A German tech outlet benchmarked four Chinese model families — Kimi, Qwen, GLM and DeepSeek — against established Western offerings such as ChatGPT and Claude. The review examines whether their strong benchmark scores translate into comparable quality in everyday practical use. It concludes that these alternatives offer notable capability at a lower price, though with trade-offs.
DeepSeek has introduced v4.1-Flash, a large-scale model with 763B total parameters built on a new causal encoder-decoder design that also handles vision input. Commentary from Latent Space and Sebastian Raschka argues the release is significant enough that it should have been branded as a new major version rather than a point update. Details on training, availability and licensing were not included in the report.
A user report says the DeepSeek v4.1 Flash model can be loaded and run locally on a 2020 M1 Mac Mini with 16GB of unified memory, but generation is extremely slow at roughly 23 seconds per token. That pace makes the setup impractical for interactive use, though it shows the model can technically execute on older consumer hardware. The result highlights how limited RAM and memory bandwidth constrain local inference of large models.
A new arXiv paper proposes HISA, a hierarchical indexing approach for token-level sparse attention of the kind used in DeepSeek Sparse Attention. Such methods score every past key with a lightweight indexer before running attention on a selected subset, which becomes costly as context grows. HISA aims to make that key-selection step more efficient while keeping fine-grained selection quality.
DeepSeek has introduced V4.1-Flash, a multimodal model with 552 billion parameters, though only 16 billion are active per token. The release cuts KV cache memory requirements to a quarter of the previous generation. On the DeepSWE coding benchmark, it slightly outperforms Opus 5 and GPT-5.6 Sol.
DeepSeek has released V4.1-Flash, a multimodal model with 552 billion total parameters that activates only 16 billion per token. The company says it cuts KV-cache memory to roughly a quarter of what its previous model required, which matters for agent workloads. It slightly edges out Opus 5 and GPT-5.6 Sol on the DeepSWE coding benchmark.
DeepSeek has added a new model called V4.1 Flash to its Hugging Face repository, where the weights and model card are hosted. The listing drew attention on Hacker News, though the report gives no further detail on capabilities, size, or licensing. It appears to be a lighter or faster variant in the company's V4 series.
Three US agencies have named DeepSeek, Moonshot AI, Alibaba, MiniMax, StepFun and Z.AI as allegedly distilling American AI models at industrial scale since late 2024. The agencies also disputed DeepSeek's claim that its model was trained for about $5.6 million. They advised US providers to quietly reduce the quality of responses served to accounts flagged as suspicious.
US authorities say Chinese AI developers including DeepSeek and Alibaba are systematically pulling knowledge out of American models such as those built by OpenAI and Google. The agencies frame the activity as industrial espionage aimed at closing the capability gap at low cost. No details were given on which specific countermeasures might follow.
DeepSeek intends to launch a new model called V4.1 Flash on or near September 10, 2026, Beijing time. According to the company, internal and external testing showed the Flash variant outperforming its V4 Pro model on performance, cost, speed, and task completion time. No independent benchmarks or pricing details were included in the announcement.