Open source maximalist. Chinese AI models you can actually own. DeepSeek, Qwen, Kimi , Xiaomi and all the good stuff.
Joined June 2026
- Tweets414
- Following11
- Followers9
- Likes263
CNopen retweeted
NVIDIA put four Cosmos 3 world-model checkpoints on Hugging Face, open weights, 16B trainable parameters each.
They generate driving and robot simulation data, and NVIDIA shipped the four-stage post-training recipe alongside them.
What you get:
1. Transfer-Auto: text prompts and per-camera scenario maps in, up to 11 synchronized camera views including fisheyes, plus joint RGB-LiDAR out. Weather, lighting and time of day are controls.
2. Sim-Auto: the streaming variant, causally distilled to four denoising steps, built for closed-loop driving rollouts.
3. Sim-Transfer: takes a physics simulator's depth, segmentation and blur controls plus a prompt and renders RGB for a policy to consume.
4. Sim-Bimanual: predicts the next observation from two-arm robot actions, for YAM and AgiBot embodiments, 59 control values per step padded to 64.
All four are post-trained from Cosmos 3 Nano. License is OpenMDW1.1, commercial and non-commercial use allowed, so the cost is your own compute. Inference needs Linux, an NVIDIA GPU, Docker and the NVIDIA Container Toolkit; the published vLLM-Omni setup was tested on one GB200, which is not a minimum-memory spec.
The useful part is adaptation, not inference. Post-train Transfer-Auto on your own camera layout and LiDAR config first, then distill it into Sim-Auto if you need streaming. NVIDIA says the open problem is measuring where these models actually support policy learning.
Point your agent at the open skill, run a short training job, and check quality on held-out episodes before a long run. A clean short run proves execution, not convergence.
CNopen retweeted
Gemini Omni 1.2 Flash now makes video clips up to 15 seconds long, up from 10, and costs 30% less than the last version.
It's rolling out today in the Gemini API, Google Vids, and Gemini for business. Video costs 7 cents per second at 720p. You can change a finished clip with a follow-up message instead of generating it again.
If you don't code, you'll probably meet it first in Google Vids and Flow. Faces and brand colors stay the same from the first frame to the last, which makes it a good fit for product clips and ads.
Moonshot AI launched Kimi Hosted Agents, a managed runtime that runs agents on its own cloud.
Isolated sandboxes, stoppable sessions, exportable logs, 140+ skills and 30+ plugins.
Sandbox time costs 0.072 yuan an hour, with no runtime to maintain.
CNopen retweeted
BREAKING: Claude introduce personal finance for us market
CNopen retweeted
Proactive agents are a different product than chat. Once they can run computer tasks with their own memory and tools, the useful ones will be the ones you leave overnight, not the ones you babysit.
How to run ERNIE-5.5-Flash on vLLM-ERNIE:
1. Install vLLM 0.32.0, then Baidu's plugin from GitHub.
2. Serve FP8 on 8 80GB GPUs, data parallel 8, expert parallel on.
3. Send prompts to /v1/completions, not chat: the Base model continues text.
A 128K context fits on those GPUs.
CNopen retweeted
For teams that kept regulated data away from the API, this changes the conversation. The records stay on your side and an automated agent does the review, so the open question is which workloads legal will now approve.
CNopen retweeted
Asking how a change can be verified gets the agent to name the check before it writes code. Real signups and real clicks catch bugs that a passing unit test never sees.
Alibaba's Qwen team released BabelArena, a benchmark that puts the same 699 agent tasks into 23 languages.
Its tests show low-resource languages burn up to twice the input tokens of English, and agents drift back to English.
Code and data are open: score your own agent.
CNopen retweeted
A 65 MB text-to-speech model now tops the CPU streaming benchmark for time to first audio.
Phonon shipped two weeks ago and has passed 200,000 downloads. On Hugging Face's Open TTS Leaderboard it ranks first on intelligibility in German and Spanish.
1. Two sizes: 31m for English at 65 MB, 40m for English, French, German, Spanish and Portuguese at 75 MB.
2. Phonon 31m is Apache 2.0, free including commercial use. The 40m model is CC BY-NC, priced by user count if you ship it commercially.
3. Top three on intelligibility in every language evaluated, against models up to 100 times its size.
4. Runs on phones, laptops, single-board computers and in a browser tab. No server, no network, the text stays on the device.
5. In Transformers from v5.21, with a ptts runtime covering CPU, CUDA, Vulkan, Metal, WebGPU, Core ML and Snapdragon.
At 65 MB and Apache 2.0, it is the one to try for a voice that works offline. Load gradium/phonon-31m-en in Transformers and generate a wav before you design around it.
ByteDance-Seed open-sourced a cross-layer transcoder for Qwen3-8B-Base: 32,768 features that show what drives a prediction.
Its features read the stream before each MLP and write into later layers, so one feature spans the model.
Apache 2.0: encoding fits on one 80 GB GPU.
CNopen retweeted
Breaking!🔥OpenAI released GPT-Image-2.6, in two versions: Flare and Sunburst. Both are in ChatGPT and Codex from today, on every plan, desktop, mobile and web.
The two things I care about: it does GIFs now, and it accepts reference images.
On speed they say 50% faster. We'll see if that holds once you throw a real prompt at it.
Pricing: $8/$5 input, $30 output, same for text-to-image and image-to-image.
How to serve mixed GPUs with GPUWorks:
1. Install from FlitAI's GitHub: 40MB, one command, no Kubernetes.
2. Pick from 1,000+ launch configs for nearly 400 models.
3. Turn on idle-release: models share GPUs, waking in about 4 seconds.
One pool, one OpenAI-compatible API.
CNopen retweeted
And just like I said in my article , this is going to be come also a PERMANENT cost for business , to be able to defend properly.
xitter.gq/altcoindaily/status/21…
We chat started cooking too
One small open model can find objects, read text, and cut the subject out of a photo, all in one go.
Tencent's WeChat Vision team released WeVisOne-VL in 2B and 4B sizes under Apache 2.0. Hosted demos are live for both, so you can try it today.
What it does:
1. It draws boxes around the objects you name, points at them, and counts them. If the object isn't in the picture, it answers "None" instead of guessing.
2. It reads text in images, so you can pull words off signs and scans.
3. On dense detection, the 4B model scores 71.3 against 58.7 for LocateAnything-3B, per Tencent's own tables.
The weak spot is saying no. It correctly rejects absent targets only 52.6% of the time on PR-Bench, and those numbers are self-reported.
Still, if you've ever stitched three tools together to find things and cut them out, this one is worth an afternoon of poking.
Super nice !
Tencent released Hy4 today, its new flagship open model, and the API price stayed where it was.
770 billion parameters in total, 49 billion active at once, 1 million tokens of context. First Hy model that reads images natively, charts and visual math included.
It also does Computer Use: terminal commands and clicks inside desktop apps, trained on Blender, CAD and long jobs across several programs.
Hallucinations fell from 15.76% to 5.45% against the August preview. GameDev +19.1%, GameLogic +9.8%, multimodal agent +36.8%, document tasks ~97% done.
On 579 real financial tasks it leads the open models at 3.04, ahead of DeepSeek-V4.1-Flash at 2.97. On Tencent's own agent benchmark it matches or beats Claude Opus 5.5 at about one eighth of the output price.
Still behind the top closed models on research math and very long coding sessions. $0.83 per million input tokens, $2.50 output.
Lenovo's Kaitian line shipped its first AI host, pairing a Zhaoxin CPU with a Houmo NPU to run 35B models locally.
The LQ50 NPU has 24GB of memory: Qwen3.6-35B-A3B runs at about 29 tokens per second.
Inference stays on the machine, so no cloud tokens and data stays local.
