Tom's Hardware reports that Nvidia's RTX Spark Superchip in the Surface Laptop Ultra outperforms Apple's M5 Max MacBook Pro by about 25% in agentic AI completion benchmarks using Qwen models. Nvidia CEO Jensen Huang announced the RTX Spark platform, which Nvidia describes as laying the foundation for reinventing the PC in the agentic AI era. The figure comes from a single report, and the supplied text does not describe the benchmark methodology or system configurations.
Zhidx reported on October 15 that Apple board executive chairman Tim Cook, speaking to outlets including Jiemian News and The Paper during a visit to Shanghai, confirmed that Apple Intelligence will launch a test version in China by the end of October and that the related Alibaba Qwen models have been approved. The first features include live translation, writing tools and a photo cleanup tool, but not Siri AI. The report also relayed Bloomberg reporter Mark Gurman's claim that the official version is expected in December alongside iOS 27.3 in China, still without the new Siri AI, Image Playground or the latest Genmoji, and with limited text-generation tools.
At a Houmo AI ecosystem partner conference on Oct. 14, CEO Wu Qiang said the company's M50 compute-in-memory edge chip has won more than 150 product design wins and serves over 100 customers less than a year after entering mass production, according to TMTPost. Wu also said the surge in AI agents has raised token consumption by two to three orders of magnitude. Other speakers, including Alibaba Cloud's Qwen smart-hardware lead Xu Chao, argued that shrinking models and raising intelligence density are key to moving compute from cloud to device, while the report notes edge AI still faces thin use cases, deployment barriers and unclear business models.
A 市值榜 article republished by TMTPost and 36Kr says agent competition is shifting from single-task execution to continuous personal service: after Meta launched Muse in September, ByteDance was reported in late September to be accelerating its personal agent project Spell (named “小豆”, planned as a standalone app, with the final form undecided), Tencent's Yuanbao was reported to plan a similar product, Ant Group's Lingguang began small-scale testing of a new version on October 11, and Qwen had earlier announced at its Yunqi conference that it is accelerating a Personal Agent. The article cites The Information as saying that about three weeks after launch Muse had more than 3 million users sending at least one instruction a week and over 1 million daily, and notes Amazon began blocking Muse from its shopping site on September 20, while on September 25 an outside security researcher reported a flaw that could expose emails and files in users' dedicated cloud VMs. It also cites The Wall Street Journal and Sensor Tower data saying Muse's global downloads approached 9 million but growth was slowing, with 22.5% 14-day retention versus ChatGPT's 26.8%.
Transsion listed in Hong Kong on October 15 at HK$35.30 per share, below the bottom of its offering range and at a roughly 37% discount to its A shares. Its shares closed 5.72% below the issue price on debut. Reports by TMTPost and Chyeyebang examine how rising memory costs, partly linked to AI-driven demand for upstream capacity, are squeezing Transsion’s thin-margin, high-volume handset model. About 40% of net IPO proceeds are earmarked for AI research and development. The reports question whether AI can support premium pricing in price-sensitive markets or will mainly improve efficiency and retention. A separate Chyeyebang commentary said the H shares subsequently recovered and were admitted to Stock Connect.
Tencent Hunyuan's Frontier team and university researchers released Terminal-RSI, a training framework that alternates Recursive Self-Rewrite of verified trajectories with critic-inheriting Recursive Policy-Refine. Starting from Qwen3.8-27B, the five-stage process produced Terminal-RSI-27B, which the authors report scored 19.84% on Terminal-Bench 4.0, 86.52% on Terminal-Bench 2.1, 16.95% on Terminal-Bench 3.0 and 79.00% on Terminal-Bench Hard. Hugging Face · Papers describes the paper and release, while 钛媒体 and 智东西 provide additional reporting on its methods and limitations.
Multiple China-market iPhones received updated regulatory filings on September 29 carrying three large-model registration numbers: Apple Intelligence (Shanghai-AppleZhiNeng-202506160057), Tongyi Wanxiang (Shanghai-TongYiWanXiang-202410090024) and Tongyi Qianwen (ZheJiang-TongYiQianWen-20230901), according to IT之家 and cnBeta. The filings cover iPhone 15 Pro / Pro Max and later models, plus iPhone Air and the foldable iPhone Duo — matching the device list that supports Apple Intelligence. Apple Technology Development (Shanghai)'s Apple Intelligence model was registered on July 8, 2026 for use on Apple phones, and Alibaba has said Qwen will be integrated into Apple Intelligence to bring text and image understanding and content generation to Chinese users across iOS, iPadOS, macOS and visionOS. Bloomberg reports Apple plans to launch Apple Intelligence in mainland China in December via iOS 27.3, without Siri AI features.
Qwen's official X account congratulated LightSeek Foundation after its TokenSpeed inference engine reached 685 output tokens per second per user on Qwen3.8-Flash-Next. The run used 4× NVIDIA GB300 GPUs with NVFP4, and LightSeek published a blog post on the kernel and serving optimizations behind the result. Qwen said it is excited to see the community unlock more of Qwen's inference potential; the figure comes from LightSeek's own benchmark and has not been independently verified.
Alibaba's Qwen opened limited invitation testing for its Qwen Go personal assistant in the Qwen app on October 13, according to 36Kr and TMTPost. The reports place the launch in a wider personal-agent race: ByteDance is testing a project code-named Spell for Doubao, Tencent has TX Handy Bot and Yuanbao assistant projects, while Meta and OpenAI have Muse and dots; Meta also released Muse Wonder models with Active Memory and Active Focus on October 12. Qwen's team detailed QwenGyre, an online reinforcement-learning framework for long-horizon agent tasks that it says raised Qwen 3.8 2.4T's NL2RepoBench pass rate from about 52.5% to 58.5% and cut training time by about 44% versus an Async baseline, while TMTPost analyses question how agent execution costs can be recovered.
A TMTPost piece in Alibaba Cloud's co-produced Yunqi Strategy Reference describes how Dazhihui and Alibaba Cloud launched an AI-native overhaul in October 2024 to turn the two-decade-old stock-quote software into an assistant that understands intent, breaks tasks into steps and supports investment decisions, rather than just adding an AI chat box. The article says the teams tested and rejected feeding financial data to a model through RAG or web search alone, and instead started with data governance to unify metric definitions, then built task orchestration, a Skill tool layer and a security sandbox on Alibaba Cloud's FinXScope 2.0 agent platform, so the model only plans while factual data is returned by controlled tools. It also describes a four-layer memory system (profile, behaviour, interaction history and short-term context) and a human-plus-model evaluation process; the article says intent recognition fell from about 10 seconds to milliseconds by shrinking the model and training intent classification into a 4B model, but these figures are the project teams' own account with no independent measurement.
A roundup republished by Huxiu AI and 36Kr reports that Tencent Yuanqi issued a shutdown notice on Oct. 8, saying the zero-code agent platform will stop all service on Nov. 9 and that agents built by users will cease to work, with Tencent directing users to its Tencent Cloud agent development platform. Many WeChat public-account operators had used the platform to turn past articles into knowledge bases and run automated replies, the article says, and now need to migrate. The piece also lists other closures, including Tencent's QClaw (shutting down Dec. 24), Doubao and Qwen removing user-built agents, and OpenAI retiring Sora and the ChatGPT Atlas browser, and groups the causes into five patterns.
Alibaba's Qwen began selling the Qwen Clip AI clip-on earbuds on JD on October 13 at 1,099 yuan after subsidies and limited-time discounts, according to IT之家 and 36Kr; Qwen says Bose tuned the open-ear design, which offers 9 hours of battery life, 40 hours with the charging case, LHDC 5.0 and Hi-Res Wireless support. Qwen also claims the earbuds support voice control, real-time recording transcription, 92-language transcription notes and AI translation across 89 languages. From October 15, the earbuds can be bound to the Qwen App for voice Q&A and for assigning research and other tasks to a background AI assistant that alerts users when the work is done, 36Kr reported.
Axera officially launched its AX8860 series AI chip at the 2026 China International Industry Fair, positioning it as a compute base for running large models at the edge and on-device. The 6nm chip pairs a 16-core Arm Cortex-A55 CPU with a self-developed NPU rated at 256 TOPS mixed precision (native FP4, no sparsity), supports up to 128GB of LPDDR5X memory with 273 GB/s bandwidth on a 256-bit bus, and offers PCIe 5.0 connectivity. Product forms include a 40W module, a single-slot Gen5x4 PCIe card, a single-slot four-chip card, and planned Mini PC and pocket PC designs; clustering over a 256 Gbps interface is claimed at 512 TOPS for two chips and 1024 TOPS for four at about 80% scaling efficiency. Axera and Sipeed claim the AX8860 runs Qwen 27B with 25% faster prefill and 30% faster decode than NVIDIA's DGX Spark, at system pricing "close to half" of DGX Spark, though no independent testing of these performance or price figures was provided. The chip supports vLLM, SGLang and PyTorch, ONNX and Hugging Face model formats, and an OpenAI-compatible API.
QuestMobile's report puts China's AI-native app user base at 537 million as of August 2026, with 85.1 uses per user per month, up 67.1% year over year. It says users' first entry point for finding information is shifting from search boxes and news feeds to AI chat boxes, while comprehensive news apps and search engine apps saw monthly per-user usage of 123.6 and 35.8 times, down 7.6% and 21.7% respectively. By active penetration in August, the top five were Doubao at 77.2%, Qwen at 29.4%, DeepSeek at 22.8%, Yuanbao at 8.2% and Ant Afu at 5.8%, with competition described as extending from model capability toward ecosystems and scenarios.
A Databricks blog post describes how Flo Health, a women's health app with 81 million monthly active users, built safer and more useful custom open models on Databricks AI Runtime. Flo first created an evaluation system of 26 AI judges (12 medical-safety, 6 micro-behavior, 8 usefulness) to turn medical standards into a training signal, generated more than 15,000 privacy-first synthetic questions, then ran supervised fine-tuning and DPO on 32 NVIDIA H100 GPUs using LLM Foundry and Compose-RL with MLflow tracking. Flo reports medical safety rising from 34% to about 61%, usefulness from 42% to 76%, and the share of responses meeting both from 16% to 47%, and estimates tailored models can cut inference costs by an order of magnitude versus frontier APIs; it plans to apply the pipeline to newer open-weight models such as Gemma, Qwen and Kimi.
The vLLM community project vLLM Semantic Router released and open-sourced the Decision 3.0 family of multimodal decision models: d3, d3-flash, d3-mini, d3-nano, d3-lite and d3-edge, spanning roughly 0.6B to 27B parameters, with weights published on Hugging Face. The flagship d3 is built on Qwen3.8-27B and accepts multiple images and video, while the other five are based on Qwen3.5 at different sizes; unlike Jev, which is served mainly via cloud API, the series is open-weight and locally deployable. According to team-provided figures, on the Jev Decision Index 0.3.1 leaderboard d3 ranks first with a text decision index of 64.7 and a vision score of 71.6, ahead of Jev's 60.1 and Microsoft Decision-1's 53.2 (9B, fine-tuned from Qwen3.5-9B); the team also reports median latency on a single AMD Instinct MI325X GPU of 55 ms for text, 273 ms with images and 553 ms for 10-second video, while one developer ran d3-lite via Core ML on an M5 Pro Mac to classify and route 1,000 pull requests in about two minutes.
In September 2026, Meta, OpenAI, Manus and others launched personal agent products within weeks: Meta's Muse passed 2.5 million downloads in 13 days, Manus's Cue gives each agent its own email, phone number, digital wallet and cloud computer, and OpenAI's Dots runs in isolated cloud environments with access to more than 4,000 apps through ChatGPT plugins. In China, Alibaba said at its Yunqi conference it is accelerating a Qwen Personal Agent, ByteDance is reported to be merging its Spell project more deeply into Doubao, Tencent's Yuanbao is rumored to get a personal agent, and Tencent's lightweight cloud team already runs LightVela. The article argues this wave differs from earlier ones because agents shift from chat tools to executors, with ecosystem depth and monetization deciding the winners.
A local-AI wave is growing as small developers and hardware hobbyists run high-end GPU machines in basements, attics and home offices for cost, privacy and convenience, cnBeta reports. Former Microsoft employee and repeat founder Gary Fleck built a machine with discounted parts for under $6,000 to train custom models on terabytes of EEG data, saying it beats paying monthly AWS or Azure bills — though cooling his basement required $10,000 to $20,000 in electrical work and HVAC. Local agent platforms such as OpenClaw are helping drive the trend, Mac mini and Mac Studio are popular with agent developers, and Dell says the home-AI hobbyist market is booming. Crazy Egg CEO Hiten Shah runs 18 devices including 14 Mac minis and Mac Studios, with his home office often hitting 85°F (29.4°C), while one YouTube creator runs Qwen-3.8-27B locally on five AMD GPUs in his garage. Some enthusiasts say they want to keep compute in their own hands in case governments restrict AI access.
NextFin News, published via TMTPost, reports that Ulanqab in northern China has become one of the country's densest data-center clusters, with Alibaba, Huawei, Kuaishou and specialist operators taking space because of relatively cheap electricity, available land and a climate that eases cooling loads; one abandoned school is being converted into a computing facility. The building boom coincides with falling model prices: DeepSeek made a 75% cut to its V4-Pro API rates permanent in May, and a Token price index compiled by Silicon Data and cited by the Financial Times showed composite prices more than 52% below their May peak by October, though not every model fell equally. Compute sellers are still spending more — Alibaba's capital expenditure for the quarter ended June 30 reached 67.678 billion yuan, up 75% year on year, while AI cloud and computing services revenue rose 45% to 48.437 billion yuan — and the article's hypothetical example of a 10-megawatt IT load at PUE 1.2 implies a roughly 42 million yuan annual electricity-cost gap for every 0.40 yuan per kWh difference, over 200 million yuan cumulatively across five years.
Ghost emerged from stealth on Oct. 5 with an $11 million seed round led by Andreessen Horowitz, with Abstract, Audacious Ventures, SV Angel and Nova participating; the valuation was not disclosed. The startup also unveiled Core, a $3,499 screenless computer built to run AI agents continuously, controlled through a phone app or voice, with NVIDIA RTX PRO 4000 Blackwell SFF graphics, 64GB of DDR5 memory, a 1TB SSD, an AMD Ryzen 5 7600 (reported later upgraded to a 9600X) and preinstalled Qwen and Gemma models. Reports say the first preorder batch is sold out with shipping expected in the last week of October, while Ghost's claims that models, memory and personal data stay local, that there is no metered AI subscription and that a forthcoming agent firewall will police outbound requests remain vendor claims; one report notes search and messaging still send necessary information to external services.
The Jetta M6, FAW-Volkswagen's Jetta brand's first new-energy strategic model, will go on sale on October 22 after presales opened September 28 starting at 80,800 yuan (465 Pro at 80,800; 555 Max at 90,800; 555 Ultra at 99,800). The electric sedan measures 4806/1895/1500 mm with a 2820 mm wheelbase and offers 465 km and 555 km range versions. It uses Horizon Robotics' Journey 6M chip for urban and highway NOA plus cross-floor memory parking, and integrates the Qwen, DeepSeek and Doubao models, along with lane-level navigation, phone-car connectivity and a standard heat-pump air conditioner.
An analysis published by TMTPost says Meta's personal agent Muse, launched on Sept 8, 2026, topped the U.S. App Store free chart within two weeks, citing about 2.8 million downloads in the U.S. and Canada over 12 days, roughly 1.8 million iOS downloads in the same period, about 642,000 U.S. mobile daily active users, and a more than 20% rise in Meta's share price (about $200 billion in market value). Zuckerberg described Muse as a personal agent that understands a user's goals and works around the clock to complete tasks, calling it a first step toward "personal superintelligence"; the article also cites 100 million free tokens a week, a cloud-based Muse Secure VM, and a plan to earn commissions on transactions rather than rely on subscriptions. It relays the follow-on moves: OpenAI's always-on agent Dots built on GPT-6 Astra with 4,000-plus app connections (Sept 30), Manus's Cue (Sept 28), ByteDance's "Spell" personal-assistant project, Tencent's LightVela and Yuanbao, and Alibaba's Qwen Personal Agent — while Amazon blocked Muse's purchasing feature, citing unauthorized AI agent access.
An analysis published by TMTPost argues the 18th Double 11 festival has been stretched into a shopping season of more than a month, with platforms cutting back coupon-stacking rules while linking subsidies to how much merchants spend on advertising. Alimama's Double 11 coupons are allocated by AI according to a merchant's ad spending, with qualifying brands receiving up to 55% back in performance-ad credits, while JD.com's "Steady Profit" scheme compensates for gaps against system-recommended ROI targets and returns up to 12% for higher ad spend. The piece also says Douyin added a Doubao section to its merchant dashboard, and that Taobao's Qwen shopping assistant and JD's smart shopping guide are competing for a new discovery entry point, citing JD figures that order costs fell 15%-27% for assets optimized to its product-data standard versus generic AI assets.
OpenRSI Index, launched by the OpenRSI Foundation, has drawn attention on X, where its announcement post has passed 133,000 views. The benchmark draws tasks from real open-source research projects such as Marin, Qwen and GPIC, handing agents existing models, code, data and training recipes to improve, spanning pretraining, post-training and image generation, with the site listing representative single runs at roughly 5,815 to 20,864 H100-hours. Evaluation uses an RSI-Harness that separates the agent's workspace from a clean judging environment, pausing and snapshotting the workspace for scoring; advisers include Luke Zettlemoyer, Karthik Narasimhan, Yejin Choi and Dawn Song, and a companion RSI-Anything pipeline lets researchers turn their own open problems into runnable RSI tasks.
A joint team from the National University of Singapore, Nanyang Technological University and GaugeForge released Quantum-Harbor, a simulated quantum engineering lab, plus QIQCBench, a set of 49 real-world tasks spanning measurement and characterization, control, error correction, compilation, simulation and sensing. Seventeen frontier AI agents were dropped into the physics-driven environment with instrument APIs and hidden hardware parameters, then graded under an "evidence-bound" verified-autonomy standard that checks whether conclusions are supported by data the agent actually collected. Pass rates ranged from 78% at the top down to 3%, 937 of 1,510 failures came from exhausting budget or looping, and the many-body physics category averaged only 15%.
Two Chinese outlets republished the same report describing how US enterprise AI budgets have split into three tiers under token-cost pressure: giants such as Microsoft and Meta are shifting to in-house models (MAI-Code-1-Flash, MetaCode, Muse Code) and cutting external model budgets; mid-size firms like Uber spent a full year's AI budget in under four months and are trimming seats or downgrading; and small firms and startups are turning to Chinese open-weight models. It cites OpenRouter data putting Chinese models at a 46.4% weekly share peak, with DeepSeek leading at 17.6%, and says OpenAI has cut prices twice (GPT-5.6 Luna, GPT-6 Sol) in response. The report also mentions Anthropic billing third-party agent frameworks separately in April and GitHub pausing new Copilot Pro/Pro+ signups.
On October 13, Alibaba's Qwen unveiled three Personal Agent hardware products — the AI glasses N1 and N1 Pro and its first AI ear-clip headphones — all going on sale the same day. The N1 series packs a 50-megapixel sensor with 4K video, live photos and always-on environmental awareness, while the N1 Pro adds eye-tracking interaction, iris-verified payments (co-developed with Ant Group's GPASS, with an official false-acceptance rate below one in a million) and monitoring of heart rate, blood oxygen and skin temperature. The headphones, tuned with Bose, offer real-time translation in 89 languages and AI note-taking; after subsidies and limited-time discounts, the glasses start at 2,218 yuan and the headphones at 1,099 yuan.
The open-source Unsloth project is trending on GitHub's Python list, and its README now describes three ways to use it: the native Unsloth Desktop app for Windows, macOS and Linux, the Unsloth Studio web UI, and code-based Unsloth Core. The project says it can run and train LLMs plus diffusion, embedding and audio models, naming Qwen3.8, DeepSeek-V4, Gemma 4, GLM-5.3-Flash, Kimi K3 and MiniMax-H3, and its Unsloth Start command wires local models into agents such as Claude Code and Codex. The README also claims 2x faster training with up to 70% less VRAM and no accuracy loss — figures that come from the project's own documentation.
An analysis republished by TMTPost (originally by 巨潮WAVE) argues that as China's ride-hailing market matures — roughly RMB 411.9 billion in 2025, up 6.07% year on year, with users flat at 539 million — system-level AI agents are replacing aggregator apps as the entry point for trips. It cites Cao Cao Mobility integrating AI ride-hailing into Doubao's phone assistant on a Nubia handset, Honor's YOYO, OPPO's Xiaobuo Next and Qwen AI glasses, so users can book without opening a travel app. The piece also describes Robotaxi operators (Apollo Go, Pony.ai, WeRide) and platforms (DiDi, Cao Cao, Ruqi, T3) moving toward mixed human-plus-driverless operation and eventual consolidation.
SUSE announced an "enterprise private AI" strategy and an AI Factory offering with NVIDIA that combines NVIDIA AI Enterprise with SUSE's infrastructure and security governance capabilities, for AI workloads on-premises, in multi-cloud and in offline environments. SUSE says the combination can cut deployment time from months to days, though it acknowledges actual timelines depend on existing systems, data readiness and integration complexity. The release also includes Liz, a Kubernetes operations agent integrated into Rancher Prime that supports Chinese-language interaction and troubleshooting under a human-in-the-loop design; the China version is adapted to models including Qwen, DeepSeek, Kimi, Zhipu and MiniMax. SUSE also floated a possible evolution toward a "Token Factory," described as a product direction rather than a delivered capability.
What problem it solves
Developers need abstractions that can accommodate increasingly agentic LLMs without limiting their reasoning and tool-use abilities through rigid prompts or opinionated orchestration. They also need a way to turn agent components into multi-tenant, multi-session applications with supporting infrastructure.
Key features
Composable agent building blocks — Compose agents from a…
What problem it solves
Developers using AI coding CLIs need a way to delegate work among multiple agents and coordinate their execution. CAO addresses this by running provider CLIs in isolated terminal sessions and giving a supervisor tools for coordinating workers.
Key features
Multi-agent delegation — A supervisor delegates work to specialist agents, with parallel and sequential execution suppor…
What problem it solves
Heretic addresses the difficulty of modifying language-model behavior to remove refusals, suppress slop, reduce sycophancy, or apply other behavioral changes. It provides a training-free approach that steers models through mathematical techniques rather than expensive retraining.
Key features
Configurable behavior changes — The default configuration suppresses refusals, whil…
What problem it solves
Running and training models locally requires tools for model execution, fine-tuning, agent integration, and deployment across different model formats and hardware. The project brings these workflows into a desktop app, a web UI, and a code-based version.
Key features
Run and train models — Run and train language, diffusion, decision, embedding, audio, image, and video models…
What problem it solves
Developers often need to understand why model applications are slow or costly, but system decisions require reasoning across model architecture, hardware limits, memory, computation, communication, and scheduling. Easy-to-miss factors such as KV cache, bandwidth limits, and inter-device communication can shift estimates by orders of magnitude, so system design should begin f…
What problem it solves
Researchers, red-teamers, and system tinkerers need a chat interface that can work with multiple hosted and local models while exposing evaluation, perturbation, and sampling controls. They also need clear visibility into data flows, storage, telemetry, and deployment boundaries so they can make informed choices about privacy and self-hosting.
Key features
Multi-model races …
What problem it solves
Machine learning workflows need a shared model definition so that models can work consistently across training frameworks, inference engines, and adjacent modeling libraries. Transformers centralizes these definitions and exposes a unified interface for loading pretrained models, running inference, and adapting model internals.
Key features
Broad Model Support — The framewor…
What problem it solves
Serving language models requires reading checkpoints, tokenizing requests, and running accelerated kernels while managing a local model server. TensorFold addresses these requirements with a native binary that can run on Apple Silicon and NVIDIA GPUs without requiring Python or MLX for serving.
Key features
Native Metal and CUDA inference — The Zig-based engine executes mode…
What problem it solves
Interactive multimodal and speech systems often require generation to span heterogeneous stages with different compute patterns, dependency structures, and resource needs. Serving these pipelines requires coordinated stage lifecycles, inter-stage transport, model integrations, and a unified API surface.
Key features
Coordinated multi-stage runtime — Models generation as coor…
What problem it solves
Many AI applications send messages to remote servers, where they may be logged and analyzed outside the user's view. This project addresses that concern by running language models, conversations, voices, and tools on the device without requiring an account or internet connection.
Key features
On-device language model chat — Run GGUF language models fully offline, including G…
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…
Handing a problem to four Agents most easily produces four attempts. Increasing the number to eight may raise the success rate further. But if every Agent searches along similar lines, a significant portion of the additional computational power brought by a larger team will be spent on duplicate work. If we instead require them to communicate continuously, messages will consume context and reasoni…
What problem it solves
A frozen LLM can only use information that fits in its context window, making it difficult to answer questions from large document collections or long interaction histories. Memorilla addresses this by reading a collection and supplying the LLM with a fixed number of query-dependent memory tokens.
Key features
Query-dependent memory compression — The system reads a complete …
What problem it solves
Voice input and voice output are usually handled by separate cloud services, which means audio and transcripts leave the machine and the two halves of the voice loop do not connect. Voicebox aims to cover both directions locally: cloning voices from short samples, generating speech, dictating into any text field, and letting MCP-aware agents speak back in a cloned voice.
Key…
What problem it solves
Developers need a coding agent that can fit into different workflows without requiring separate processes for each interface or model provider. They also need access to agentic features, automation, and shared sessions while retaining control over model and provider selection.
Key features
Agentic workflows — The agent includes Auto-Memory, Auto-Skills, SubAgents, Agent Team…
What problem it solves
Developers need to run AI models inside applications while handling model downloads, inference configuration, and chat sessions. The project presents local inference with data privacy and no inference costs as its deployment model.
Key features
High-level application API — The engine exposes a simple API for creating an engine, loading a model, opening a chat session, and re…
What problem it solves
Jan provides a desktop application for downloading and running LLMs with local control and privacy. It also connects to supported cloud models and exposes an OpenAI-compatible local API for use by other applications.
Key features
Local AI Models — Download and run supported LLMs from HuggingFace on the local machine. The README lists Llama, Gemma, Qwen, and GPT-oss as exampl…
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…
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…