AITOPAI热报
Models

DeepSeek 据报道计划用华为 Ascend 950 芯片预训练 V5

TestingCatalog@testingcatalog
Post
View the post
Why we picked it

DeepSeek 用华为芯片练 V5，绕开 NVIDIA，2027 年跑全量，这个动向值得盯着看。

DeepSeek 计划在华为 Ascend 950 芯片上对 V5 模型进行预训练，完整训练预计在 2027 年进行。此举意味着 V5 训练将不依赖 NVIDIA GPU。此前 Cohere Model Vault 已上线 DeepSeek V4.1 Flash 等开源模型，腾讯 Hy4 也已在 OpenRouter 提供。

Original · TestingCatalog

DAILY AI BRIEF 🗞 — Nov 4

APPLE 🔥: - Apple opened Siri AI to all developers via an App Intents developer preview in the iOS and macOS 27.3 betas.

TENCENT 🔥: - Tencent launched Hy4 with vision and computer use on Tencent Cloud and OpenRouter, with open weights coming soon.

OPENAI 🔥: - ChatGPT Voice now works inside ChatGPT Finances for US users. - ChatGPT side chats no longer expire after an app restart or update.

ANTHROPIC 🔥: - Claude plugins now auto-connect at login to MCP servers that trust Claude as an issuer. - Cat Wu has left Anthropic.

SPACEXSI 🔥: - Grok Bot rolled out a new shopping experience to find deals, compare items and track orders. - Grok Bot passkey approvals for purchases are coming soon (via alexmichael). - Grok Bot can now post on X for you (via alexmichael). - Grok Bot is testing image markup to circle, pin notes or paint over parts of a photo for edits (via alexmichael). - Grok Bot is testing a usage breakdown chart by source or by bot (via blankspeaker).

GOOGLE 🔥: - Google's Dreambeans app is now available to everyone in India in 14 languages. - Google made Guided Learning globally available in Google Classroom.

PERPLEXITY 🔥: - Perplexity Computer now supports Agent Plugins imported from GitHub or a .zip file.

META 🔥: - Meta Muse added a dance feature with sharing to Instagram, Facebook and WhatsApp.

COHERE 🔥: - Cohere Model Vault now hosts open models like DeepSeek V4.1 Flash, GLM 5.3 Flash and Gemma 4.

DEEPSEEK 🔥: - DeepSeek reportedly plans to pretrain V5 on Huawei Ascend 950 chips, with the full run in 2027.

MICROSOFT 🔥: - GitHub Copilot is now available in JetBrains Air.

CLOUDFLARE 🔥: - Cloudflare launched Zero Wire, a public vulnerability intelligence feed on Cloudflare Radar.

OPENROUTER 🔥: - Stealth models Crouton Alpha and Pulse Alpha are free on OpenRouter and OpenCode with 1M context.

* Used Grok to compose this brief, cherry-picking the news and doing some post-editing.

#DeepSeek#华为#Ascend 950#开源模型

Keep reading

Recent stories with the same tag or section
PapersUS vs ChinaSelected10:50
Multi-layer Indexer Speeds DeepSeek Sparse Attention Decode

The indexer scan, not sparse attention itself, turns out to be the long-context bottleneck, and reusing selections across layers removes it without any retraining. Teams serving DeepSeek-V3.2 or GLM-5 long-context workloads on SGLang can adopt MLI for faster decode at unchanged accuracy, and it composes with the IndexShare reuse GLM-5.2 already ships.

#DeepSeek#DeepSeek Sparse Attention#GLM-5#SGLang
arXiv: DeepSeek@Serdarcan Dilbaz et al. (4)Source
PapersBuildersMulti-sourceSelected10:49
GC-MoE Operator Accelerates MoE Inference with CUDA Green Contexts

The technique shows that partitioning a single GPU with CUDA Green Contexts removes the resource contention between dispatch warps and GEMM warps that limits fused MoE operators like Mega MoE. Practitioners serving DeepSeek-V4.1-Flash or GLM-5.2 on Hopper or Blackwell clusters can adopt the open-source operator to cut decode latency without extra hardware. The results on H20 GPUs, a China-market chip, make it especially relevant for deployments where H100-class hardware is unavailable.

#GC-MoE#DeepSeek#GLM#CUDA Green Contexts
Event
arXiv cs.AI@Yong Li et al. (3)11 sourcesSource