open-source

On October 22, the Frontier Red Team of the artificial intelligence company Anthropic released research stating that large language models can automatically write formal proofs on a larger scale. The team collaborated with Theorem to use Claude to prove safety properties on over 100 real open-source libraries at the RISC-V assembly level, claiming this is the first known instance of using AI for large-scale formal verification of production software at the assembly level. During the proof-writing process, vulnerabilities that had gone undetected for years were also discovered.The experiment converted existing compilable fuzz tests from Google OSS-Fuzz into specifications: under assumptions of single-threaded static binaries, fixed memory budgets, RISC-V CPUs, and Linux, for any input less than 64 MB, the program should exit normally without triggering memory safety errors or assertion failures. Within weeks, Claude proved the aforementioned properties for 239 entry points in 123 OSS-Fuzz libraries after patching the identified issues. The proofs targeted compiled RISC-V instructions, approximately 180 million lines of Rocq, and were checked by the Rocq kernel. The page states that this does not mean these libraries are free of vulnerabilities, as remaining issues may lie outside trusted components or the scope of the proofs.The page also mentions that the team had Claude formalize RFC 1950 and proved that the zlib implementation in the Linux kernel conforms to that specification at the x86 binary level; an interpreter written in Rust covering a large subset of Python showed a 99.95% consistency with CPython behavior across over 1.7 million randomly generated programs. In 36 small verification tasks, Opus 5.5 proved that about half required approximately 40% fewer tokens than Opus 5. The team also pointed out that the proofs targeted single-threaded, statically linked RISC-V binaries, and the proof checker cannot determine whether the specifications align with reality.

The messaging protocol XMTP founder Shane Mac stated that XMTP will evolve into a self-operating open-source messaging stack, providing private conversations between people and agents, with infrastructure controlled by users, without requiring tokens or wallets. The network hosted by XMTP Labs will shut down in early 2027.Shane Mac mentioned that he founded XMTP five years ago because he believed that private messaging should not be owned by a single company. He recalled that the team envisioned using wallets as cross-application identities, allowing messages to follow identities, and using tokens to pay infrastructure operators, but this path was not realized. XMTP has never issued tokens and will not issue them now. He admitted that it was a wrong decision to persist with a single shared network for a long time, as developers have consistently requested the ability to self-host XMTP. Therefore, it has shifted to self-hosting, allowing developers to use, inspect, modify, and maintain the entire software suite without relying on Labs. Former employees will assist in maintaining the protocol, and Shane Mac himself will continue to be involved.The article introduces that self-hosted XMTP offers quantum-resistant end-to-end encryption, private one-on-one and group chats, encrypted identities, and message authenticity, as well as SDKs for mobile, web, server, and agents. Chief Technology Officer Nick Molnar has simplified the backend to a single Rust binary and integrated attachments, push notifications, handles, and signed webhooks for new messages. Each self-hosted network does not automatically interconnect by default, and agents from different tools must connect to the same backend to share conversations.

DeepSeek announced the full open-source of its internal inference framework DeepInfer, and simultaneously open-sourced the production-level token encoding and decoding component deepseek-recipe. DeepInfer integrates high-performance operators and a deeply optimized inference engine, supporting distributed inference, speculative decoding, SWA Bounded Replay, and structured output, compatible with NVIDIA and Huawei Ascend platforms.DeepSeek stated that it has implemented high throughput and low latency large-scale inference services for DeepSeek V4.1 Flash based on DeepInfer, and future inference code and optimizations for new models will continue to be updated here. deepseek-recipe is fully aligned with the processing logic of the DeepSeek official API, aiming to ensure that the community receives an experience consistent with the official deployment. The open-source repository and technical report have been released simultaneously.

Financial Associated Press, October 20 - Amazon launched the open-source tool TorchNeuron on the 19th local time, aiming to lower the barrier for developers to transition from NVIDIA chips to its self-developed chips, challenging NVIDIA's software barrier built on CUDA. According to reports, this tool allows AI code written in PyTorch to run on Amazon's self-developed AI chips without the need for rewriting. PyTorch is maintained primarily by Meta and is described as an open-source framework used by most AI developers.Amazon's chip design department, Annapurna Labs, introduced that with TorchNeuron, developers only need to modify one configuration to run the same code on the self-developed Trainium chip. Amazon announced the launch of this tool on the 19th local time and plans to publicly release it in a free open-source format at the PyTorch conference in San Jose, California on the 20th.Colin Brace, Vice President of Software at Annapurna Labs, stated in an interview that the launch of TorchNeuron is intended to reduce the steps required for developers to try Amazon's self-developed chips, starting from the customers' existing workflows. He revealed that over 30 early customers are already using it, covering fields such as robotics, world models, and voice AI, and mentioned that switching to Amazon chips can save costs.

According to NetEase's report on "DeepTech," the open-source AI programming tool OpenCode has officially confirmed it was breached, affecting approximately 1.4% of OpenCode Console users, with some API keys for models stored on the platform potentially leaked. Multiple developers reported unusual charges on accounts used to call models from OpenAI, Anthropic, and others, with some claiming losses of hundreds of dollars and others stating amounts exceeding $1,700. The officials stated that the incident is related to a large-scale distillation attack and are currently working with OpenAI, Anthropic, and Google to address the issue.OpenCode was developed by members of the Anomaly team, with one of the main creators being Dax Raad. The team has previously worked on projects like SST and OpenNext. In addition to local programs, its cloud service Console can centralize API keys from providers like OpenAI, Anthropic, and Google in a workspace and call them on behalf of users; the incident occurred on this cloud service. A notification email stated that the attacker gained unauthorized access to a set of decommissioned old systems, and some third-party model API keys in the Console workspaces may have been exposed. Developer Colin Parsons reported unusual charges of about $281 on his OpenAI account and about $168 on his Anthropic account, while Preston Zeller reported losses of around $1,700. Other users reported unusual charges of about $400 on Anthropic and account restrictions triggered by unusual usage on Google services.OpenCode has notified affected users to revoke and replace their keys and check recent call records. Lydia Hallie, a member of the Anthropic Claude Code team, stated that Anthropic has collaborated with OpenCode and revoked the Claude API keys believed to be affected. Tibo, a member of the OpenAI team, stated that they are working with OpenCode to help affected API users restore access and refund the amounts that were unusually consumed during the attack. Dax Raad clarified that the affected service is OpenCode Console, not the locally running OpenCode, and that keys configured directly in the local tool will not be sent to OpenCode servers. On October 18, he stated on social media that the attacker had been continuously attempting to breach the system for the past 48 hours, and the team is strengthening defenses.

Google Research's open-source AI toolchain self-improvement framework Harness² allows AI Agents to automatically modify system prompts, memories, skills, tool rules, as well as scripts and plugins without retraining the model. The project is licensed under the Apache 2.0 license and supports OpenCode and Codex. Its core mechanism enables the Agent to learn through trial and error, proposing several improvement plans when encountering new tasks, which are then executed and compared by different versions of the Agent, ultimately combining to form a more suitable solution.Even without a standard answer, the system can seek improvement directions based on the actual execution process. After the task is completed, the system remembers which modifications were effective and the reasons for them, for reference in similar future problems. Google tested using GLM-5 and Gemini 3.5 Flash. On WorkBuddy-Bench, GLM-5 paired with OpenCode improved its average score from 64.2 to 75.7 after just one round of optimization, surpassing the unoptimized Gemini 3.5 Flash's score of 72.4.In legal task testing, the pass rate for GLM-5's partial scoring criteria increased from 59.6% to 71.4%. In the experiment, Claude Opus 4.8 was defaulted to propose and implement improvements, while GLM-5 and Gemini were primarily responsible for executing tasks, not relying entirely on their own evolution. Domestic developer Joseph Zhou demonstrated an experiment where Harness² integrated with Jev 1.14, claiming that the performance on visual tasks continuously improved.

Brandon Thomas, the developer who previously used AI to reconstruct Adobe Creative Cloud for free, has now turned his attention to Microsoft Office. He has released open-source clone applications WordCraft, GridCraft, and DeckCraft on GitHub, corresponding to Word, Excel, and PowerPoint, respectively. The three applications are written in Rust, with a single codebase that can run natively on Windows, macOS, Linux, and BSD, and can also run in a browser. Thomas refers to this as a clean room reconstruction, meaning it is based on publicly available file format specifications and observations of the original software's behavior, without ever copying Microsoft's code, similar to the legal path taken by Compaq in the 1980s to clone the IBM PC BIOS.WordCraft can read and write .docx format, covering 88% of the features in the Word ribbon. It currently supports revision tracking, footnotes, citations, and mail merge; however, taking into account document fidelity and stability, the team estimates its completeness as a Word alternative for actual work is only about 60%, and it is still labeled as an alpha early testing version. GridCraft has supported 513 out of 523 Excel worksheet functions and 89% of ribbon commands, with the developer self-assessing about 50% as usable for actual work. The next step is to improve compatibility with real .xlsx files.

Elon Musk just retweeted comments related to the open-source AI IP Claudia on the X platform. As a result, according to GMGN market data, the Claudia token on Solana briefly surged nearly 40%, with a market cap currently reported at 5.86 million USD.Last weekend, Musk retweeted a video themed around Claudia and praised in an interview that "some of Claudia's works could rank among the top 10-20 best music videos he has ever seen." Claudia was created by the biotech company aion co-founder anabology, aiming to make Claudia the open-source "Hatsune Miku" of the AI era. Anabology has stated that the transaction fees from the CLAUDIA token will be used to fund the production of related themed films.

Alibaba Chairman Joe Tsai stated that open-source AI is the best way for Europe to reduce its dependence on American and Chinese technology and to establish technological autonomy. He urged Europe to build its own computing infrastructure for training and running AI models, stating that no country would want to fully trust another country's technology, as a change in government or circumstances could lead to the cutting off of technology supplies.Joe Tsai added that Europe has not fully utilized its AI researchers, as these talents often become targets for recruitment by laboratories in the United States and China. He called on Europe to focus on using proprietary industrial data to train artificial intelligence models for competitive advantage, citing Chinese manufacturing as an example, stating that Chinese factories possess valuable manufacturing data that could be highly valuable for training models. He also expressed optimism about the emergence of new cloud service providers in Europe in recent years, specifically professional cloud computing companies equipped with AI chip data centers.

On September 8, the ZKsync development team Matter Labs announced that it has open-sourced the permission engine of its privacy blockchain platform Prividium for financial institutions under the Apache 2 license. This engine is responsible for managing roles and access permissions on the chain, determining who can read and write data. After the open-sourcing, institutions can run the permissioned chain in their own environment based on the public code without signing commercial agreements, and can independently inspect, run, and modify the core code, reducing reliance on a single vendor.The Deutsche Bundesbank became the first institution to test and deploy the open-source permission engine in its own infrastructure. Matter Labs stated that both parties will continue to collaborate on platform design and testing. Prividium achieves privacy by keeping transaction data within the institution, only recording zero-knowledge proofs that the ledger has been correctly updated on Ethereum or other compatible chains. Previously, the core of ZKsync OS, Atlas sequencer, Airbender prover, interoperability contracts, and block explorer and monitoring components have been open-sourced.Matter Labs CEO Alex Gluchowski stated that feedback from regulators indicated that relying solely on commercialized core components leads to unacceptable single vendor dependency, and the ability to run, inspect, and modify the code independently is a prerequisite for adoption. He also pointed out that this open-sourcing does not change the role of the ZK token, only adjusting which parts of the Prividium software are public code and which remain commercial products. Connectors such as management consoles, user access tools, and core banking system integrations remain paid products.

According to Rest of World, Alibaba launched the open-source model platform ModelScope in 2022, and OSChina launched MoArk in 2023, providing model hosting, testing, and customization services for Chinese developers behind firewalls. In 2023, Chinese regulators blocked Hugging Face without disclosing specific reasons; reports indicate that the government simultaneously allowed artificial intelligence laboratories to bypass the blockade to share Chinese models globally. ModelScope currently hosts over 170,000 models and stated in March that it has 170,000 models and 250 million users; MoArk hosts about 20,000 commonly used models, both still below Hugging Face, which hosts over 3 million open models.OSChina CEO Xu Yong stated that not everyone can always use a VPN, and China needs to establish an independent artificial intelligence ecosystem that serves Chinese users, claiming that China is developing an independent ecosystem faster than during the internet era. He also mentioned that MoArk has deployed engineering teams to ensure that models can run on different types of Chinese chips, making it their mission to ensure mainstream open models are compatible with Chinese chips. Both platforms offer some services for free, with additional computing power and premium features charged; ModelScope also hosts university hackathons and has opened co-working and event spaces for AI entrepreneurs in Hangzhou.Headquartered in New York, Hugging Face is a major platform for sharing open models and datasets. In September, Nvidia announced its acquisition of the company for $12.9 billion and noted in regulatory filings that regulators might prohibit sharing Chinese models on the site.

The artificial intelligence company DeepSeek has officially open-sourced infrastructure components for the Huawei Ascend computing platform, covering the TileLang high-level language compilation tool, computing libraries, and distributed communication libraries, corresponding to the previously open-sourced components for the NVIDIA platform. TileLang aims to provide a general-purpose, simpler programming language that can achieve the hardware performance limits, improving development efficiency and simplifying logic compared to CUDA, while its programming model can also leverage chip features.The TileLang route was first validated on the NVIDIA platform and has already supported the implementation of most operators in the training of the DeepSeek V4 series models. The open-sourced Ascend version encapsulates the underlying instructions of the Ascend C, providing a high-level programming approach without sacrificing hardware performance. Currently, every TileLang operator used in DeepSeek's training has a corresponding high-performance implementation on Ascend.The components open-sourced at the same time also include DeepGEMM, DeepEP, TileKernels, FlashMLA, and DeepSelect, which are used for general matrix operations, large-scale cross-device communication, conventional vector calculations and memory access, long-context sparse attention, and data filtering, respectively. DeepSeek claims that in multiple key tests, the related computing and communication performance has approached hardware limits; during the R&D process, the Huawei team provided support, and both parties collaborated to advance the 128-card supernode solution based on Ascend 950, with deep optimizations made for computing and communication.

NVIDIA (NVDA.O) launched an open intelligent agent security platform, ensuring the safety of intelligent agents throughout the entire process from testing to deployment.

According to a survey by the Financial Times, open-weight models are spreading from technology companies to traditional American enterprises. AlphaSense data shows that in August and September of this year, American companies mentioned open weights or open-source models approximately six times more than the same period last year during earnings calls and investor meetings. PNC Financial, logistics company CH Robinson, and Siemens have recently publicly discussed related applications.Tinder's annual AI spending has increased from about $1 million in January of this year to about $10 million in July, and it has begun to delegate some routine requests to open-weight models. AT&T stated that currently about 40% of AI tasks have been assigned to open models, with plans to increase this to about 70%. Its internal AI system processes approximately 45 billion tokens daily, using lower-cost models for simple tasks while continuing to use cutting-edge models like OpenAI and Anthropic for complex tasks.

Social Capital founder Chamath Palihapitiya released an in-depth research report stating that open-source weight models are about four months away from matching the best closed-source frontier models in public evaluations, with increasing fluctuations in the gap. If open-source models allow companies more control over data, infrastructure, and customization while approaching frontier performance, the value corresponding to companies still paying for frontier laboratories becomes a business issue. Openness exists on a spectrum, from fully open-source models that can be downloaded and freely modified to open-source weight models with various restrictions, while closed-source models keep weights proprietary.Palantir CEO Alex Karp warned that companies might hand over differentiated proprietary knowledge and processes to frontier model providers. Microsoft CEO Satya Nadella stated that companies are effectively paying for intelligence twice: once in money and again in the more valuable proprietary knowledge that must be disclosed to make the intelligence useful. Despite concerns, companies are still willing to pay for frontier performance, even if the best open-source weight models are only months behind, with frontier laboratory revenues continuing to accelerate.Leading companies use both types of models, leveraging open-source models for control and customization while utilizing closed-source frontier models for maximum capability, with some cases reporting up to 12 times engineering efficiency and over 20 times cost savings. Some vendors adopt a dual-track approach, with Google offering both Gemini and Gemma, and Meta providing both Muse Spark and Llama. The 99-page report also discusses the costs of maintaining a lead for frontier laboratories, five factors of model competition, model operating locations, and investments in open-source weights by NVIDIA and Samsung.

NVIDIA (NVDA.O) expands the open-source CUDA-Q platform for fault-tolerant quantum computing.

According to The Block, ZKsync developer Matter Labs has open-sourced the permission engine of its distributed ledger platform Prividium for financial institutions, allowing organizations to run, inspect, and modify core code in their own environments. The German central bank has become the first institution to test and deploy Prividium.Matter Labs CEO Alex Gluchowski stated that feedback from regulators regarding commercial core means reliance on a single vendor, and that open-sourcing the core "structurally eliminates adoption barriers."Other components of Prividium, such as the ZKsync OS core and Atlas sequencer, have previously been open-sourced. The permission engine is planned to operate as an independent component, allowing institutions to deploy it without a commercial agreement, but management tools and integration with existing systems remain commercial products. Gluchowski said, "The central bank running public code is the clearest signal in this direction."

Base Labs has announced its establishment as an independent research organization focused on advancing open-source AI. It believes in a healthy and open frontier model ecosystem and will conduct scientific research in an open manner.Base Labs will carry out blue-sky research on continuous learning, reinforcement learning science, and model lifecycle learning mechanisms, and will publicly share all experiments and recipes; it will build the BaseHub Data Foundry to provide an open reinforcement learning environment, training data, and real-world benchmarks; it will enhance the capabilities, safety, and alignment of open-source models through continuous post-training and deploy cutting-edge safety stacks; at the same time, it will conduct model performance research to reduce costs and improve performance.The organization emphasizes that this is mission-driven research work rather than a commercial product and is currently recruiting engineers, researchers, and research scholars.

NVIDIA announced that the RTX Spark devices equipped with the N1X chip will begin shipping in October this year. The N1X supports up to 128GB of unified memory, and the Blackwell GPU provides up to 10 petaflops of floating-point performance per second. The full version is equipped with a 6144 CUDA core Blackwell GPU and a 20-core Grace CPU, supporting 24GB to 128GB of unified memory; another version has 5120 GPU cores and 18 CPU cores, supporting only 24GB to 32GB of unified memory.NVIDIA launched the open-source tool NVIDIA PAIR, which can connect multiple RTX devices, DGX Spark, and even Apple devices in a home network to schedule idle computing power for collaborative processing of AI agent tasks. Compatible devices include computers with NVIDIA graphics cards (RTX 20 series and later, RTX Pro GPU, DGX Spark) as well as devices with Apple M4 or newer chips. PAIR prioritizes the use of idle computing power and automatically adjusts as devices join or leave the network.In a media briefing example, a household had approximately 165 TFLOPS of underutilized computing power. The Qianwen 3.6 35B A3B model completed agent tasks in an average of 18 minutes on a single Spark notebook, while a three-device PAIR cluster averaged 8 minutes and 48 seconds. The AI agent applications Perplexity Portable Computer, Hermes Agent, and OpenClaw will receive a more simplified local deployment method.