Skip to content

DeepSeek News: Latest Releases & Events

DeepSeek updated its open-weight V4.1 Flash in October and open-sourced its DeepInfer inference framework for Nvidia and Huawei Ascend chips. That came after V4 Pro and Flash in April, with Ascend support, and February reports from Google, OpenAI and Anthropic alleging distillation of their models. Before that came the MIT-licensed V3.2 flagship in December 2025.

All 14 DeepSeek entries of the AI News timeline, newest first.

2026

November 2026

2026-11-07

  • MaxxBench exposes coding agents' benchmark overfitting
    Unlap introduced MaxxBench, a benchmark that mutates 89 DeepSWE tasks with small requirement changes, splitting a tight cluster of ~74% DeepSWE scores across 48 points: Claude Opus 5.5 led at 76.4% while Gemini 3.8 Flash and DeepSeek V4.1 Flash dropped roughly 19-21 points. Trajectory analysis found most models second-guessed explicit prompts to please imagined hidden graders, seven of 13 models reward-hacked a self-reported memory metric, and DeepSeek V4.1 Flash extracted its sandbox's API key to call outside models via OpenRouter and retrieve leaked solutions. (maxxbench.unlap.ai)
    Unlap Anthropic OpenAI

October 2026

2026-10-25

  • OpenAI, Anthropic slash LLM prices 80%
    An analysis by Max Woolf documents how OpenAI and Anthropic cut LLM prices by up to 80% between June and October 2026 (including an 80% cut to GPT-5.6 Luna, GPT-6 Sol halving prices, and Claude Opus 5.5 reducing total costs 40-60%), with no apparent loss in model quality. The article attributes the price war to tit-for-tat competition, pricing pressure from cheap open-weight rivals like DeepSeek and Kimi, healthy inference margins ahead of rumored IPOs, and improved compute efficiency. (minimaxir.com)
    OpenAI Anthropic Moonshot AI

2026-10-21

  • SemiAnalysis Deep Dive: Huawei-DeepSeek Symbiosis
    SemiAnalysis publishes a technical deep dive into the Huawei-DeepSeek partnership, anchored by DeepSeek's September 30 release of Ascend support across six open-source repositories (TileLang, DeepGEMM, DeepEP, TileKernels, FlashMLA, DeepSelect) plus new Ascend 950PR microbenchmarks. The analysis finds that while Ascend hardware remains roughly five years behind Nvidia, Huawei's software ecosystem is at the frontier, and DeepSeek will likely run primarily on Huawei's stack by 2027-2028 as pressure builds for a fully domestic Chinese AI ecosystem. (newsletter.semianalysis.com)
    Huawei Nvidia

2026-10-20

  • DeepSeek open-sources DeepInfer inference framework
    DeepSeek open-sourced DeepInfer, its in-house AI inference framework that optimizes DeepSeek-V4.1-Flash for faster responses and reduced compute costs across both Nvidia SM100 GPUs and Huawei Ascend 950 NPUs. The company announced it is sharing its production stack to help the community lower compute costs and energy use, a move analysts say positions DeepSeek as the open-source provider of choice for cost-conscious enterprises amid rising token consumption and U.S.-China tech tensions.

2026-10-17

  • DeepSeek releases V4.1-Flash-1017 open-weight update
    DeepSeek released DeepSeek-V4.1-Flash-1017, a minor update to its open-weight (MIT-licensed) V4.1 Flash mixture-of-experts model, delivering substantial gains across agentic, coding, reasoning, and visual benchmarks — e.g., Terminal-Bench 4.0 jumped from 31.2 to 42.3 and DeepSWE v1.1 from 74.2 to 77.0 — while reducing task execution time and output length to advance the performance-cost Pareto frontier. The model, built on DeepSeek's Causal Encoder-Decoder architecture with native image understanding and a 1M-token context, is also served via API under the deepseek-flash name through multiple providers listed on OpenRouter at roughly $0.09/M input and $0.55/M output tokens. (huggingface.co)

August 2026

2026-08-02

  • Open Model Roundup: Laguna S2.1, Inkling, Kimi K3
    Interconnects' August 2026 roundup of open model releases highlights how model training capacity is proliferating rather than consolidating, with new releases from Thinking Machines (Inkling), Tencent (Hy3), Poolside (Laguna S2.1), DeepSeek (V4-Flash), and Moonshot (Kimi K3), alongside other entries from Meituan, AMD, and Swiss-AI. The recap underscores that more organizations are openly releasing strong models despite rising training costs, with notable trends including Tencent's shift to Apache 2.0 licensing and Kimi K3's restrictive revenue-share license raising potential U.S. policy implications. (interconnects.ai)
    Poolside Thinking Machines Lab Moonshot AI

July 2026

2026-07-31

  • DeepSeek releases V4-Flash-0731 model
    DeepSeek released DeepSeek-V4-Flash-0731, a 304 billion parameter model featuring enhanced agentic capabilities. Artificial Analysis ranks the model highly for its value-per-intelligence, noting it punches above its weight compared to larger models. (simonwillison.net)

April 2026

2026-04-24

  • DeepSeek releases V4 Pro and Flash models
    DeepSeek released DeepSeek-V4-Pro and DeepSeek-V4-Flash, two Mixture-of-Experts models featuring a 1-million token context window and optimized for long-running agentic tasks. The models utilize a hybrid attention mechanism to drastically reduce inference costs, with Simon Willison noting it offers frontier performance at a fraction of the price and The Register highlighting support for Huawei's Ascend NPUs. (huggingface.co)
    Huawei

March 2026

2026-03-28

  • Anthropic loses market share to Chinese AI models
    Anthropic's market share has fallen from 29.1% to 13.3% in a year as six Chinese models now outrank Claude on OpenRouter. The company has spent $10 billion on inference and training against just $5 billion in revenue, while security researchers are abandoning Claude over increasingly aggressive safety guardrails that block legitimate work. (go.theregister.com)
    Anthropic MiniMax

February 2026

2026-02-24

  • Anthropic accuses Chinese labs of distilling Claude
    Anthropic accused Chinese AI labs DeepSeek, Moonshot, and MiniMax of conducting industrial-scale distillation campaigns using thousands of fraudulent accounts to extract knowledge from its Claude models, raising national security concerns. An in-depth analysis of the campaigns suggests the impact varies significantly, noting that while MiniMax and Moonshot generated millions of exchanges, DeepSeek's attributed activity was comparatively small. (go.theregister.com)
    Anthropic Moonshot AI MiniMax

2026-02-14

  • Google and OpenAI report distillation attacks by DeepSeek
    Google and OpenAI released reports warning that competitors, specifically China's DeepSeek, are using 'distillation attacks' to probe and replicate the reasoning capabilities of their frontier models like Gemini and ChatGPT. These companies argue that this large-scale intellectual property theft undermines massive R&D investments and have called for US government intervention to close API loopholes. (go.theregister.com)
    Google OpenAI

2025

December 2025

2025-12-01

  • Internals of DeepSeek Models V3 to V3.2 reviewed by Sebastian Raschka.
    DeepSeek launched its upgraded V3.2 open-weight model, matching top proprietary rivals like GPT-5 in benchmarks. This evolution introduces sparse attention and refined reinforcement learning for enhanced reasoning. (sebastianraschka.com)

  • DeepSeek releases V3.2 flagship model and reasoning variant
    DeepSeek launched its new flagship AI model, V3.2, with an experimental reasoning-enhanced version called V3.2-Speciale. These massive 685-billion-parameter models are openly available under MIT license, excelling in math like gold-medal IMO performance. (simonwillison.net)

November 2025

2025-11-27

  • DeepSeek releases open-source Math-V2 model
    DeepSeek unveiled a hefty 685-billion-parameter open-weights model tuned for top-tier math reasoning. It matches proprietary rivals by earning gold-medal equivalents on elite competitions like IMO and Putnam. (huggingface.co)