

3. Fooocus
Midjourney级的图像生成。Midjourney每月30美元。Fooocus免费无限生成。

仓库 → https://t.co/JIHuXGYGV1

4. Whisper
OpenAI的转录模型开源版。Otter每月20美元，Whisper免费支持99种语言。

仓库 → https://t.co/XCel1qj9ud

5. Plausible Analytics
隐私优先的谷歌分析替代品。Google Analytics 360企业版年费15万美元。自托管Plausible免费。
仓库 → https://t.co/euffyRbVR4

6. AppFlowy
开源Notion。Notion团队版每用户每月20美元。AppFlowy在你的服务器上无限用户免费。
仓库 → https://t.co/blaPL1JcdJ

7. Penpot
开源Figma。Figma每编辑者每月45美元。Penpot自托管永久免费。

仓库 → https://t.co/y85SRFnVtV

8. n8n
开源Zapier。Zapier Pro月费600美元。n8n自托管无限自动化免费。

仓库 → https://t.co/UCCGx45D9W

9. https://t.co/cLSNPPAAb5
开源Calendly。Calendly Teams每用户每月16美元。https://t.co/cLSNPPAAb5对个人免费，团队版开源。
仓库 → https://t.co/XF0jFFHPaM

10. Bitwarden
开源1Password。密码管理器每用户8美元。Bitwarden无限永久免费。

仓库 → https://t.co/SDpxdaSSol

最疯狂的部分：

这些 GitHub 仓库，正在悄悄干掉每年 500 亿美元的商业软件收入。

它们不违法。

但离谱的是：

它们免费到像是应该违法。

收藏这条，顺手转给那个还在为“本该免费”的工具付钱的人。

100%免费。100%开源。

[HankAI @hank_aibtc]

模型地址：https://t.co/4PyFCBkxTl
浏览器 Demo：https://t.co/dhcnThhbkR

[Kirill @kirillk_web3]

instead of watching 2 hours of Netflix tonight, watch this 40-minute masterclass from the founder of a $20B China AI company

it's the clearest explanation I've seen of how Agent Swarms and AI systems actually work at scale

useful whether you've never built an agent in your life or have been using Claude every day for the past year

I took the key ideas and turned them into a practical guide on how to actually build with Kimi

find it below

[Kyle Hessling @KyleHessling1]

BREAKING! Qwopus 3.6 27B is LIVE!

Thank you for your patience on this one, but I believe you'll find the wait was worth it!

We've benchmarked this thing up and down, verified that it holds at least a 75.25% (152/202) in the initial 202 SWE bench solves. Not a full run of 500, but it shows the agentic coding quality from the original 27B is retained while adding all of the additional Qwopus benefits across many domains. As always, Jackrong is absolutely cooking here!

COT quality has improved significantly through the inversion techniques from our Negentropy proof of concept. It also went through thorough curriculum training. You can check out the MMLU pro benchmarks on the model card, but it improved a whopping 10 points over the base model in physics, as well as meaningful jumps in Chemistry, business, and computer science.

However, the best part is that I was able to build an entire survival shooter game using this local model entirely. I genuinely was blown away by the results, which you can play right now on my HF space (link in comments below). "Qwopus Commander" was completed in 9 turns of Qwopus 3.6! To test the new long context training, I made it re-output the entire 3000+ line program each turn, and it would make fixes and add features that I requested in large prompts, while perfectly replicating the entire rest of the game from context. What's more is that I did it all at Q8 KV cache quantization, and never had an issue over the entire 303k token run!

IMPORTANT: Run it at --temp 0.75 to 1. Mess with it in that range for your use case. Higher temp actually lets the fine-tune shine and be exploratory and is also more stable. Swe Bench was run at temp 1, the game was built mostly at 0.8!

We're so blessed to have all of you here and using the models! The support means so much! Please let me know what you build with it in the comments! Or if you have any issues getting it up and running, I will try my best to get back to you!

Looking forward to seeing what you legends produce with it this weekend!

https://t.co/AEl3APtTLk

[Nav Toor @heynavtoor]

Repo: https://t.co/2KTjosu0ly

[tuna🍣 @tunahorse21]

early checkpoint if a bigger run, getting expensive so im taking it slow ( im just a wagie doing this for fun) if anyone wants to throw me some gpus for sponsored run dm plz

no evals yet, but my smilar training run like the 9b and those evals did well

https://t.co/EN5guR8TLg

[Ahmad @TheAhmadOsman]

Current models rotation (mix of API & local)

> GPT 5.4 Pro (Subscription)

> MiniMax M2.7 (API) / M2.5 (local)

> GLM 5.1 (API) / 4.7 (local)

> Kimi K2.5

> Qwen 3.5 397B MoE

> Qwen 3.5 27B Dense

[Kyle Hessling @KyleHessling1]

The Opus distilled Qwen 27B is going viral again, but everyone is promoting the old version! You MUST get the v2 version, which Jackrong currently does not have listed in his Opus Fintune collection, so no one is finding it. I discussed the improvements with v2 in my last week's test. Please go check out that post with the video, pinned on my page, if you get a second, and download the v2 here: https://t.co/tBMGR7yJVB

[Eric ⚡️ Building... @outsource_]

For those wanting an Opus distill but in a smaller, more accessible 9B size that easily fits on GPUs with 16GB VRAM or less:

Jackrong just dropped Qwopus3.5-9B-v3 — a fast-iterated reasoning-enhanced model based on Qwen3.5-9B with Claude Opus-style patterns distilled in.

The Q4 quant is only 5.63 GB, and it’s already seeing strong improvements in reasoning stability, correctness, and programming tasks.

Perfect for local runs without needing massive hardware.

https://t.co/4ZELVa5L9Z

[Unsloth AI @UnslothAI]

This model has been #1 trending for 3 weeks now.

It's Qwen3.5-27B fine-tuned on distilled data from Claude-4.6-Opus (reasoning). Trained via Unsloth.

Runs locally on 16GB in 4-bit or 32GB in 8-bit.

Model: https://t.co/6KgPDHCJZ3

[GitHub Projects Community @GithubProjects]

https://t.co/jHidbOhg37