#开源生态

Building on open source AI? Hugging Face’s policy chief on the costs, risks, and trade-offs

On TechCrunch's Equity podcast, Rebecca Bellan spoke with Irene Solaiman, Chief Policy Officer and SVP of Strategy at Hugging Face, about what startups need to know before building on open models, from choosing tools and infrastructure to navigating trade-offs in open AI. The episode covers why startups, enterprises and governments are switching to open models to cut costs, who makes money when models are free to use, why the U.S. may need deep-pocketed players like Nvidia to build an open frontier model that can compete with China's, and how recent attacks on open AI platforms raise questions about keeping an open ecosystem secure.

HumanLayer is joining Factory

HumanLayer announced it is joining Factory to accelerate building autonomous, self-improving software for enterprises and developers. The company says it will open source HumanLayer and bring features such as multiplayer tasks/workstreams, in-app collaboration on docs and plans, and live session diff streaming into Factory. The current HumanLayer platform will remain running for a transition period of at least 30 days, with support for exporting data, migrating to Factory, or self-hosting the open source, and paying customers will be refunded for the last several months.

AI Is Getting Better. Running It Yourself Is Getting Harder

The article argues that large Mixture-of-Experts models such as DeepSeek-V3 (671B total, 37B active parameters) and Llama 4 Maverick (400B total, 17B active) can be efficient to compute yet costly to host, since all weights must remain available even when only a subset is activated per token. It contends that batching and shared deployments favor large providers, so open weights may become centralized in practice as self-hosting costs rise, and it suggests shared colocated clusters and research into compression, distillation, and expert offloading as responses.

vLLM Delivers Day-0 Support for Kimi K3.1 Pro

vLLM announces day-0 support for Kimi K3.1 Pro, Moonshot AI's 2.8T-parameter (104B active) multimodal MoE with a 1M-token context that adds a DSA sparse-attention indexer to all 24 full-attention layers of its KDA + MLA hybrid. The post details a packed KV cache for three kinds of state, token-sharded prefill attention (up to 1.69× faster TTFT at 340K tokens on 8× GB300), decode context parallelism with an exact top-k merge, a fused low-latency decode path, a DSpark speculator preview, AgentX benchmark results on GB300 and Rubin, zero-mismatch RL, and support on NVIDIA, AMD, TPU and Alibaba T-HEAD Zhenwu PPUs.

Ai2 and Google collaborate on a new generation of climate models

Ai2 is launching a climate modeling initiative with multimillion-dollar, three-year funding from Google.org and technical support from Google Research to develop NeuralGCM+, an open-source model extending Google Research's NeuralGCM to simulate atmosphere and ocean together for faster, more reliable decade-scale climate projections. Ai2 will focus on climate trends over years to decades, such as how rainfall and water availability could change through mid-century in regions like the southwestern United States, while Google Research continues developing NeuralGCM for nearer-term forecasts over weeks to months. The work will proceed along three fronts: connecting the atmosphere and ocean, improving the response to a warming world, and making the tools easier to use and extend, including improved performance on widely used GPUs.

Polymorphic Inference: self-tuning inference for open models

Together AI introduces "Polymorphic Inference," an approach for adapting inference to fast-changing model architectures, traffic shapes, and hardware, contrasted with the older "Static Inference" paradigm. The company announces Together Compiler, an AI-agent-driven "Continuous Compilation" optimization loop already running in production on some serverless endpoints, and previews upcoming work including self-serve compilation, shape-shifting speculative decoding, the ThunderAgent Router, post-training, and TogetherLink workload-aware model routing.

The biggest modding site in PC gaming is taking a stance on AI

Nexus Mods announced that mod authors must now declare whether AI was used in their work, with "AI-assisted" mods also requiring a public explanation of how and where AI was used, and AI tags made more prominent on mod pages. Site owner Victor Folmann said the platform is "pro-human creativity" but won't ban AI-generated mods outright, arguing a ban would only push AI use underground; Nexus Mods also said it won't use mods to train AI or sell or license them to anyone doing so. Mod authors have 60 days, ending January 2nd, to update existing mods before enforcement begins, and the site is adding controls to hide AI-generated content and limits on bulk uploads.

Why Every AI Model Will Ship as a Torrent

A Hacker News submission links to the article "One Model, One Swarm" at pirateface.co, arguing that every AI model will ship as a torrent. The submission has 14 points and 2 comments.

Revolut swaps Claude for GLM open-weight model, cutting AI coding bill by 90%

Revolut has moved its roughly 2,000 software engineers off Anthropic's Claude Code and OpenAI's Codex as default coding tools onto GLM, an open-weight model from Chinese lab Z.ai that Revolut hosts itself, bringing its bill to about a tenth of what it paid Anthropic and OpenAI. David Abitbol, who runs AI operations and infrastructure at Revolut, said the company serves GLM from about 56 Nvidia B200 GPUs at a blended cost of around $0.30 per million tokens, and reported roughly 60% shorter coding time and 70% shorter lead time in a sample of about 2,000 engineers, though the comparison does not isolate the model change from other causes. Abitbol said the 90% saving excludes cloud, staff and monitoring costs and depended on heavy tuning, with out-of-the-box inference "terrible" until time to first token fell from about 14 seconds to 1.7 seconds on the same GPUs.

KVCache Meetup Korea 2026: Event Recap

PyTorchKR and XCENA co-hosted KVCache Meetup Korea on Sep 30, 2026, with Tensormesh sponsoring, featuring talks from XCENA, SK hynix, FuriosaAI, and LMCache on LLM inference and KV Cache optimization. Sessions covered CXL memory expansion and the Maru KV Cache storage engine, GPU and HBM utilization analysis with H100 GPUs and vLLM, request scheduling and KV Cache management in Furiosa-LLM on RNGD, and LMCache's MP Mode for unified KV Cache management. A survey of 72 responses cited the program (66.7%) and venue and facilities (41.7%) as positive points, with requests for better audio, screen visibility, interpretation, and real production examples.

Transformers Model Integration Contracts

Transformers introduces Model Integration Contracts (MICs), which pair a shrunk fixture model built from a real checkpoint's configuration with a runnable usage path and human-reviewed baselines, run on cheap CPU runners in PRs. Replaying frozen MICs over five months of commits showed a small UMT5 fixture flagged a decoder-attends-to-future-tokens regression from July 30 to September 28, 2026, which a forward pass alone would have missed. The tmic CLI and agents help draft contracts, and the MIC CI job currently reports failures without blocking PRs.

[AINews] LLMs can read & write neural network weights: "Weight Space Programming" cuts Autoresearch out of the RSI loop

AINews reports a paper in which a smart model directly reads and edits the weights of a smaller model, described as one-shot fixing bugs from failed traces, improving in a general direction over about 50 steps, and writing entire network weights from scratch that beat some independently trained models. The post frames this as a new recursive-self-improvement capability that emerged suddenly, notes an independent reproduction appears to have already found it, and cautions it is unlikely to scale beyond simple games and RL policies for now. The same issue also covers DeepSeek's open-sourced "super-agent" harness prototype, the GDPx benchmark where the best model scores 12%, and October model releases.

Code-First Internal Tools vs. Low-Code in the Age of AI

OpenRun, an open source self-hosted GitOps platform for deploying web apps and internal tools declaratively, argues that AI coding tools have removed most of the speed advantage of low-code platforms' visual editors, since generating a small app from a prompt now takes about as long as assembling it on a canvas. The post contends that low-code downsides remain—proprietary app formats, editor-only editing, limited testing and refactoring, per-user pricing and security features gated to higher tiers—while code-first apps benefit from reviewable pull requests, easier maintenance and no lock-in to a tool or model. It also notes that code-first internal tools still need a platform for hosting, SSO, access control, audit logs, database credentials and deployment, which is the gap OpenRun targets.

The best Python AI agent frameworks in 2026

This roundup compares nine Python agent frameworks against criteria including model portability, durable execution, human approval, frontend and voice reach, MCP support, tracing and commercial support, and picks Pydantic AI overall for covering all of them in the open-source library. It then recommends LangGraph/LangChain, the OpenAI Agents SDK, Google ADK, Microsoft Agent Framework, Strands Agents, CrewAI, the Claude Agent SDK and Agno for specific vendor or design preferences, each with a stated catch. A comparison table lists model providers, durable execution engines, speech-to-speech voice, interfaces, MCP, observability, commercial support and license per framework, checked October 29, 2026.

How Olmo 3 earned the highest ethical rating from Nextcloud

Nextcloud awarded Olmo 3 its highest Ethical AI rating, green, which requires open-source training and inference software, a self-hostable trained model, and publicly available training data. Nextcloud bundles Olmo 3 with its llm2 local LLM backend, citing the 7B model's ability to run on consumer hardware without high-end GPUs and its fit for customers seeking independence from big tech. Nextcloud engineer Marcel Klehr said fully open models make AI transparent, explainable, auditable and controllable, and that the company is working on expanding its AI features.

Ask HN: "Net neutrality" for AI compute and memory?

An Ask HN post asks whether something like "net neutrality" should apply to DRAM and compute, noting that both Sam Altman and Elon Musk have raised compute/DRAM scarcity and questioned whether local open weights models are the most sustainable solution. It asks whether allocation should follow a dynamic market-based price mechanism, go to the highest and best use, or let whoever pays use it for any purpose, and whether laws should guarantee equal access to DRAM/compute for all.

Announcing the Special Interest Group (SIG) for Ray on Accelerators

The Ray community announced a Special Interest Group for hardware accelerators, aiming to bring together hardware vendors, Ray contributors, and application developers around a vendor-neutral roadmap for accelerator support. Its initial roadmap covers development and testing infrastructure, hardware-aware scheduling, and networking/transport, including Ray Direct Transport for direct device-to-device data movement. Ray currently supports more than a dozen kinds of accelerators, from NVIDIA and AMD GPUs to TPUs, Trainium, and Gaudi.

We replaced 11,000 lines of AI chat code with CopilotKit and AG-UI

fynk describes replacing its custom AI chat with the open AG-UI protocol and CopilotKit, removing about 11,000 lines of its own code. Because CopilotKit was React-only and fynk is a Vue app, engineer Abel Enekes contributed a Vue port (62,517 lines across 244 files) that was merged on 13 May 2026 and published as @copilotkit/vue. The post details the assistant's shared state, tool-call traces, human-in-the-loop confirmations, generative UI, and in-house citations and permissions, and advises adopting what every agent needs while building what differentiates the product.

Haas: Postgres 19 is late, and it's AI's fault

PostgreSQL developer Robert Haas writes that the delay in shipping PostgreSQL 19 is not particularly concerning, as it is short and stems from the project's concern for quality. He says the long-term picture is less clear: with forty years of bugs that humans could not find but LLMs can discover, it is unclear how the project will catch up or continue to innovate.

How Small Can We Slice an AI Model?

The article argues that decentralized AI has been limited by pipeline parallelism, which assigns whole layers to individual machines and excludes devices with only a few gigabytes of spare RAM. It proposes sub-layer sharding instead: hosting individual experts from sparse mixture-of-experts models and compressed attention state on everyday devices, with a routing layer sending tokens only to the relevant shards. It surveys existing decentralized projects (Petals, Aethir, Akash, Exo) and frontier-lab work on sparse architectures (DeepSeek's MLA, Shazeer's research, DeepMind's Mixture-of-Depths), while noting the proposed swarm and memory budgets have not been demonstrated.

The Limits of AI as an Economic Weapon

Max Bessler argues that the rise of capable, free open-weight models from Chinese firms such as Moonshot, Alibaba, Zhipu and DeepSeek is splitting the AI market into a commoditized tier and a smaller premium tier, undermining Washington's ability to use AI models as a coercive chokepoint. He compares this to Japan's 1980s DRAM dominance, which captured volume but yielded little leverage, and suggests the U.S. should instead keep chip export controls narrow and maintain control of the cloud and inference layer where models run.

I built an AI Tamagotchi before Meta could

The Verge's Sean Hollister bought a $50 Waveshare ESP32-S3 1.75C, a Meta-sanctioned devkit for building "Muse Gadgets," and found it nearly useless out of the box until he used Claude Code to add conversation history, spoken replies via ElevenLabs, a camera preview, multiple avatars, and a Tamagotchi-style minigame. He notes he has no coding experience and can't vouch for the code's quality, and that he keeps the device on guest Wi-Fi over privacy concerns. He concludes the appeal of dedicated hardware is reprogrammability rather than Muse itself, and wonders whether the real Muse Charm will be as freely programmable.

Open-Weight vs Closed-Weight Models: What's the Difference?

Open-weight models publish their trained weights so anyone can download, run, fine-tune or serve them, while closed-weight models are only available through their maker or licensed clouds. On OpenRouter, open-weight models grew from 26.8% of top-50 model tokens in October 2025 to 71.7% in late September 2026, and cost about a quarter as much per token, though the best closed-weight models still score about 11 points higher on the Artificial Analysis Intelligence Index.

PostgreSQL 19 Running Behind Schedule Due to AI/LLM Burden

PostgreSQL 19, typically released every September–October, will not arrive until at least November this year because developers have been overwhelmed by increased bug reports and patches from AI/LLM activity. Developer Robert Haas said LLM-drafted fixes still require human review and commitment, and time that would have gone to reviewing release contents was diverted to security issues and back-branch bugs, delaying decisions on what needed reverting before shipping v19.

Was I wrong about open models & cyber risks?

The author revisits his earlier claim that open-weight models like GLM-5.3 crossing a cyber-capability threshold would not produce a step change in societal harm, noting that over a month after release there is little public evidence of such harm and that closed models have featured more heavily in recent hacks. He concedes he should have credited AI safety voices who framed this as a new era of cyber instability, and argues the real issue is AI capability and diffusion generally rather than open versus closed weights.

Open-weight AI models can exploit the patch gap

Researchers found that GLM 5.3, a widely available open-weight model released by Z.ai in August, could autonomously build exploits against all six Firefox releases between September 17 and October 26 using only information available at each version's release, at a marginal cost of about $1,000 per version. Visiting the model-built webpage caused the browser to read a local file and send it to an attacker's server without further user interaction, and the attacks also relied on an unpublished vulnerability that Mozilla has since patched in Firefox 159.0 (released October 27, 2026). Similar attacks on Google Chrome did not succeed as of writing, and the authors expect the threat to worsen as open-weight cyber capabilities improve.

Featured coverage of this story: 《Open-weight model GLM 5.3 built full Firefox exploit chains with zero human input in under 48 hours》

ResearchOCT 29TH, 2026Prime Mega MoE: Fast Fused MoE Kernels for Training

Prime Mega MoE extends DeepSeek's Mega MoE kernel, which previously had only a forward implementation, with a fast overlapping backward pass so it can be used for training. It reports up to 3.11× faster MoE kernels than the default PyTorch-based implementation and up to 1.34× faster end-to-end training than the Prime-RL baseline (34% higher throughput). Supported data types are bfloat16, MXFP8 (Blackwell only) and blockwise FP8 (Hopper only).

AI Contributions to GHC

GHC now has a formal policy for LLM-assisted contributions with four policy points and one strong preference: prioritise people and community, make merge requests easy to review, require an identifiable human author who takes full responsibility, and disclose direct LLM generation, while preferably writing the final version yourself. The policy follows Simon Peyton Jones's draft rather than the competing July "shared statement," and the article argues it is reasonable because reviewer time is the scarce resource and a blanket ban would be hard to police and aimed at a problem GHC is not yet seeing at scale.

How MoEngage turned over 400 APIs into agent tools with AgentCore Gateway

MoEngage used Amazon Bedrock AgentCore Gateway to expose roughly 400 internal REST APIs across seven AWS Regions as permission-scoped agent tools, avoiding over 20 separate MCP servers and per-API networking. Cedar policies filter tools server-side per agent identity, managed egress over Amazon VPC Lattice keeps APIs private, and teams expose operations by annotating OpenAPI specs rather than writing MCP code. MoEngage estimates the approach saved about six months of platform engineering effort, with new API onboarding now taking hours instead of days.

Advisories in Five Months: Results from the Drupal AI Security Initiative

From June through October 2026, Drupal published 256 security advisories, more than the 250 published across the previous 30 months combined, with monthly volume rising from an average of 8 in early 2026 to 106 in October. The Drupal Association attributes the surge to cheaper AI-assisted vulnerability discovery rather than a change in Drupal's attack surface, and says the Alpha-Omega-funded SEIR team participated in 223 of the 256 advisories. The post describes Phase 1 lessons (intake and review are now the bottleneck), upstream fixes such as an Opengrep PHP parser bug, and Phase 2 plans to extend Triagebot to the GitLab API.

Welcome Le Chonk: Mistral Large 4 on Hugging Face

Mistral Large 4 ("Le Chonk") is a multimodal, multilingual Mixture-of-Experts model with about 1T total and 49B active parameters and a 1M-token context window, released in FP8 and NVFP4 variants under the Mistral Ecosystem Open-Weight License. It combines a 72-layer text decoder using MLA with a DeepSeek Sparse Attention indexer and a 1.6B Pixtral vision encoder, and the post reports 82% on the Artificial Analysis Cyber Index vulnerability reproduce-and-patch test and 93% on Cybench. Day-0 support covers Transformers, vLLM, llama.cpp, TRL fine-tuning, Inference Endpoints and Dell Enterprise Hub, with GGUFs from Unsloth and the llama.cpp team.

Announcing Project ProtoStar: Making Publicly Funded Science AI-Ready

Project ProtoStar is a new initiative bringing together universities, industry, and the open science community to release publicly funded scientific datasets, models, and code in AI-ready, machine-actionable formats on open infrastructure. Founding partners include Carnegie Mellon University, Purdue University, Georgia Institute of Technology, Arizona State University, and The George Washington University through the Open Forum for AI, the Astera Institute, Hugging Face as infrastructure partner, and ORCA as community partner. The first open challenge, hosted by Carnegie Mellon University from November 2–13, 2026, asks participants to use historical weather data from 210 U.S. cities through 2018 to predict 2019–2026 climate conditions and propose a climate adaptation policy idea.

Deepgram acquires Pipecat-maker Daily to bolster its voice AI infrastructure

Deepgram announced it is acquiring Daily, the maker and maintainer of the open-source voice AI framework Pipecat, with the deal amount undisclosed and Pipecat to remain open-source. Daily, which has raised over $62 million from investors including Tiger Global and Nat Friedman, was in the same 2016 Y Combinator batch as Deepgram, which reached unicorn status with a $130 million Series C earlier this year. Deepgram is also launching the Real-Time Context Standard (RCS), an open protocol for voice agents co-developed with the Pipecat team.

Roasting China’s AI Industry

In a ChinaTalk interview, Xianyu, the writer behind the satirical blog "Funeral AI" (葬AI), discusses why AI writing isn't funny, how he built MemeBench — a public arena where ordinary users rank video models by meme-ability — and why Five-Year Plans function as a kind of benchmark while "AGI" lacks quantitative indicators. He describes recruiting about 300 students via QQ to seed the rankings, and recounts that MiniMax's open-source H3 beat ByteDance's Seedance in the meme arena, prompting ByteDance to question the result and MiniMax to complain about an earlier article. He also argues that Chinese AI startup themes rotate every three months following Silicon Valley, with many "world models" actually being post-trained versions of Alibaba's Wan video model.

Block Unveils Dessa, a Full-Duplex Speech Model Trained Only on Speech

Block announced on October 23, 2026 a research preview of Dessa, a full-duplex speech-to-speech model trained only on human conversation audio, with no text and no pretrained models. Block says it trained Dessa on one NVIDIA GB200 NVL72 rack and that it runs locally on a MacBook Pro M4 Max or better, faster than real time. No benchmark scores, parameter count, weights, paper, API or price have been published, and access is via a waitlist.

Ask HN: Why not AI assisted ZFS rewrite?

A Hacker News poster asks why no filesystem developer has done a clean-room rewrite of ZFS, using LLMs under experienced engineers' guidance, to create a ZFS-like filesystem free of Oracle's licensing and upstreamable to the Linux kernel. They cite clean-room rewrites of Photoshop, MS Office and Fusion CAD, and note BcacheFS's promise ended when its core developer left the Linux kernel. The poster acknowledges filesystems need maturity and battle testing and that vibe-coding one would be foolish, framing the idea as an open-source project with the right contributors.

NVIDIA Releases Nemotron 3.5 Transcribe: An Open 30B MoE Model for 40-Language Transcription with ~4B Active Parameters

NVIDIA released Nemotron 3.5 Transcribe, an open-weights speech LLM with about 34B total and 4.3B active parameters that handles 40 languages and audio up to 1 hour, with text instructions selecting verbatim, speaker-attributed, contextual or translated output. NVIDIA reports 4.71% mean WER on 9 public English Open ASR sets, the best open-weights score, while commercial APIs still lead, and notes weak Mandarin meeting accuracy (34.95% cpCER on AliMeeting) and offline-only operation. Weights ship in BF16, FP8, NVFP4 and GGUF under the OpenMDW-1.1 license, with runtimes including vLLM 0.32.0, Transformers and SGLang-Omni 0.2.0.

MicroZig to Adopt Strict No LLM / No AI Policy

The Zig Embedded Group has published a Code of Conduct for its GitHub organization and Discord server that includes a strict No LLM / No AI policy taking effect October 28th, 2026. It prohibits LLM-generated code or prose, paraphrasing LLM output, using LLMs for editing, translation, brainstorming, or finding bugs, and discussing chatbot/LLM use. The document also sets out expected and unacceptable behavior and enforcement contacts Matthew Knight and Felix Queißner.

Akai is embracing vibe-coded mods with an open-source MPC

Akai is giving hackers and custom firmware developers early access to the not-yet-announced MPC Sample Key, which will have open-source firmware and likely run a modified version of the MPC Sample software adapted for keys, pitch, and mod wheels. Instagram Reels show vibe-coded mods adding synths such as Native Instruments Massive, virtual analog and FM synths, and one running a fully custom OS, with the same community features reportedly coming to the regular MPC Sample. CEO Jack O'Donnell told The Verge that inMusic brands, especially Akai, would seed early devices to developers and open-source some software while providing an SDK.

Why Agents Love Supabase

The post discusses a podcast with Supabase founders Paul Copplestone and Ant Wilson about why agents favor Supabase, citing its near-sweep as first-choice app backend across frontier models. It highlights practical announcements from Supabase Select—running Supabase locally without Docker, declarative schemas kept in Git, and Supabase Compute in private alpha—plus Scale efforts including OrioleDB, Multigres and the Turso acquisition. The founders say Supabase launches about five million databases a month and that AI-driven usage improved activation and conversion.