AI Briefing#deepseek
The latest AI and developer news about #deepseek, with the original source and a short summary.
The latest AI and developer news about #deepseek, with the original source and a short summary.
Unit 42 research reveals Chinese-speaking attackers autonomously compromised 10 organizations in semiconductor, defense, and civil society sectors using DeepSeek and Zhipu GLM agents in Aug-Sep 2026.
Gemini 4 Hydrogen leads the proprietary field with 13.5%, while GPT Astra 6 follows at 9.5% on the same challenging benchmark.
The credit applies to featured open-weight models on AI Gateway, including Kimi K3, Kimi K3.1 Pro, GLM 5.3, GLM 5.3 Flash, and DeepSeek V4.1 Flash.
V4.1-Flash enables server-side tool calls for search, file input, and safety checks, with standalone APIs available for OCR, ASR, and TTS.
A forensic report identifies NVIDIA Blackwell hardware signatures in the initial weights of six Chinese models, including DeepSeek-V4-Flash and Kimi K3, indicating access to export-controlled chips.
In a new benchmark requiring simultaneous fixes for two real bugs, DeepSeek V4.1 Flash achieved a 90% success rate at a cost of 6 cents per task, surpassing GPT-6 Luna.
Removing unnecessary draft-token reservations allows larger prefill chunks and more efficient CUDA graph execution, raising output throughput from 1,373 to 3,390 tok/s.
The model features a 1M-token context window and is available in public preview without waitrooms or session limits.
GPT-6.1-Sol ranked first in coding and mathematics, while Kimi-K3.1-Pro took the official overall first place.
The tools work with any model supporting tool calling via a single AI Gateway API key, requiring no separate DeepSeek account.
Released under the MIT license, the agent operates in observe mode by default and requires human approval for auto-merges.
The move would mark the first time Ascend chips are used for full pre-training, following a September 2026 release of open-source infrastructure components adapted for the platform.
DeepSeek V5 is planned to be pre-trained on Huawei Ascend chips, marking the first time a top-tier Chinese model uses domestic hardware for this phase.
The expert-pruned build retains 98.7% of LiveCodeBench v6 performance while reducing transformer weights to 72.1 GB.
Enabling the vLLM 0.31 server-side constrained decoding flag improved tool-call schema accuracy from 92% to 100% for Kimi K2.6 and DeepSeek models.
ISTA DASLab releases GSQ-RCO quantized GGUFs for the 763.2B parameter DeepSeek-V4.1-Flash-1017, achieving 100% benchmark retention at 2.00 bpw on a 12GB GPU.
DeepSeek charges $0.75 per 1,000 requests for its API, while Tencent claims its engine handles hundreds of millions of daily calls across 157 authoritative sources.
The model is now accessible for use with agents within the OpenClaw platform.
The clearance grants explicit regulatory approval to continue open-sourcing model weights and securing authorized international commercial partnerships.
DeepSeek has released the offline inference code for DeepInfer, its inference framework for V4.1-Flash.
The library supports NVIDIA Turing and later architectures, achieving significant speedups over vLLM Marlin and DeepGEMM across dense and MoE workloads.
The open-source DeepInfer framework supports NVIDIA and Huawei Ascend platforms with a single-cluster throughput of 155 trillion tokens per day.
The release supports NVIDIA SM100 and Huawei Ascend 950 hardware, featuring optimized kernels and asynchronous scheduling for offline inference.
Output costs for DeepSeek V4 Flash 0731 increased from $0.40 to $1.32 per million tokens, part of a broader wave of API price hikes across major labs.