We got a new 26B in the house. 💪
Inkling 2 Mini is a Mixture-of-Experts model with 26.4B total parameters.
thinkingmachines.ai/news/ink…
Some of you thought that Inkling 2 Small was too big at 594B parameters. We heard you: Inkling 2 Mini is a looped transformer with just 26B parameters, small enough to run on your laptop, powerful enough for agentic work with a score of 20.1 on AA Index.
thinkingmachines.ai/news/ink…
1
18
tarlow 🌱 retweeted
Today we’re launching Shield: AI-native defense against offensive agents.
Attacker agents probing your site? Don’t block them. Deceive them.
We pointed ARTEX — the autonomous AI pentester the FBI just warned about—at our own prod. 5 minutes in, it pulled /.env and /.git/config (200 OK) with credentials inside.
Shield turned ARTEX attack surface into a synthetic maze with fake endpoints, decoy credentials, and poisoned responses that hijacked its agent memory.👇
It takes time to rebuild and create something worth it.
Give yourself time.
A small flow of water can carve a river through a landscape. It just takes time.
Just go with the flow.
You are on the right path my friend. 💚
14
Planning, drafting ideas, and building the hand over prompts with Gemini 4 Hydrogen.
Then pushing out the work with DeepSeek V4.1 Flash 1017 via API is feeling really solid.
Design, depth of scope, prompt, and end result much better.
38
Cool new paper: just like people, AI learns a lot faster when it keeps notes on what it tried and why, not just what worked in the end.
I have been experimenting with this on my Marimo Bot project too.
arxiv.org/html/2611.18856v1
16
Don't sleep on it. 💪
Banger paper from Google on RSI.
(bookmark it)
They propose an approach for continual recursive self-improvement at the harness level. They also propose their own self-improving memory system for the harness optimizer.
This clearly highlights the importance of traces and having a functional, adaptive, and structured persistent memory system in your harness.
They show that a harness optimizer improves 9 of 10 continued runs when it remembers why it made each past edit.
They discuss that a common problem shows up when you change the model or the benchmark. An Opus-evolved harness started at 68.41 on Gemini, below the 73.89 of the plain seed harness.
Self-Improvement Memory groups past edits by the harness problem they addressed. For each edit, it stores the rationale, predicted outcome, observed outcome, and evaluation context. The proposer and the selector both read it before making new changes.
Across four domains, including Terminal-Bench 2.1 to Terminal-Bench 4, SIM beats continuing from the harness alone in six cross-dataset and four Opus-to-Gemini settings. Across ten coding runs, it adds 2.36 points on average (p=0.003).
If you evolve harnesses automatically, store the reasons and results of each edit next to the code.
arxiv.org/abs/2611.18856
Chat with Paper: academy.dair.ai/papers/self-…
11
tarlow 🌱 retweeted
让 AI 边看视频边理解,门槛终于被一脚踹到了地板上!
Hugging Face 刚开源了目前全世界体积最小的视频多模态模型:SmolVLM2-256M
很多人以为给 AI 装上“眼睛”去读长视频,动不动要吞几十 GB 显存。这个仅有 256M(约 0.25B)参数的小家伙,直接把配置要求杀到了极限:
· 只要 1.38GB 显存:无论是轻薄本、手机端、树莓派还是几年前的老旧显卡,不到 1.5G 显存就能直接在本地完成视频推理!
· 专职“看视频”而非“生成视频”:吃进一段视频、一组连续帧或游戏屏幕录屏,它能实时用自然语言理解剧情、描述动作、做视觉问答(VQA)以及 OCR 文字提取。
· 跑分扎实的小钢炮:Video-MME 达到 33.7,MLVU 达 40.6,对于一个只有 256M 的模型来说,这理解能力相当凶悍。
· 开箱即用全免费:Apache 2.0 纯开源协议,GGUF 量化版也已经就绪,拉进 Ollama 或者 llama.cpp 就能单机离线跑。
拿来做实时游戏画面的屏幕助手、本地监控动作识别,或者挂在低算力边缘设备上,简直是神仙底座。
想在千元机或轻薄本上折腾“视频视觉 Agent”的兄弟,这个微型模型建议先存一份备用!
👉 模型地址与 GGUF 入口见评论区一楼
#AI #开源 #HuggingFace #大模型 #多模态
If you are using Gemini this is a huge time saver.
They have built automated tasks directly into gemini chat and it can easily connect to everything you already have.
I have already automated a few things and the process is smooth.
Put Gemini to work for you with Tasks.
While Chat is for quick back-and-forth answers, Tasks are for longer projects and recurring schedules.
Beginning today, Tasks are now free of charge for hundreds of millions of eligible users, starting in the U.S., JP, KR, CA, AU & IN. 🧵
2
55
Since I mostly use the direct DeepSeek API.
I like to get most of my work done during Deepseek's off-peak timers.
It's cheaper too.
But during peak hours I don't stop.
Instead, I either start getting future prompts ready or switch to a host on Openrouter serving the model.
2
1
124
Remember, if you are enjoying another creators work to hit that like or comment button and leave a little note or emoji.
Helping each other grow is how we change. And, the people sharing or grabbing your attention deserve a little bit of love.
So, don't be shy. Spread the love.
1
1
35
I posted this because I was listening to a great music set.
And I used to never leave a like or comment on stuff...
Now, if I feel it. I say it or do it.
Why hold back from sharing some love?
It helps people trying to grow.
Oh, here is the set: youtube.com/watch?v=ivVjmNP4…
9
tarlow 🌱 retweeted
📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery!
Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window.
Highlights:
- Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation.
- Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables.
- Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery.
- $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching — all lower than Qwen3.8-Max, with implicit caching down by 70%.
The API is live on QwenCloud. You can also try Qwen4-Max on Qwen Studio, QwenWork and Qoder.
Start building today! 🚀
- Blog: qwen.ai/blog?id=qwen4-max
- API: qwencloud.com/models/qwen4-m…
- Token Plan: qwencloud.com/pricing/token-…
More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀
If you think that everyone will build and launch software then you gotta step outside for a bit.
It's easy to get caught up with what you see right in front of you, and think that is all there is.
Take time to reconnect.
Not everyone is a builder, some just follow and consume.
21
Oh very nice, I have been enjoying this model. It's been helping me refine a lot of content and sections in my site.
谷歌悄悄开放了 Gemini 4 Hydrogen 免费 API,又有羊毛可以薅了!😂
支持直接接入各种 Agent 客户端,日常开发又多了个免费模型可用。
📌 核心参数与免费额度:
1、上下文:100 万 Token
2、RPM:每分钟最多 15 次请求
3、RPD:每天最多 1,500 次请求
4、额度按 Project 计算,同一项目下的多个 API Key 共享配额。
获取方法:
1、打开 aistudio.google.com 登录
2、创建 API Key,复制密钥
3、在支持 Gemini API 的客户端中配置密钥,模型 ID 填写 gemini-4-hydrogen
免费额度拿来跑跑日常 Agent 任务,还是挺香的。
23
tarlow 🌱 retweeted
llama.cpp v1.0.0 is now available
huggingface.co/blog/ggml-org…
















