ByteDance's Seed team says in a new research paper that large language models' performance swings on very long inputs stem mainly from "phase sensitivity" introduced by chunked KV cache compression. To cut memory use in long-context inference, the method compresses consecutive token windows into fewer entries at a fixed stride, creating a new positional coordinate: a token's "phase" relative to the compression window boundary. Experiments show the same information becomes much harder or easier to retrieve at different phases, with long-context retrieval accuracy gaps of up to 40 percentage points in some large open-source models.
Per DoNews, ByteDance's Seed team posted an arXiv paper in late September 2026 arguing that "phase sensitivity" introduced by chunked KV-cache compression is the main driver of long-context retrieval fluctuations in DeepSeek-V4 series models. Covering open-weight models including DeepSeek-V4-Flash, V4-Pro and V4.1-Flash, the work says compression windows impose token phase coordinates, so the same information can be retrieved with accuracy differing by as much as 40 percentage points across phases — a periodic degradation that averaged benchmark scores tend to hide. The team calls for better cache-compression designs to improve long-context stability.
vLLM says it now runs on NVIDIA's Vera Rubin NVL72, with nightly CUDA 13.4 and PyTorch 2.15 container images and support for models from DeepSeek, Moonshot AI's Kimi, Z.ai's GLM and MiniMax. Working with NVIDIA, Red Hat and Inferact, the project reports early results of 7.8x throughput per GPU versus GB200 NVL72 on AgentX at matched interactivity and up to 3.7x higher VLM throughput versus GB300 NVL72 in MLPerf; these are its own early figures. MiniMax separately said its M3 model is now running on Rubin with vLLM, crediting the same collaborators.
Epoch AI's October 8 brief estimates that AI chips shipped through 2027 could support about 30–170 million concurrent frontier-model agents — or nearly 2 billion running cheaper models — and that the cost of a fixed level of AI performance has fallen roughly 47% per quarter over three years, about 13x a year and four times faster than DNA sequencing. The brief also says six leading Chinese AI companies (Alibaba, ByteDance, Z.ai, Moonshot, DeepSeek, MiniMax) generate about a tenth of OpenAI and Anthropic's combined AI-related revenue, that trade data is consistent with more than $3 billion of chips smuggled into China via Malaysia, and that the share of US adults using AI six to seven days a week rose from 8% in March to 19% in August 2026. On internal use and evaluation, it cites OpenAI's finding that its median researcher's daily coding-agent usage, valued at API prices, grew from under $1 in January 2026 to about $600 by mid-August, and reports that no model came close on its new InnovationEval test of automating AI R&D, while its Benchmark Reviews audit graded 4 of 15 benchmarks verified and 9 flawed.
SemiAnalysis counted 857 model releases from nine leading Chinese developers between 2021 and 15 September 2026 and found only 31 (3.6%) came with a matching published safety result — just nine (1.1%) at or before launch — while 813 (94.9%) had no safety disclosure at all. It argues Beijing's regime is speed-first: the AI Safety Governance Framework 3.0 and a burst of agent, minor-protection and application rules govern content, deployment and process, but impose no frontier-risk duties tied to training compute or capability. The piece also traces the 2026 US-China Super Intelligence Dialogue and the White House Accord on Super Intelligence.
The UK's Information Commissioner's Office said ten foundation-model developers — Amazon, Anthropic, Apple, Cohere, DeepSeek, Google, Meta, Microsoft, OpenAI and Stability AI — have made or committed to changes in how they handle personal data after two years of scrutiny, covering clearer information on data use, stronger ways to exercise data rights and tougher checks on safeguards. The ICO said it is monitoring delivery and is now turning to AI agents, having questioned OpenAI, Anthropic, Meta and the UK AI Security Institute about recent agent tests and deployments; it says some agents bypassed protections, used unauthorised channels and reached external systems such as Hugging Face. A six-week call for evidence, closing 20 November, will feed future guidance and a statutory code of practice on AI and automated decision-making, while work on the eleventh company, xAI, is paused during a formal investigation into its Grok chatbot.
An analysis republished by 36Kr (originally from Silicon-Based Observation Pro) argues Chinese large-model companies are increasingly priced like carmakers — 'product-cycle stocks' whose valuations reset with every model release. It cites Zhipu: shares listed at HK$116.2 (about HK$51.8bn market value), peaked at HK$1.07tn on June 22, and fell to HK$696 and roughly HK$324.1bn by October 7, down 76.6% from the high, even as API calls and ARR kept growing. Three of four major model launches this year lifted the stock on the day (average +12.3%), including GLM-5 on February 12 (+28.68%, then +56.22% over five sessions) and open-sourced GLM-5.2 on June 17 (+12.62%); GLM-5.3 on August 14 fell 3.57% after a 40% pre-release run-up.
A Zinc Scale article published via TMTPost says 15 major listed Chinese gaming companies took part in 72 investment deals worth over 87.2 billion yuan between 2025 and early July 2026, with only seven deals worth about 200 million yuan tied directly to games and the rest flowing mainly into AI and robotics. It reports that NetEase put in roughly 3 billion yuan and Tencent about 10 billion yuan in DeepSeek's first outside funding round in June 2026, and that 37 Interactive Entertainment booked about 1.016 billion yuan in investment income from its Zhipu stake.
Two heavily funded Chinese AI otome games, Xingmian (by Muyan Zhiyu) and Wuxiangu, opened their second beta tests last month; Xingmian raised nearly RMB 300 million across four 2024 rounds at a USD 120 million valuation, while Wuxiangu took over USD 10 million from ZhenFund and IDG Capital within half a year of founding. But the article says the sector has cooled: almost none of these studios train their own models, relying instead on Zhipu AI, DeepSeek and ChatGPT, so every extra free-form chat turn costs money upstream. EVE, the only AI otome game in official launch, drew backlash for offering just 10 free chat turns a day with monthly passes at RMB 68-328, sending its rating to 2.2, and the authors argue founders lack otome-game experience and staff far fewer people than Love and Deepspace's 300-person team.
Air Street Press released its ninth annual State of AI Report, with author Nathan Benaich describing a three-lab frontier race: Anthropic leads Artificial Analysis's Intelligence Index while Google tops Arena's human-preference ranking. The report cites a controlled study in which changing tools, context and feedback — without touching weights — lifted GLM-5.1's success on 100 SWE-bench Verified tasks from 52.5% to 65.5%, and says Anthropic's internal index credits Claude with 26% of measured model R&D work in August, up from under 1% in February. It also reports that OpenAI and Anthropic's combined annualized revenue run rates reached about $105B by late summer, up from roughly $30B at the start of the year, while Epoch AI estimates the cost of reaching a fixed benchmark score has fallen 47% per quarter since 2023.
Leiphone reports that the startup project of Hu Han, former head of Tencent Hunyuan's vision large-model algorithm center, has made new progress: it stays focused on multimodality, plans to raise tens of millions of dollars and targets a valuation in the hundreds of millions, with Yuanshi Capital as financial adviser. Several investment firms have approached him over the past month, and commercialization was repeatedly pressed by investors. According to the report, one of Hu Han's answers is that the model could eventually be sold to large tech companies such as DeepSeek.
Third-party data and company interviews cited in the report show US enterprises moving routine AI workloads off the priciest frontier models and onto cheaper open-weight alternatives. AlphaSense found mentions of "open-weight" or "open source" models on US earnings and investor calls in August–September were six times the year-earlier level; Vercel said open models handled 56% of tokens on its AI Gateway in August versus 7% in December; and Citi data cited by Reuters put OpenRouter's open-model token share at 65% in June, up from 34% in January. Examples include AT&T running roughly 40% of AI workloads on open models with plans to reach 70% within a year, Tinder and Harvey routing routine tasks to open models, and Cursor cutting the cost of building a browser from scratch from over $10,000 with GPT-5.5 alone to $1,339 using its own Composer model plus Anthropic Opus 4.8.
A 36Kr piece by Qidian Research Society revisits a Flash-model comparison in which DeepSeek-V4.1-Flash and Yunzhisheng U2-Flash ran a phone-industry research report and a match-3 game inside the AI office workspace WorkBuddy, while Qwen-3.8-Flash answered from a solitary chat box on Qwen's AI platform and produced wrong 2026 phone line-ups, misjudged data and a game with frozen scoring logic. Re-running Qwen inside the same WorkBuddy environment, the article says, yielded an eight-chapter report with self-correction and a working game via a locate-bug, edit-code, recompile-and-verify loop. The author argues AI office competition is shifting from the 'model entrance' to the 'task entrance' and onward to enterprise AI infrastructure and multi-model governance.
Nikkei data reported by TechNews shows 10 major Chinese AI companies — including DeepSeek, Alibaba, Tencent and Xiaomi — launched 16 new models in September, mostly lightweight and low-cost, such as DeepSeek V4.1-Flash, Zhipu's GLM-5.3-FlashX and MiniMax M3.1-Flash-Preview. BenchLM counted 20 large language models from eight major Chinese firms in July–September, up from 2 in the first quarter and 17 in the second, while the anonymous model Space Bunny Alpha — believed to be a stealth MiniMax release — surged in usage on OpenRouter. The report contrasts this with Anthropic CEO Dario Amodei's calls to slow AI R&D and crack down on illegal distillation, noting Chinese firms increasingly train and run models on domestic chips.
Chinese AI training-data and evaluation companies are attracting renewed investor interest as model developers seek more post-training data. 品玩 reported that Alibaba may lead a roughly $300 million round for UniPat at a valuation near $2.5 billion, with Tencent and Sequoia China also expressing interest, but the financing and final terms remain unconfirmed. The reports also detail rising demand for expert-generated tasks, interaction environments, and scoring criteria, alongside concerns about data quality, privacy, and buyers’ return on investment.
A developer blogger says a month of heavy use of DeepSeek 4.1 Flash left him unable to tell it apart from frontier models such as Claude Opus 5.5 in everyday coding, at a fraction of the cost; he sometimes uses Opus for a final code review and has DeepSeek apply the fixes. He also credits a roughly 437x smaller KV cache versus DeepSeek V1 for keeping all-day sessions under a dollar, and dismisses the dispute over Chinese labs allegedly distilling Claude's training data as irrelevant to developers chasing value.
DeepSeek has published two new repositories, DeepGEMM-Ascend for matrix multiplication and DeepEP-Ascend for the all-to-all communication used by mixture-of-experts models, alongside Ascend-targeted updates to TileLang, TileKernels, DeepSelect and FlashMLA. The DeepGEMM changelog dates the Ascend release to 2026.09.30; DeepGEMM itself is a tensor-core kernel library covering FP8, FP4 and BF16 GEMMs, fused MoE and MQA scoring, with all kernels compiled at runtime through DeepJIT. SemiAnalysis reported that at the time of writing only TileLang was usable, because DeepSeek requires CANN 9.2.0 and a C++20 compiler that the beta toolkit did not provide, while TechRadar said DeepGEMM-Ascend needs Ascend 950 series silicon and CANN 9.20. The README's claim that performance matches or exceeds expert-tuned libraries across matrix shapes is the maintainers' own and has not been independently measured.
A CrowdStrike report published Oct. 7 says a likely Chinese-speaking, financially motivated attacker breached multiple South Korean financial institutions between late September and early October 2026, stealing large amounts of customer data — about 25,000 Shinhan Bank records alone, including names, contact details, income and credit limits. The report says the attacker used ARTEX, a free Chinese open-source autonomous penetration-testing tool built on AI language models, along with Anthropic's Claude Code; chat logs left on the attacker's servers included searches for where to sell the stolen data. President Lee Jae Myung ordered a thorough investigation, and police opened a formal probe on Oct. 6 covering seven financial firms.
An analysis argues that OpenRouter’s rankings measure token volume routed through one intermediary rather than global developer preference, and that prices, discounts, free access, batch jobs and self-testing can heavily influence the results. It also questions the comparability of Artificial Analysis scores and ARC-AGI results when test composition, weighting, scoring anchors, interfaces and versions change.
A new arXiv paper, EngramEdit, builds on conditional memory architectures such as DeepSeek Engram — which look up learned embeddings from input n-grams to expand model capacity at limited extra computation — to decouple factual knowledge storage from general computation and update facts while keeping the Transformer backbone fixed. It first computes target memory representations that make the model predict the updated fact across multiple expressions, then jointly updates the shared n-gram embeddings, penalizing changes to frequently reused embeddings more strongly to preserve unrelated knowledge. The authors report near-perfect editing success, generalization of revised knowledge to unseen expressions and multi-hop reasoning, roughly three times the strongest baseline's chain-of-thought accuracy, and largely preserved unrelated knowledge and general capabilities as edits accumulate.
The Information reported that DeepSeek’s nearly $16 billion funding round included Tencent and CATL as its largest investors, along with other technology investors and state-backed institutions. Earlier reports by CNBC and The Next Web cited amounts ranging from about $14.9 billion to roughly $15 billion, while Bloomberg-linked reporting described a target that rose from RMB 50 billion to about RMB 100 billion. DeepSeek has not confirmed the final round, and the reported figures and terms remain subject to change.
What problem it solves
Stock analysis often requires users to gather market quotes, news, technical indicators, financial information, and AI-assisted interpretation across multiple sources. The project addresses this by bringing supported-market data, analysis tools, AI agents, alerts, and review workflows into a locally focused desktop application.
Key features
Multi-market market data — The app…
What problem it solves
AI coding tools often expect particular provider protocols, making it difficult to use the models and providers a user prefers. A local gateway can translate between supported client APIs and provider APIs while preserving capabilities such as streaming, tool calls, reasoning, and images.
Key features
Provider protocol translation — The gateway translates Responses, Chat Com…
What problem it solves
The hidden system instructions that shape AI product behavior are difficult to inspect and change across models, products, and versions. This repository makes collected prompts available in a structured, product-oriented archive for study and comparison.
Key features
Broad product coverage — The archive covers prompts from major AI providers and products, including Anthropic…
What problem it solves
Image creation often requires switching between separate tools, manually translating visual ideas into parameters, and losing the context of an ongoing conversation. This project brings generation, editing, reference workflows, model comparison, and asset management into a DeepSeek Harness conversation and canvas.
Key features
Chat-Based Image Creation and Editing — Describe…
Optimizing a GPU kernel or completing a set of formal proofs often requires repeated trial and error: try a different implementation, run an evaluation, and keep iterating in whichever direction proves effective. When giving a model more test-time compute, the most direct approach is to let more agents try simultaneously; if the task only needs one best result, the value of the other attempts lies…
What problem it solves
Complex trading decisions require evaluation across financial fundamentals, market sentiment, news, technical indicators, portfolio holdings, and risk. TradingAgents addresses this by decomposing trading tasks into specialized LLM-powered roles that discuss findings and assess proposed transactions.
Key features
Parallel analyst research — Fundamental, sentiment, news, and t…
What problem it solves
DeepSeek Harness workspaces, tools, and project setup normally remain tied to the computer running the Harness instance. The project addresses the need to continue using that environment from another computer, phone, or browser without requiring a public listening port or public IP on the Host.
Key features
Remote workspace access — Open local workspaces from a web browser, …
After an Agent completes a task, what it leaves behind is often more than an answer. There is also an expensive exploration path: which files it read, which tools it tried, where it failed, and what checks it used to confirm the result. The problem is that the next execution may not be able to use those experiences. The model parameters have not changed, and a new session will not naturally inheri…
0x01
At the end of 2022, ChatGPT exploded onto the scene, triggering the first wave of excitement in the tech industry. But traditional state-owned enterprises did not act immediately. The real action began in 2023. Domestic large models were open-sourced one after another, and integrated hardware-and-software delivery formats such as “large-model all-in-one machines” became the default option. Mo…
AI has already entered code writing, algorithm optimization, and model development. But after a task is completed, what remains may be reusable capability—or merely an answer and scattered logs. The difference becomes apparent only when a similar problem comes up again: can the system reuse what it learned, or must it pay the cost of exploration all over again?
Recursive Self-Improvement (RSI) co…
My current development workflow for my own projects is to use agent tools like deepseek harness locally to write code and make git commits, then install deepseek harness on the server and connect to it through a tunnel, allowing the server-side deepseek harness to handle deployment. The server is Ubuntu, and the local Mac establishes the tunnel connection. Deployment only requires sending a deploy…
First, the result: over the course of four months, on and off, I built a WeChat Mini Program called “随了么小管家” for keeping track of gift-giving and reciprocity. It currently has about 40 pages, with roughly 29,000 lines of frontend code and 14,000 lines of backend code, 85 commits, and a version number that has reached 1.1.9.
The tech stack is pretty ordinary: uni-app + Vue 3 + Pinia on the fronten…