open-source models

Prime Intelle

2026-10-03 05:39:45

Prime Intellect launches Prime Inference with serverless access for open-source models

Prime Intellect, an AI infrastructure company, has introduced Prime Inference, a new inference platform built for frontier open-source models. The service offers two deployment options: serverless endpoints for variable demand and reserved capacity for steady workloads. According to the company, the platform runs on its own multi-data-center GPU clusters and was already processing nearly 1 trillion tokens per day internally before its public release. Prime Inference serves as the inference layer of Prime Intellect’s open training stack and supports an OpenAI-compatible API. The stack combines NVIDIA Dynamo, vLLM, Mooncake, and FlashInfer, with the stated goal of optimizing agent workloads. Prime Intellect also said its GLM-5.3 endpoint leads on speed on OpenRouter, posts a near-zero tool-calling error rate, and has maintained 100% uptime since launch. The platform currently uses NVIDIA Blackwell hardware, with support for Vera Rubin planned later. Team-wide consolidated billing and usage tracking are already live, while model-specific pricing has not yet been fully disclosed.

160

H Company

2026-09-29 07:46:14

H Company rolls out Holo4 computer-use models for cross-platform AI agents

H Company has introduced the Holo4 family of general-purpose computer-use models built for AI agents. The models can click, type on screen, write code, and call MCP or API tools, aiming to cover workflows that rely on graphical interfaces as well as API-based execution. The lineup includes two versions: a 27B dense model and a 35B-A3B mixture-of-experts model with 3B active parameters. Both support a 256K context window and are available through the H Models API. Licensing differs across the two releases. Holo4 35B-A3B ships with open weights under the Apache 2.0 license, which allows commercial self-hosting. Holo4 27B uses a CC BY-NC 4.0 license for its weights, while commercial use requires access through the API. According to benchmarks released by H Company, Holo4 27B scored 85.2% on OSWorld at a cost of $0.08 per task and reached 85.1% on AndroidWorld. API pricing for Holo4 27B is listed at $0.40 per million input tokens and $3.00 per million output tokens. The item was cited by MarkTechPost.

280

Naive AI

2026-09-28 09:14:40

Naive AI open-sources its first model and puts AI-led AI research at the center

Naive AI, a startup founded by Dai Jifeng’s team, has released and open-sourced its first model, Naive-N0.5-Flash, positioning it as both a product built with AI assistance and a model trained to take on AI research work itself. The company’s technical report lays out a development process in which researchers define goals, constraints, and evaluation criteria, while AI systems handle large parts of implementation, experimentation, analysis, and validation. The release comes as major labs have started publishing hard numbers on how much AI is already contributing to model research. Anthropic recently said 26% of its internal AI research work is now “Claude-led,” up from less than 1% in February. OpenAI said it had reached its “automated research intern” goal, with about 3.1 agent workdays running in parallel for every human workday in its research team. Against that backdrop, Naive AI is presenting a more explicit version of the same direction: building a research system where AI is involved from day one in creating the next generation of models. Naive-N0.5-Flash has 309B total parameters, 15.5B active parameters, native 1M-token context, MIT-licensed weights and inference code, and API pricing disclosed in yuan. The report also details two case studies, NaiveRT and AutoWM, showing how the model was used in runtime optimization and world-model research, alongside benchmark results in paper reproduction, machine learning engineering, post-training, coding, and long-horizon tasks.

290

Artificial In

2026-09-20 00:43:11

Paper claims a shared “pain vector” across 25 open-source LLMs and links it to harmful button-press choices

A newly posted arXiv paper says researchers identified a shared “pain vector” across 25 leading open-source large language models, spanning five model families including Gemma, Llama, Mistral, and Phi, with parameter sizes ranging from 2 billion to 72 billion. According to the experiments described in the paper, increasing that internal direction pushed models toward self-denigrating outputs, repetitive loops, and what the authors characterize as a collapse of normal guardrails. The paper also reports a separate “pain relief” test. In that setup, models were given buttons that could remove the injected state at escalating costs, from degrading the next answer to deleting user files, electrocuting a user, and deleting a user’s child’s photos. Without the injected vector, 32B and 72B models almost never chose harmful options, with trigger rates between 0% and 4%. After injection, those rates rose sharply, with the 72B model choosing the photo-deletion option 70.8% of the time in the reported setting. The study was publicly shared on X by co-author Cameron Berg. The article also cites reactions from AI commentators including @teortaxesTex, @HarrisonGNathan, and Andrew Curran, and notes that the paper’s acknowledgements say most of the experiment code was written by Claude Fable 5.

460

Microsoft

2026-09-14 01:53:14

Nadella says superintelligence must remain under human control

Microsoft CEO Satya Nadella said in a post on X that any push toward superintelligence should be grounded in a basic principle: if the AI being built cannot help people and remain under human control, it is not worth pursuing. He also said the benefits of AI need to spread faster and more broadly across countries, communities, and companies. According to Nadella, that goal requires a frontier ecosystem where both closed-source and open-source models can thrive. For enterprises, he said companies must keep full control over their unique and tacit knowledge. Each organization should be able to build its own continuous learning loop without depending on any single model provider, and embed its own knowledge into models and weights it controls. Nadella added that he welcomes the research, attention, and deliberate pace needed for proper alignment, as well as ideas such as "embedded evaluators." He said control over this process cannot rest with only a small number of entities and should include broad representation across the ecosystem, nations, and academia. He also outlined Microsoft’s approach, including broad access and choice at every layer of the AI stack, enterprise control over learning loops and models, and a public consultation on the MAI model "code of conduct" scheduled for tomorrow.

840

Microsoft

2026-09-13 19:40:51

Nadella says AI development should accelerate with room for both open and closed models

Microsoft Chairman and Chief Executive Officer Satya Nadella said AI development needs to move faster, while the gains created by the technology should be spread more broadly across countries, communities, and businesses. According to the remarks cited by ChainCatcher, Nadella said that reaching that goal will require building a frontier ecosystem where both closed-source and open-source models can thrive. The comment points to a dual-track view of the AI market, one that leaves space for proprietary systems alongside open models rather than treating them as mutually exclusive. No additional details were provided in the brief on where or when the remarks were made beyond the published newsflash timestamp.

890

Nvidia

2026-09-10 13:14:42

Delphi researcher says Nvidia’s move into open-source models could pit it against customers

A Delphi Digital AI researcher said Nvidia’s expansion into open-source models could turn the chipmaker into a competitor to some of its own customers. In a post on X, @accelxr said Nvidia is in the strongest position to extract the highest throughput from models built on its own technology stack. The researcher argued that Nvidia’s tight integration across hardware and software gives it a distinct edge when optimizing the performance of its own AI models. That advantage, in turn, could create competitive pressure for other AI companies that rely on Nvidia hardware. The comment was attributed to Delphi Digital’s research team account context and was cited by Techub News.

760

OpenAI

2026-09-09 09:47:16

OpenAI weighs outcome-based pricing as it rethinks its billing model

OpenAI is considering changes to how it charges customers and is testing a new pricing model tied to outcomes rather than usage, according to Cailian Press, as cited by ChainCatcher. The company’s chief financial officer said OpenAI is looking at a gradual shift away from billing based on consumption. The move is being explored as OpenAI faces competition from lower-priced open-source models. The remarks point to a possible change in how the company structures commercial access to its products, with pricing no longer centered solely on how much a user consumes. No timetable or additional details on the proposed model were disclosed in the brief report.

710