That said, these H200s won’t sit idle forever.
My off-the-cuff guess: as domestic chips’ adaptation and supply get on track, it will gradually open up from Q1 next year to ease shortages, maybe every added H200 is matched by equivalent-compute domestic chips, so basically 1:2.
Show more
The northernmost data center in China, enjoy free cooling for most of the year.
Oh, in case you didn't know, data centers in China have green electricity ratio and PUE requirements, instead of FOMO and being completely unregulated.
Show more
This is why I believe the visual capabilities of V4.1-visual exp/V4.2 will improve, and 1017 has already seen signs of this.
Simply because Deepseek began recruiting multimodal experts several months ago, and some recently joined the team.
Show more
There won't be a V4.1 Pro in the short term🥲
dsh.v0.2.1.rc.1 ready, may be released in a few hours along with V4.1 Flash 1017
The official Bilibili account for Ascend CANN will be introducing a reference practice for pre-training Deepseek V4/V4.1F using Ascend 950 supernodes in live stream tomorrow.
Deepseek has effectively become part of China's digital infrastructure;
It wasn't until its rare, prolonged daytime outage that you realized how many people use it at work, and how many platforms are simply routing requests to its official API.
Show more
A team led by Xu Disheng, a visiting scholar at the Shanghai Artificial Intelligence Laboratory, has completed a proof of the 3D Ghys conjecture using the Intern-S2 model.
20 kW is not enough for EUV.
However, Langdao is not a major supplier of EUV light sources. They don't even work as a backup option, which is why they are showcasing it.
One example is code. A guy on Bilibili makes review shows using real bugs from his company's "shit mountain." V4.1 Flash might resolve "diamond-tier" bugs that only Grok 4.7/Opus 5-level could handle, or might fail on "silver-tier" bugs EVEN Doubao can one-shot.
It's just random
Show more
Pro 20x quota halved for same reason:
GPT-6 Astra launch gave Plus users too many bank resets. Relay sites stockpiled Plus accounts; when used, upgraded to Pro 20x, instantly burned natural weekly quota + one bank reset to max it, then rolled accounts to sell quota in bulk.
Show more
DeepSeek employee Xu Ruifan shares some thoughts on developing Ascend infrastructure components.
Looks like they really are on holiday.
Deployment of DeepSeek-V4.1-Flash using FP8-FP4 mixed on 32 Ascend 950DT.
For 128K sequence scenarios, single-card decoding throughput reaches 5,800 TPS at 13.79 ms. In low-latency inference scenarios, latency is kept under 5 ms while achieving a throughput of 2,727.
Show more
“950 super-node enables seamless scaling from 550B and 1.6T models to 10T-class models.”
Off-peak hours: Cache hit price: 0.02 yuan; Cache miss price: 1 yuan; Output price: 4 yuan.
Peak hours price is twice the off-peak hour price.
Started on September 10th.
Astra Blender MCP is largely a gimmick
Dedicated 3D generators like Tripo already far exceed non-experts' expectations in both precision and speed.
By comparison, Astra's point-by-point MCP drawing is to these models what hand-forging is to an electric arc furnace.
Show more
@teortaxesTex GLM-5.3 Omni, I heard about it back in the ox alpha
Jihao Liu, an algebraic geometer from Peking University, used LLM to disprove the Yau-Tian-Donaldson conjecture.
Most importantly, this is not a "low-hanging fruit" combining existing human achievements; LLM creates innovative mathematical methods.
Show more
So many routers offering Ox Alpha for free is ridiculous; it doesn't seem like something could provide.
Furthermore, to be fair, Ox Alpha isn't better than GLM-5.3; it's more like a post-trained version of GLM-5.2. Could this be an Nemotron 4 or MAI 3?
Show more
The community reports that the official DeepSeek API appears to have changed checkpoints or enabled new A/B testing, significantly better performance in some cases compared to the current V4 Pro.
They seem to be working on fixing the issue.
Show more
