Search “DeepSeek”

Yuyuan Tantian explains the rumors of China's AI technology export ban

In response to outside claims that China is about to shut the door on AI technology exports, Yuyuan Tantian sorted through the background and regulatory boundaries of the public consultation opened on October 16 by the Ministry of Commerce and the Ministry of Science and Technology on revising the Catalog of Technologies Prohibited and Restricted from Export in China. The article points out that regulation focuses on three categories of technology: data security, sensitive uses, and computing power foundations. When enterprises use restricted large models to provide external API calls, APP/webpage interaction, SaaS applications, and the like, this does not fall within the scope of large-model technology export management; open-source technologies and specific international businesses are eligible for facilitation measures such as automatic licensing. The article also mentions that two companies, DeepSeek and MiniMax, have completed registration procedures and finished open-sourcing and overseas authorized use.

Artificial Analysis Coding Agent Index v2.0: Claude Opus 5.5 tops with 61.7

Artificial Analysis released Coding Agent Index v2.0, with Claude Code paired with Claude Opus 5.5 (max) topping at 61.7 points, followed by Claude Sonnet 5.5 (max) at 59.3 points. The new version replaces DeepSWE with Cognition's FrontierCode v1.1; Opus 5.5 scores 56.0% on that benchmark, ahead of Codex paired with DeepSeek V4.1 Flash at 52.6%. SWE-Atlas-QnA-AA switched to GPT-6 Luna (xhigh) for judging, with five votes per item; the average score across 21 configurations fell from 58.8% to 57.1%.

Mistral Research Lead Discusses Le Chonk: The Inside Story of the West's Largest Open Source Model Catching Up

Matt Turck speaks with Mistral research lead Pierre Stock about Le Chonk, described as the West's largest open source model. Stock discusses the team training on 3,800 GPUs in its own data center, the fact that Le Chonk is still in training, what Mistral borrowed from DeepSeek, and the trade-off between open-sourcing weights and refusal behavior for cybersecurity. The episode also mentions that Mistral Large 5 will be released later.

LMSYS Releases SGLang AWS Trainium Plugin Beta

LMSYS, in collaboration with AWS, has released the SGLang Trainium plugin Beta, built on the newly released TorchNeuron, supporting LLM inference serving and RL rollout generation on Trainium2 and Trainium3. Initial coverage includes Qwen3 (dense), Qwen3-MoE-30B-A3B, DeepSeek-V2-Lite, GLM-4, and GLM-5.2/5.3, and supports data parallelism, tensor parallelism, expert parallelism, Radix Cache, BF16/FP8, and asynchronous scheduling. In the first end-to-end heterogeneous RL run, SGLang generated rollouts for Qwen3-MoE-30B-A3B on Trainium3, while Miles trained a GRPO policy with Megatron on a single H200 node. The Beta is currently limited to a small number of teams.

New paper proposes BackUpper: making coding agents back up before they act

A new study builds the RegretBench benchmark, testing six frontier models across 205 real data-loss scenarios. Results show agents complete tasks at an average rate of 86.1%, but back up before critical operations only 15.2% of the time—GPT-5.6-Terra at 4.4%, GPT-5.6-Sol at 5.9%, Gemini 3.7-Flash at 2.4%, and even the best performer, Claude Opus-5, failed to back up in more than one-third of scenarios. The authors' proposed BackUpper runs in parallel with the agent as an external plugin, using a four-tier strategy: replacing destructive commands, pairing backup commands, generating snapshots, or calling the GPT-5.6-Luna sub-model to determine which data needs saving. In testing it covered all critical operations, with recovered data averaging 92.2% similarity to the original state, at an additional cost of $0.014 per task. BackUpper already supports Claude Code, Codex, Hermes, OpenHands, DeepSeek Harness, and OpenClaw, and its code and dataset have been open-sourced.

TorchTitan Update: Support for Multiple New SOTA Models, TitanRL, and Large-Scale Performance Improvements

PyTorch's native training platform TorchTitan has released an update that adds full training-stack support for Muse Spark Open, Muse Glimmer, Qwen3-VL, Qwen3.5, Kimi K2.7, Kimi K3, and DeepSeek-V4, and introduces the reinforcement learning component TitanRL. TitanRL lets pretraining, post-training, and vLLM inference share the same model code, and supports asynchronous rollout as well as agent-task training such as long-context and code execution. On performance, FlexShard on 256 H100 GPUs reduces DeepSeekV3 optimizer time by up to 79%, and reduces end-to-end training step time on Llama3 405B by up to 20%.

SGLang's Elastic Expert Parallelism Scaling: Online Scaling of MoE Deployments with NIXL

SGLang's Elastic EP Scaling lets MoE deployments add GPUs in fine-grained increments without full replicas or restarts, scaling from 20 GPUs to 24 GPUs in the example. max_ep_size reserves metadata and communication buffer capacity at startup, and new ranks start on demand; tested on 8×H100 with DeepSeek-V3-Lite FP8, the first completed request was 5.3×–8.2× faster than a restart. After a joining rank enters the Mooncake WORLD, automatic rollback is not currently supported, and a failure during finalization may require restarting the deployment.

Modal Releases LLM Training Library Modal Dojo

Modal has released Modal Dojo, a training library that provides tuned default recipes, modifiable parameters, and an out-of-the-box metrics observability dashboard, supporting models from 500 million to trillions of parameters. The library is built on open-source frameworks such as Miles and Slime, can load datasets from Hugging Face, and lets agents query run status, logs, and trajectories via a CLI. Axiom Bio says it has used it to launch training of DeepSeek V4.1 Flash with a 64k context, running on 64 B300s.

Alibaba Cloud Model Studio Token Plan Team Edition Subscription and API Access Guide

Alibaba Cloud Model Studio Token Plan Team Edition uses Credits as a unified billing unit. A single subscription provides access to more than ten text and image models, including Qwen3.7-Max, GLM-5.2, DeepSeek-V4-Pro, and Kimi-K2.6, and is compatible with both the OpenAI and Anthropic protocols. The plan comes in three tiers: Standard with 25,000 Credits, Advanced with 100,000 Credits, and Premium with 250,000 Credits. A dedicated API Key with the sk-sp- prefix can be generated only after seats are allocated. Access requires using the corresponding Base URL, and the plan is completely isolated from the Coding Plan and pay-as-you-go systems.

AI Tries to Take Over Scanlation Groups' Jobs, Only to Find They Don't Rely on Translation to Make a Living

Models such as DeepSeek, GPT, and Qwen can already handle the collection, upscaling, translation, and typesetting of manga scanlation, but when they encounter explicit content or copyrighted content they refuse to translate outright, so a "human feedback" step is still needed to fill the gaps. The work of scanlation groups has therefore shifted to proofreading tone, handling puns and cultural references, adjusting for differences between Chinese and Japanese typesetting, and even drawing clothes onto exposed characters to pass review.

Franchise Recruitment GEO: An Analysis of the New Customer Acquisition Logic for Franchise Brands in the AI Era

Franchise recruitment marketing is shifting from search bidding and feed ads to GEO (Generative Engine Optimization), with the goal of ensuring that brand facts are accurately mentioned and reasonably recommended in generative Q&A systems such as DeepSeek, Doubao, Tongyi Qianwen, ERNIE Bot, Kimi, and ChatGPT. The article compares the business logic of three models—search advertising, feed advertising, and GEO—and points out that GEO competes on the quality of facts and the credibility of information sources, not on advertising budgets. To judge whether a GEO service provider is reliable, it is recommended to prepare 20 to 50 real franchise recruitment questions and require an on-site demonstration of multi-model queries, brand mentions, cited sources, and erroneous information identification.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

Flatkey Gateway Lets Codex/Claude Code Call Multimodal Models

Flatkey, as a unified API gateway, lets Codex and Claude Code call text models such as GPT, Claude, DeepSeek, GLM, and Kimi as well as image and video models such as GPT Image and Seedance via CLI. One Key provides access to 100+ official models and 1,000+ tool APIs, with costs consolidated into a single bill. Everyday prices are at least 20% lower than official pricing. During the October 19–23 promotion, GPT, GLM 5.2, Kimi K3, and DeepSeek are 40% off, Claude is 20% off, and Seedance 2.0 is 15% off.

MiniMax M3.1 Flash Tops Search Efficiency Leaderboard

Parallel Web Systems' search capability efficiency leaderboard shows MiniMax M3.1 Flash in first place, with web research performance comparable to DeepSeek V4.1 Flash (1017) at roughly 3x lower cost. MiniMax officially reposted and thanked them, emphasizing that the model is very fast.

English translation is being prepared.

Memory Tensor Open-Sources MemT 1.0 Series Memory-Native Models

Memory Tensor has open-sourced and released the MemT 1.0 series of memory-native models, including MemT-1.0-Nano with 36B parameters and 3B activated (based on Qwen3.6-35B-A3B) and MemT-1.0-Flash with 288B parameters and 17B activated (based on DeepSeek-V4-Flash), adding a memory computation pathway parallel to attention in selected Transformer modules. The team built a 34-billion-token memory training corpus, of which 20 billion tokens were used for continued pre-training and 14 billion for post-training. In a comprehensive evaluation across 11 memory benchmarks, Flash and Nano achieved average scores of 50.6 and 47.0, respectively, ranking in the top two among the parametric memory methods evaluated; on 500 SWE-bench Verified tasks, the proportion of tasks that Nano succeeded on all three attempts rose from 54.00% for the no-experience baseline to 62.80%.

SGLang-TorchTPU Open-Sourced: Bringing the SGLang Serving Stack to TPU

The SGLang team, in collaboration with Google and Meta, has open-sourced SGLang-TorchTPU, a PyTorch-based native backend that brings the full SGLang serving stack to Google Cloud TPU and is already in large-scale production use at Meta. Initial support covers models including Muse Spark Open, GLM-5.3, Kimi K3, DeepSeek V4, Qwen3.8, and Gemma 4, with features such as speculative decoding, KV cache offloading, PD disaggregation, and FP8 KV cache. The team also used TorchTitan and Miles to complete GRPO training of Qwen3-4B on a single 4-chip TPU v7x, with the reward curve matching a 6× H200 baseline within random fluctuation.