Reflection could challenge DeepSeek and Qwen with a US model
7 sourcesAxios reports a release expected this month; the model's weights, license and benchmarks are not public yet.
New publication notifications
All DeepSeek news
All 31 AlexTech.ai stories on DeepSeek (the Chinese open-weight AI models) since June 9, 2026, newest first, each with its original sources.
Beijing probes DeepSeek and Moonshot over data routed to Claude
5 sourcesChina's Cyberspace Administration is investigating the firms for potentially exposing sensitive data to US servers via Claude.
CLOSEDQUORUM: Windows malware lets four AI models vote on the attack
4 sourcesCisco Talos reports a Go binary that queries Gemini, DeepSeek, Qwen, and Mistral to decide post-compromise actions.
DeepSeek training a 2T model, eyes an 8T successor
7 sourcesAn X post claims the firm is scaling up from 552B parameters to a 2T system, potentially reaching 8T via MoE architecture.
Needle 3 brings DeepSeek v4 Flash automation performance to 29MB
6 sourcesDesigned for tool calling, this open-weights model matches DeepSeek v4 Flash while running entirely on-device.
DeepSeek V4.1 Flash beats GPT-6 Astra in new intelligence index
2 sourcesSurpassing GPT-6 Astra, the 552B parameter MoE model earned 40 points in reasoning via Artificial Analysis's latest benchmark.
DeepSeek replaces Pro model with V4.1 Flash
3 sourcesDeepSeek officially launches V4.1 Flash, replacing the Pro model with superior performance and lower costs.
DeepSeek V4.1 Flash Beta with Native Multimodal Architecture
2 sourcesDeepSeek opens internal beta for V4.1 Flash, featuring a native multimodal architecture and faster speed.
DeepSeek builds 160,000 Huawei chip cluster for AI
2 sourcesDeepSeek plans a massive 160,000 Huawei Ascend 950DT chip cluster in Inner Mongolia for AI inference.
A 27B local model matches trillion-parameter giants
6 sourcesStartLux's 27B model trailed the 1.6T-parameter DeepSeek-V4-Pro by only 1.3% in CAICT's agentic MCP specialized test.
DeepSeek-V4-Flash-Vision regains sight in llama.cpp
3 sourcesllama.cpp adds vision support for DeepSeek-V4-Flash-Vision-Exp, enabling local multimodal inference via GGUF.
DeepSeek and GLM-5.3 outperform closed models in cyber discovery
1 sourceAikido's benchmark on 32 real vulnerabilities confirms DeepSeek V4 Pro and GLM-5.3 outperform closed frontier models in cyber discovery, at lower cost with higher reliability via pooling.
Info
AI-Generated Content Notice: The reports, articles, headlines, and images published in this section are produced with the aid of artificial intelligence tools and are subject to human supervision. Despite reviews and checks, content may contain errors, omissions, or missing updates. Materials are for informational purposes only and do not replace official sources or professional advice.
Follow our channels