Open source maximalist. Chinese AI models you can actually own. DeepSeek, Qwen, Kimi , Xiaomi and all the good stuff.
Joined June 2026
- Tweets302
- Following11
- Followers10
- Likes184
Qwenputer puts a 32-bit RISC-V CPU inside Qwen3.5-4B's weights: 11 compiled layers using its DeltaNet state as RAM.
It runs C and MicroPython in the forward passes, no tool call or sandbox.
Apache-2.0, loads as stock Qwen3.5: 97.5 on integer arithmetic where base gets 75.0.
How to run Baidu's Unlimited OCR 2:
1. Weights and code are on Hugging Face and GitHub, MIT license.
2. Start with Transformers for a quick test.
3. Serve production volume with SGLang.
4. Feed the whole document in one pass.
A fixed 512-token working set keeps memory flat.
MiniMax open-sourced mcode-stack, the skills its Code team uses daily.
Each is a SKILL.md file that makes a coding agent check its work: when to hand a task to a subagent, what tools cost in context, how to review a diff cleanly.
MIT, works in MiniMax Code and Claude Code.
One MCP server as the control plane for every machine is the right shape. The next bottleneck is not running 8 agents at once, it's reviewing what 8 agents produced.
ZwZ2-27B is a new vision model from Ant Group's inclusionAI, built on Qwen3.6-27B.
It learns fine details from cropped regions during training, so at inference it reads the full image in one pass, no zooming tools.
It beat its base model on all 25 benchmarks, +4.99 points.
OpenBMB released MiniCPM-OCR, a 1.3B model that turns Chinese and English PDFs and scans into Markdown, plus Jev, a 2B model that scores its output.
Jev flags the wrong regions, so you review only those, not the whole page.
Apache 2.0 weights, runs on vLLM and llama.cpp.
inclusionAI released SingProbe, an 8.13M-parameter guardrail built into GLM-5.3-Flash.
It reads the model's hidden states to flag unsafe replies and hallucination risk per token, adding under 0.5% decode time.
It beats Qwen3Guard-Gen-8B on safety and loads in SGLang or vLLM.
CNopen retweeted
Anthropic has written its no-ads promise into its IPO filing.
The cover letter says Claude's only incentive should be giving a helpful answer. It warns that once advertising enters a company's revenue targets, "the pressure to optimize for engagement tends to grow." It also takes a jab at ChatGPT, which already runs chat ads.
Keeping the promise has a cost, and Anthropic lists it:
1. Extra safeguards on Claude Mythos Preview, its most powerful model at the time, delayed its release as Claude Fable 5. That cost revenue and market share.
2. About $100 million in free Claude credits went to cyber defenders.
3. Anthropic expects more decisions like this as its models get stronger.
There is nothing to download or try today. It matters because the company has now admitted what the no-ads promise costs.
Ant Group's inclusionAI refreshed Ling-3.1-flash as Ling-3.1-flash-1019, a 560B model that runs 25B parameters at a time.
Extra post-training lifted its agentic benchmark scores and it keeps a 1M-token context window.
MIT license, and SGLang serves it day-0 on 8 H200s.
A team incubated by Xiaomi's technology committee showed HachimoDock, a desktop AI pet you build for about 150 yuan.
Talk to it and it turns the request into tasks for Codex or WorkBuddy on your PC.
Code and parts list are on GitHub, but the license is personal use only.
CNopen retweeted
AI didn't break the internet. It sent the bill.
Millions of attack attempts. No organization says it was breached. Two vendors withdrew their claims.
Inside: 8 AI predictions that missed, and a checklist if you run a site.
Has your host patched yet?
Tencent Hunyuan released Hy-SimulMT, a 30B model that translates Chinese and English speech live, returning text and cloned-voice audio.
It ran about 3 seconds behind the speaker on long-form audio, the lowest in Tencent's tests.
Apache 2.0, with GGUF builds for local use.
CNopen retweeted
Weekly AI Recap, Sunday edition. Here are the seven biggest breakthroughs from the week.
1. Terminal X Eval Agent → 88.6% match with senior investor judgment, and 25% more mandate deviations caught on GPT-6.1 Sol. Every graded output is scored against a 54-criterion rubric.
2. Nemotron 3.5 Bolt 74B → 74B total parameters with 9B active, shipped as native 2-bit (IQ2) weights that fit in 24GB. A 74B dense model must keep all 74B parameters resident in memory.
3. Mistral Voxtral 3 → speech split into three layers: 3 transcription models, 1 text-to-speech model, and a LiveKit-based voice agent SDK that wires them together.
4. Google Cloud Agent Compute Engine → one hardware-isolated microVM per AI agent, with no credentials inside it.
5. Nemotron 3.5 Bolt E4B 4GB → an open small language model for local voice agents, game characters and robots. NVIDIA starts from Gemma 4 E4B and changes its behavior through post-training rather than a new base.
6. Kimi K3.1 Pro → a post-trained successor to Kimi K3 with a 1,048,576-token context. The story is post-training, not a new base release.
7. Agent skills donated to the Agentic AI Foundation → procedures no longer need rewriting for each platform, so a skill can survive the next release instead of depending on one company's roadmap.
Bookmark this recap to catch up on the week in 60 seconds.
DS-Bot, a community plugin, turns DeepSeek Harness into a team of long-lived bots: each keeps its own memory, and a main bot called Chief runs the rest.
Group chats get an admin that calls in other bots for you.
Apache 2.0; runs on Windows, Linux, or Docker on a NAS.
This week in Chinese open AI:
1. DeepSeek, MiniMax cleared export rules; local runs abroad need a license.
2. Kimi K3.1 Pro: 1M context, $2 per million input.
3. Amap's Lambda-0: humanoid from human video, 62.5% real tasks.
4. Hy-Nomas: steer a video world, Apache-2.0.
BAAI released MS-Human-700, a simulated human body in MuJoCo: 206 joints, 700 muscle-tendon units, for reinforcement learning.
Gymnasium environments cover walking, stairs and right-hand manipulation, plus training code.
Apache-2.0, but you train the policies yourself.
CNopen retweeted
Qwen 4: new checkpoint (massive medieval voxel world)
this is checkpoint 3, still codenamed “Kalmia” on Arena.
> three checkpoints have now appeared in just four days
from many tests, Kimi K3.1 Pro is already getting mogged in 3D coding but it’s still not as good as DeepSeek’s Flash 1017