Open models

Liquid AI released LFM2.5-24B-A2B, a reasoning model with 24 billion total parameters, 2 billion active per token and a 131,072-token context window. The company reports 184 tokens per second on an Apple M5 Max and stronger results than larger rivals on several tool-use tests, while acknowledging weaker coding and knowledge scores. Downloadable weights and support for MLX and llama.cpp make it a practical option for private agents running locally.

Liquid AI ↗ · October 14 briefing · The frontier model race

AMD introduced Helion, an open rack design for agentic workloads powered by sixth-generation EPYC processors. Hyperscale configurations dedicate CPUs to agent execution, while enterprise racks combine CPUs with Instinct MI350P GPU inference nodes. An integrated token router directs model calls across frontier and open models, and partner systems are planned for November 2026.

AMD ↗ · October 14 briefing

Nvidia CEO Jensen Huang said in Paris that the company will unveil about $2 billion in new European AI investments within the next week, on top of roughly $8 billion it has put into the continent over the past year. In France, Nvidia is taking minority stakes in Doctolib, Pigment, Genesis AI and Entalpic, and Huang met President Emmanuel Macron to discuss investment, open-source models and AI safety. Huang argued that open models can be tailored to local languages and national sovereignty, saying Europe needs to get in the race. The tour continues to London and Berlin, where Nvidia holds its GTC developer conference next week.

Bloomberg Technology ↗ · October 13 briefing

French lab Mistral released Mistral Large 4, nicknamed Le Chonk, a multimodal mixture of experts model with 1.05 trillion total and 49 billion active parameters. It is in public preview through Mistral's API, with open weights planned by the end of October after safety testing. Mistral says it trained the model on about 3,800 Nvidia Grace Blackwell GPUs in its own European data centers, far less compute than Chinese rivals.

TechCrunch AI ↗ · October 6 briefing · The frontier model race

Google DeepMind launched EmbeddingGemma 2, a 740 million parameter open model that maps text, code, images, video and audio into a single shared embedding space. Released under the Apache 2.0 license, it is built on Gemma 4 and designed to run on consumer hardware. Google calls it its most capable model for on-device multimodal search and retrieval.

Unite. AI ↗ · October 6 briefing · The frontier model race

Reflection AI, founded by two former Google DeepMind researchers, released Beam, a 501 billion parameter open-weight model it pitches as America's answer to Chinese open-source AI. The company says Beam excels at coding and agentic tasks, scores close to leading Chinese open models, and runs on three to four times less compute than comparable Western open models. Valued at $25 billion, Reflection has a $6.3 billion compute deal with SpaceX, supply agreements with the US Department of Energy and Department of War, and is already training a larger successor.

Fortune AI ↗ · October 5 briefing · The frontier model race

AI inference infrastructure provider Modal Labs is closing in on a $750 million round led by Accel that would value the company at $15.75 billion, more than tripling its $4.65 billion valuation from four months ago. Modal, founded in 2021 by CEO Erik Bernhardsson and CTO Akshat Bubna, lets developers run AI models and compute heavy workloads without managing their own servers, and had surpassed $300 million in annualized revenue as of May. The rapid rise reflects surging demand for inference services as companies increasingly run open source models in production.

TechCrunch AI ↗ · September 28 briefing · Follow the money

The Caltech-founded startup unveiled a 1-bit language model called Bonsai, tuned for Qualcomm's Snapdragon AR1 Gen 1 platform used in smart glasses. The 2 billion parameter model runs locally on-device, letting wearers ask what they are looking at in real time using both vision and language. It compresses larger models roughly 4 times over while preserving performance, aiming to offer open-weight AI that runs on-device as an alternative to cloud-dependent assistants. No smart glasses product using the technology has been announced yet.

TechCrunch AI ↗ · September 24 briefing

Legal AI startup Harvey, valued at $15.6 billion, saw its gross margins collapse from about 50 percent to negative 50 percent by June as token usage under OpenAI and Anthropic's enterprise pricing rose twentyfold. Margins turned positive after Harvey launched its own in house model, post trained on Moonshot's open weight Kimi K3. Other startups including Abridge and Ramp are making similar moves toward cheaper open weight models as compute costs squeeze their businesses.

Bloomberg Technology ↗ · September 21 briefing

Base Labs, the AI safety research arm spun out of Baseten earlier this year, launched a partnership with Hugging Face and Goodfire to build safety evaluation and monitoring infrastructure for open weight models. Goodfire is contributing its model interpretability expertise to the effort. The group plans to develop and publish methods for training and monitoring open models so that safety features are built in from the start rather than added afterward.

TechCrunch AI ↗ · September 17 briefing

Arcee AI closed a $150 million Series B at a $1 billion pre money valuation to keep building its Trinity family of open weight models, including a 400 billion parameter Trinity Large model with 13 billion active parameters per token. The round was led by Vista Equity Partners, Cambium Capital, and Emergence Capital. The company says its entire 2025 model lineup cost about $20 million to build and will use the funding to expand work with US national laboratories.

Fortune ↗ · September 16 briefing · Follow the money

Pizza Bot is a self-hosted, open-source desktop app for Mac, Windows, and Linux that runs AI agents in the background and surfaces their output through an email-like inbox, with finished tasks in Unread and pending approvals in Action. It works with any model provider, including Anthropic, Amazon Bedrock, Google Gemini, OpenAI, OpenRouter, or a local Ollama model. The Apache 2.0 licensed project began as an internal Amazon tool to automate CRM data entry.

Hacker News ↗ · September 15 briefing

Cognition, the company behind the Devin coding agent, released SWE-2, its most capable coding model to date, post-trained with reinforcement learning on top of Moonshot AI's 2.8 trillion parameter open model, Kimi K3. On a leading coding benchmark, SWE-2 scored within one point of Fable 5.1 while costing 64 percent less to run. The release signals a broader trend of startups fine-tuning open weight Chinese models to compete with closed frontier labs on price.

AI Research News ↗ · September 13 briefing · The frontier model race · AI that writes code

Moonshot AI is targeting $2 billion in annualized revenue by the end of 2026, roughly doubling its August run rate, riding strong demand for its open weight Kimi K3 model, which generates about 300 billion tokens a day on OpenRouter. The target still trails far behind closed model rivals, with OpenAI reporting around $40 billion in annualized revenue and Anthropic near $65 billion. Because K3's weights are public, Moonshot operates on thinner margins than competitors selling proprietary access, making the revenue goal an aggressive bet on volume.

TechCrunch AI ↗ · September 13 briefing

Leading law firm Latham Watkins has purchased Nvidia servers to build its own in-house AI systems, moving away from dependency on commercial offerings from OpenAI and Anthropic. The firm is customizing open-weight foundation models to handle legal work rather than relying on closed commercial APIs, reducing costs and security risks from external dependencies. This trend among large enterprises suggests a broader shift toward building proprietary AI infrastructure to capture AI value internally.

Financial Times Tech ↗ · September 11 briefing

Databricks-trained AI agents are achieving answer quality comparable to Anthropic's Claude and OpenAI's GPT-5.6 Luna models while completing tasks in roughly half the time. This breakthrough demonstrates that open-weight models trained on commodity hardware can match or exceed the performance of expensive proprietary frontier models in many practical applications. The result challenges the assumption that only the best-resourced companies can build competitive large-scale AI systems.

VentureBeat AI ↗ · September 11 briefing

Mistral achieved a $24 billion valuation in a new funding round led by Samsung, establishing itself as Europe's leading AI competitor. The company is betting on open-weight AI models as its strategy to compete with OpenAI and Anthropic in the global AI market. Samsung's investment signals growing interest from device makers in securing access to foundational AI models.

CNBC Technology ↗ · September 9 briefing · Follow the money

Nvidia agreed to acquire Hugging Face, the popular open-source AI platform, in a $13 billion transaction that represents Nvidia's second-largest acquisition ever. The deal gives Nvidia control of a thriving developer community and the world's leading repository for open-source AI models and datasets. This acquisition accelerates Nvidia's vertical integration strategy, combining its hardware dominance with software and platform control to lock in enterprise customers.

Bloomberg Technology ↗ · September 3 briefing · Follow the money

Perplexity launched hybrid compute capabilities for its Computer platform, enabling a single AI agent to dynamically split work between cloud-based frontier models and local open-weight models running on Apple Silicon, keeping sensitive data on-device. The system intelligently routes confidential information to local processing while using cloud models for general web research and reasoning, solving a critical enterprise privacy concern. This hybrid approach marks a major shift in how enterprises can use AI without compromising data security.

VentureBeat AI ↗ · September 2 briefing

Perplexity today launched hybrid compute for its agentic platform, Computer, a system that lets a single AI agent split its work between frontier models running in the cloud and smaller open-weight models running locally on Apple silicon Macs — routing sensitive data to the local machine so it never leaves the device. The company says it is the first time an AI agent can begin a task in the cloud and dynamically hand off the confidential portions of that same task to a model running on the user's own hardware, without restarting the job or losing context. The feature becomes available today through Perplexity's desktop app for enterprise customers that opt in, as well as Pro and Max subscribers, on any Apple silicon Mac running macOS 15 or later."Hybrid is really compelling because it's often the work that requires confidentiality that is the most important to get right, and so the accuracy really, really matters," Jon Staff, who leads Perplexity's macOS and iOS engineering teams, said during a press briefing attended by VentureBeat. "By combining these two together, we can get that maximum intelligence from the frontier models, but we also get the security and the privacy that comes with local."How Perplexity's on-device privacy gate keeps sensitive data off the cloudThe architecture works like a dispatcher. A frontier model in the cloud breaks a task into subtasks and routes each one to the appropriate place. Web research, long-ho...

VentureBeat ↗ · September 1 briefing

Hugging Face is exploring a sale that may value the AI platform at $13 billion or more, according to Business Insider, as consolidation pressures mount in the open-weights ecosystem.

Bloomberg Technology ↗ · August 23 briefing

Google DeepMind's Gemma family of open models has passed one billion cumulative downloads, the company said on August 20, 2026, capping two years in which outside developers published more than 100,000 variants built on the models' open weights. The figures, disclosed in Google's official announcement by Clement Farabet, a vice president at Google DeepMind, and product director Olivier Lacombe, mark the first time Google has put a cumulative total on Gemma adoption since the family launched in...

Unite. AI ↗ · August 20 briefing

Chinese startup z.ai put its new frontier model GLM-5.3 on its API at $1.40 per million input tokens and $4.40 per million output tokens, with cached input at $0.26 and cache storage free for a limited period. Artificial Analysis scored the model 60 on its Intelligence Index, seven points above GLM-5.2 and tied with Kimi K3 as the strongest open-weight model available, with sharp gains on agentic coding and cyber benchmarks including Terminal-Bench 3.0. Weights have been promised but not released, so for now the model is API-only rather than genuinely open source.

VentureBeat ↗ · August 19 briefing

Alibaba released Qwen3.8-27B, a model small enough to run on a consumer laptop, and at the same time published the open weights for Qwen3.8 Max, its most powerful system. The company says the 27B model matches a model ten times its size on coding, professional tasks, research and multi-step agentic workflows. The launch lands one week after Meta open-sourced its strongest model and shipped laptop-class versions, putting the two rivals in direct competition for the same hardware. Hugging Face now counts 151,448 Qwen derivative models, roughly 2.6 times Meta's footprint.

CNBC Technology ↗ · August 17 briefing

Researchers from Japan's National Institute of Advanced Industrial Science and Technology, the University of Tsukuba, the University of Technology Nuremberg and the University of Oxford found that changing nothing but text color shifts how vision language models read meaning. Testing four open-source models, they measured sentiment swings of up to +0.42 when positive words were colored green or blue and -0.48 when negative words were colored red, with Qwen2-VL-7B the most susceptible and IDEFICS2-8B showing a 0.52 total range. In visual question answering tests, highlighting a wrong decoy answer through contrast alone pushed error rates up, and the effect grew as the surrounding text was made harder to see. The authors call visual styling a critical and previously underexplored vulnerability in model pipelines.

Unite. AI ↗ · August 17 briefing

Nvidia released Nemotron 3.5 Lightning, a 30 billion parameter mixture of experts model that companies can download, modify and deploy for free without asking permission. Nvidia claims it delivers up to four times faster output and 30 percent faster task completion than comparable models in its class, and it runs on a single PC GPU including RTX PCs and DGX Spark systems as well as across data centers. Alongside it, Nvidia open sourced Switchyard, a model routing library that automatically directs each step of an agent workflow to the cheapest suitable model. CrowdStrike, CodeRabbit and Harvey were among the companies that tested and customized the model ahead of launch.

CNBC Technology ↗ · August 11 briefing · The frontier model race

Mistral AI announced a three part expansion of its infrastructure business aimed at turning European AI sovereignty into a purchasable product. It is adding regional inference endpoints that let customers choose whether workloads run in Europe or the United States, a Priority Tier backed by an uptime guarantee for mission critical deployments, and a coalition of European enterprises signing multi year compute commitments. Mistral says those commitments will underwrite 200 megawatts of European infrastructure by the end of 2027 and a full gigawatt by the end of 2030. In a move likely to unsettle sovereignty purists, the company will also start hosting third party open models, beginning with GLM-5.2 from the Chinese lab Z.ai.

VentureBeat ↗ · August 11 briefing · The compute buildout

River AI, founded by xAI co founder Igor Babuschkin and only two months out of stealth, raised $1.1 billion led by General Catalyst and AMP PBC with participation from Nvidia, AMD Ventures, Y Combinator and Temasek. The startup builds infrastructure for personal AI agents, offering an API that lets developers fine tune open weight models using reinforcement learning and LoRA. River claims an enterprise can finish a complex reinforcement learning run in 15 to 20 minutes with no infrastructure team, at two to four times the cost savings versus closed source alternatives. Babuschkin previously worked at DeepMind and OpenAI before co founding xAI.

TechCrunch AI ↗ · August 11 briefing

LTX, the open world model company spun out of Lightricks, released LTX-2.5, an open weights video and world model that generates a 10 second clip from an image in 6.8 seconds on Nvidia superchips. The release adds multi shot support, a diffusion decoder for higher quality, new conditioning modes and better support for the autoregressive models needed for real time use and robotics. Weights are available now on Hugging Face and through the LTX API, free for organizations under $10 million in annual recurring revenue, with larger companies negotiating a license. LTX says its models have passed 33 million downloads, making it the most used open world model line on the market.

VentureBeat ↗ · August 11 briefing

Meta released Muse Glimmer, a 30 billion parameter open weight model built to run autonomous AI agents directly on consumer hardware rather than in the cloud, with weights available on Hugging Face now. It ships under the permissive Apache 2.0 license, Meta's first fully open release since it moved to the proprietary Muse Spark in April, and it drops the 700 million monthly user restriction that drew years of criticism of the Llama community license. The model is quantized to roughly 4 bits and uses block level speculative decoding so it runs fast enough inside a real agent loop on a single consumer GPU or a Mac, with Meta partnering with AMD, Arm, Dell, Intel and Nvidia on device optimization. The launch landed alongside a Mark Zuckerberg essay, The Future Is for Everyone, arguing that superintelligence controlled by a handful of companies or governments would produce worse outcomes for everyone else.

VentureBeat ↗ · August 10 briefing

Google DeepMind published its WeatherNext Cyclones model in Nature and open sourced the code and weights on GitHub under an Apache 2.0 license. The model's three-day cyclone forecasts are as accurate as what previous systems could manage only two days out, effectively handing forecasters an extra day of warning. During the 2025 season it helped the National Hurricane Center predict Hurricane Melissa's rapid intensification and Jamaica landfall in time for an early advisory. Google released three variants, including a mini version small enough to run in a free Colab notebook.

Ars Technica ↗ · August 9 briefing

Liquid AI debuted LFM2.5-2.6B, an open-weight language model designed for agentic workloads that can run entirely on local hardware from smartphones and laptops down to a Raspberry Pi, without cloud inference or GPUs.

VentureBeat ↗ · August 6 briefing · The compute buildout

Mistral AI released Shieldstral, a 3 billion parameter open weights safety classifier that screens both text and images against moderation policies written in plain language at inference time, rather than against fixed harm categories baked in during training. The model is available on Hugging Face under the Apache 2.0 license, covers 12 languages and runs on a single 16GB GPU. Mistral says it matches open guard models up to seven times its size on text safety benchmarks.

Unite. AI ↗ · August 4 briefing

OpenAI said an internal version of Astra, its next major model family, produced new results for 10 problems in mathematics and theoretical computer science that had been open for at least a decade, including Alain Connes's rigidity conjecture, Ehrhart's volume conjecture and three problems from Paul Erdos's list. The headline result is an explicit construction of a non sofic group, a question left open since Mikhail Gromov introduced soficity in 1999. OpenAI published a 249 page manuscript collection alongside Lean 4 proof certificates on GitHub under an Apache 2.0 license, with a sorry count of zero, meaning no step in any formalized proof was left unproven. The company put the compute cost of finding the solutions at roughly $2,000 at Sol API rates.

Forbes Tech ↗ · August 3 briefing

Alibaba unveiled Qwen3.8-Max on Monday, a 2.4 trillion parameter mixture of experts model that activates about 95 billion parameters per query and handles text, images and video with a one million token context window. On the Arena. AI leaderboard it became the highest ranked Chinese model for text tasks and placed second globally on vision, trailing only Anthropic's Claude Fable 5. Alibaba shares rose roughly 4.5 percent in New York premarket trading and about 7 percent in Hong Kong. The model is available through Alibaba Cloud Model Studio, with open weights due to follow next week.

CNBC Technology ↗ · August 3 briefing

Cisco's Foundation AI team has fingerprinted nearly 900 open models on Hugging Face to verify their true lineage, replacing the self-reported base model tags that uploaders type in and that are never substantiated through weight level analysis. The move responds to a growing provenance gap, as one report found that Alibaba's Qwen family is the declared parent of 69 percent of new open model derivatives as of early 2026, with Chinese labs accounting for about 70 percent overall. By offering free weight level verification, Cisco aims to replace unverified trust in a supply chain where security teams currently approve models based on unproven tags.

VentureBeat ↗ · July 30 briefing

Nvidia launched the Open Secure AI Alliance, a coalition of more than 28 companies including Adobe, Microsoft, IBM, Salesforce, and Palantir, to build open source tools for securing AI systems. The group formed in direct response to the recent Hugging Face breach, where members argued that closed AI models blocked essential forensics while an open weight model helped contain the intrusion. Early contributions include Nvidia's agent security framework, Hewlett Packard Enterprise zero trust identity standards, and a Microsoft vulnerability scanning tool, though OpenAI and Anthropic are notably absent from the alliance.

Forbes Tech ↗ · July 28 briefing

China warned it would take all necessary measures if the US moves to sanction Chinese AI companies, with the Ministry of Commerce urging Washington to end what it called a smear campaign over allegations of intellectual-property theft. The threat followed US Treasury Secretary Scott Bessent's warning that the administration would scrutinize Chinese open-weight models for industrial-scale distillation of American AI systems. The flashpoint is Moonshot AI's Kimi K3, which US officials are examining for whether it improperly extracted capabilities from a leading US model.

Bloomberg Technology ↗ · July 27 briefing

Nvidia, Microsoft, Meta and more than 20 other American technology companies signed a joint letter urging Washington not to impose premature restrictions on open-weight AI models. The signatories, which also include Dell, IBM, Palantir, Mistral and the Linux Foundation, argue the US should build a strong domestic open ecosystem rather than cede the field to China. The letter follows alarm over Moonshot's Chinese Kimi K3 model, which has topped some benchmarks and which US officials accuse of being distilled from Anthropic's Fable 5. Notably, OpenAI, Anthropic and Google did not sign.

CNBC Technology ↗ · July 26 briefing

Hugging Face's production infrastructure was compromised after a malicious dataset exploited code execution paths on processing workers, letting an autonomous AI agent escalate privileges, harvest cloud credentials, and self-migrate through internal systems while creating decoy activity to cover its tracks. Hugging Face turned to Z.ai's open-weight GLM-5.2 model for forensic analysis after frontier model safety guardrails blocked its own defenders' incident-response queries. CEO Clement Delangue said attackers are already using AI agents and that this will not be stopped by locking models behind APIs, and experts warn organizations now need their own AI-driven incident response capabilities and vetted self-hosted models.

Forbes Tech ↗ · July 21 briefing

Poolside, a San Francisco AI lab that has mostly sold coding models to governments and defense agencies over its three-year history, released Laguna S 2.1, a 118 billion parameter Mixture of Experts model that activates only 8 billion parameters per token and supports context windows up to 1 million tokens. The weights are freely available on Hugging Face under the permissive OpenMDW-1.1 license. Poolside reports the model scores 70.2% on Terminal-Bench 2.1, beating rivals many times its size including DeepSeek-V4-Pro-Max and Nvidia's Nemotron 3 Ultra, and 78.5% on SWE-Bench Multilingual. The model went from the start of training on May 22 to public launch in under nine weeks, using 4,096 Nvidia H200 GPUs.

VentureBeat ↗ · July 21 briefing · AI that writes code

Bloomberg reports that Moonshot AI's Kimi K3, a 2.8 trillion parameter open weight model released last week, may matter more for memory demand than for raw computing power. The model uses a hybrid attention architecture that routes each token to 16 of its 896 experts and claims roughly 2.5 times the scaling efficiency of its predecessor, yet its scale still demands far more memory capacity than earlier generations, pointing to sustained demand for chipmakers such as SK Hynix and Samsung Electronics. Traders compared the release to DeepSeek's R1 debut in early 2025, when almost $600 billion was wiped from Nvidia's market value in a single day on fears that AI would require less computing power than expected. Kimi K3's release triggered a similar reaction, pushing semiconductor stocks sharply lower on Friday.

Bloomberg Technology ↗ · July 20 briefing

OpenAI chair Bret Taylor told CNBC he expects concerns over AI token costs to fade as the market matures, predicting that companies will shift toward paying for outcomes rather than for tokens consumed. Taylor said much of today's tokenomics frustration stems from an immature market, and that within a year CFOs will stop thinking about tokens at all as other companies absorb that complexity. His comments came amid renewed scrutiny of AI costs following Moonshot AI's launch of the Kimi K3 model, though he argued frontier models remain more token efficient than open weight alternatives. Taylor pointed to his own startup, Sierra, which already charges businesses per outcome, such as per loan origination or medical pre-authorization, as an early example of the shift.

Business Insider Tech ↗ · July 20 briefing

Chinese labs like Moonshot AI are cornering the market for cheap, customizable intelligence with open-weight models, threatening to turn America's prestige models into expensive niche products. On OpenRouter, Chinese models now occupy the top five spots by weekly token usage, and businesses are shifting away from premium AI models toward cheaper alternatives.

Axios ↗ · July 19 briefing

Databricks announced a new funding round led by Coatue that values the company at $188 billion, up from $134 billion in February, as it has successfully transitioned its image from big data provider to AI company and championed cost-saving open-weight models.

TechCrunch AI ↗ · July 19 briefing · Follow the money

Moonshot AI, the Beijing-based startup backed by Alibaba, released Kimi K3, a 2.8-trillion-parameter open-source model that the company says is the largest ever released, with benchmarks showing it rivals top proprietary systems from Anthropic and OpenAI. The model features a 1-million-token context window, native visual understanding, and an always-on reasoning mode, with full weights scheduled for release on July 27. On benchmarks such as GDPval-AA v2 and AA-Briefcase, K3 placed third and second respectively, while topping the BrowseComp long-horizon information-seeking benchmark and ranking first on Arena. AI's Frontend Code Arena.

VentureBeat ↗ · July 16 briefing · The frontier model race

Thinking Machines, the AI startup founded by former OpenAI CTO Mira Murati, released Inkling, its first major language model, under an enterprise-friendly Apache 2.0 open source license. The 975 billion parameter mixture-of-experts system reasons natively across text, images and audio, scoring 77.6% on SWE-bench Verified to beat Nvidia's Nemotron 3 among US open-weights rivals, and 91.4% on a leading voice understanding benchmark. The company says Inkling was designed to answer directly on topics that may be subject to censorship, and it balances cost against performance through a controllable thinking effort mechanism. The weights are already available on Hugging Face and the company's Tinker API.

VentureBeat ↗ · July 15 briefing

At the VB Transform 2026 conference in Menlo Park, Rachad Alao, vice president of product engineering at Canadian enterprise AI startup Cohere, argued that AI sovereignty means far more than downloading an open model or running an app behind a corporate firewall. For banks, hospitals and governments running mission-critical systems, he said organizations need tight control over where data resides and jurisdiction over the AI itself, extending from GPUs and private-cloud infrastructure through the governance systems, connectors, search tools and agent frameworks acting on enterprise data. Alao, who previously led responsible AI and trust and safety engineering at Google and Meta, warned that surging agent workloads could outrun falling token prices.

VentureBeat ↗ · July 15 briefing

The startup says it compressed Alibaba's Qwen3.6 27B model from roughly 54GB to under 4GB, letting its new Bonsai 27B model run fully on-device on Apple hardware via MLX at around 11 tokens per second, and its CEO says Apple is evaluating the technology. The compressed model reportedly retains more than 95% of full-precision benchmark performance while using up to 15 times less memory. The weights are available under an Apache 2.0 license, and a deal could accelerate Apple's push to run powerful AI locally on its devices.

CNBC Technology ↗ · July 14 briefing

Open source AI is booming, according to Hugging Face CEO Clem Delangue. The company has grown into something like a GitHub for AI, where builders can share and download open models and datasets, and it is now used by roughly half the Fortune 500. Delangue said he has seen the same pattern repeat across companies, which often start out renting AI capabilities before eventually building and owning more of their own AI infrastructure.

TechCrunch AI ↗ · July 12 briefing

Zhipu founder Tang Jie argued in an internal staff memo reviewed by Bloomberg that frontier artificial intelligence should remain broadly accessible rather than controlled by a select few, saying genuine security comes from broad participation and oversight rather than technological barriers. In line with that view, Zhipu released its latest model, GLM 5.2, under an open source license free for users to download and commercialize. The comments came shortly after Reuters reported that Beijing was weighing measures to limit overseas access to China's most advanced open source AI models, putting Tang's stance at odds with potential government restrictions.

Bloomberg Technology ↗ · July 12 briefing

Open source AI tool Ollama raised a $65 million Series B led by Theory Ventures, adding to a $15 million Series A led by Benchmark's Peter Fenton for $88 million raised total. The company, which lets developers run open-weight AI models on personal computers, now counts 8.9 million monthly active developers, 176,000 GitHub stars, and roughly 17,000 forks, and says it is used at 85 percent of Fortune 500 companies. Ollama runs with just 14 employees and also offers a paid cloud tier priced by GPU usage rather than token limits.

TechCrunch AI ↗ · July 9 briefing · Follow the money · AI that writes code

A new in-depth profile breaks down how the French startup, founded in 2023, has become Europe's leading rival to OpenAI, most recently raising about 1.7 billion euros at an 11.7 billion euro valuation in a round led by ASML, with reports it is now raising roughly $3.5 billion at a $23 billion valuation. The company spans open-weight models, its Vibe chat platform, and enterprise tools, with annual recurring revenue reportedly on track to pass $1 billion. A new open-weight model is slated for summer 2026, and CEO says the company is aiming for an IPO rather than a sale.

TechCrunch AI ↗ · July 5 briefing

A June 12 U. S. export-control order pulled Anthropic's Claude Fable 5, the most capable model on the market, offline for every customer with no warning, and it returned this week wrapped in tighter safeguards after China's Z.ai released its open-weights GLM-5.2 into the vacuum. New survey data covering 145 enterprises shows two-thirds had already hedged their model strategy before the order came down: 51% blend closed frontier models with open-weight models on their own infrastructure, and another 16% are moving core workflows off closed APIs entirely. The blackout also exposed a deeper gap, as just 1 in 10 enterprises has automated monitoring that would catch an AI model drifting, misbehaving, or failing in production.

VentureBeat ↗ · July 5 briefing

Even as the geopolitical conversation around AI continues to grow more fraught following the U. S. government's actions to limit the new models from Anthropic and OpenAI, Chinese open source darling DeepSeek is back with yet another open release that could once again change AI development around the globe. Over the weekend, the firm released DSpark, a new, MIT-Licensed system designed to make large language models answer faster without changing what the underlying model is trying to say. The easiest way to think about it is this: most AI chatbots write like someone crossing a river one stepping stone at a time. They choose one small chunk of text, then the next, then the next. DSpark gives the system a scout that runs a few steps ahead, guesses the likely path, and lets the larger model quickly check which steps are safe. When the guesses are good, the model moves faster. When the guesses are weak, DSpark tries not to waste time checking them. DeepSeek published the work with a technical paper, model checkpoints and DeepSpec, a codebase for training and evaluating speculative decoding systems. The release is available through DeepSeek’s public GitHub and Hugging Face pages, both under the permissive, friendly, commonplace MIT license, making the new technique broadly usable by developers, researchers and commercial enterprise operations that want to study or adapt the approach. The system is aimed at one of the most expensive problems in AI deployment: serving large models ...

VentureBeat ↗ · June 29 briefing

Washington is racing to sell the world on American AI just as China's cheap and capable models are becoming harder to ignore. Why it matters: Chinese models don't have to beat OpenAI or Anthropic to reshape the global AI order. They just have to be useful, available and widely adopted. Between the lines: Experts argue that two key things are kneecapping the U. S. government's desire to export American AI: An erratic export controls strategy that involves making decisions about access to advanced models on the fly. Not paying sufficient attention to China's efforts to spread its open-source AI models abroad while deploying AI at scale across manufacturing, health care and other industries domestically. Driving the news: The State Department this week expanded its Pax Silica initiative, an effort to build a U. S.-led AI and chip supply-chain bloc and reduce reliance on Chinese technology. The move came as the AI industry continues to grapple with the fallout of the government's decision to put export controls on Anthropic's newest AI models. Meanwhile, Chinese AI is closing the capability gap, as Axios' Sam Sabin has reported, and the models are much cheaper. What they're saying: "I think we're seeing another example of the Huawei strategy in the context of open-source AI models," Emily Weinstein, a former Commerce Department staffer now at the U. S.-China Economic and Security Review Commission, told Axios."China is able to offer not even just the models, but often the underlying or ass...

Axios ↗ · June 28 briefing

GLM-5.2 — the latest Chinese open-source model capturing Silicon Valley's attention — is raising fresh concerns among security researchers that advanced AI hacking capabilities are becoming dramatically cheaper and more accessible. Why it matters: The barrier to entry for malicious hackers eager to automate and personalize their attacks is getting lower and lower. Driving the news: Z.ai's GLM-5.2, which was released last week, has agentic capabilities that rival those of Claude Opus 4.8 and OpenAI's GPT-5.5 while costing roughly half as much to run. Two separate security evaluations from Graphistry and Semgrep found that GLM-5.2 performed on par with leading U. S. models on cybersecurity investigation and vulnerability-discovery benchmarks. Researchers at Graphistry also suggested that GLM-5.2 may be an "illegal distillation of both GPT-5.5 and Opus 4.8" — a claim that, if true, could help explain how Chinese models have been rapidly narrowing the gap with U. S. competitors. Z.ai did not respond to a request for comment. The big picture: Unlike Claude or ChatGPT, open-weight models like GLM-5.2 can be downloaded and modified directly, allowing users to remove safety controls, fine-tune them for specific tasks, and operate them without relying on a commercial provider. Graphistry said GLM-5.2 is the first open-weight model it has tested that it would recommend for a "frontier-like" cybersecurity experience. Threat level: Hackers are already talking in Russian-language forums about how ...

Axios ↗ · June 25 briefing

GLM-5.2, a new open-source model from Z.ai (formerly Zhipu AI), is drawing serious attention from Silicon Valley engineers for its performance on long-form coding tasks, with benchmarks placing it competitive with GPT-class models. The Latent Space newsletter reports Z.ai is forecasting an open-source Fable-equivalent model by December 2026, intensifying the race between Chinese labs and Western frontier models. The release lands at a moment when US export controls on Anthropic's models are heightening scrutiny of the open-source Chinese AI pipeline.

Business Insider Tech ↗ · June 21 briefing

Chinese startup Z.ai (formerly Zhipu AI) released GLM-5.2, a 753B-parameter open-weights model targeting long-horizon coding with a 1M-token context window and API tiers starting at $12.60 per month.

VentureBeat ↗ · June 16 briefing · AI that writes code

NVIDIA released Nemotron 3 Ultra, a 550-billion-parameter open mixture-of-experts model built for long-running agents with a one-million-token context. The release ships open weights, training data, and recipes, with NVIDIA claiming up to sixfold higher inference throughput than comparable open LLMs.

MarkTechPost ↗ · June 4 briefing

While many AI open source model providers are pursuing larger and more powerful models, Google is still giving attention to the smaller, more local side of the market. Today, the tech giant released Gemma 4 12B, an 11.95-billion-parameter open-weights model with permissive Apache 2.0 license optimized to execute locally on a standard enterprise laptop using just 16GB of VRAM or unified memory. That means those enterprise users looking to keep working with AI while on a flight without WiFi, or try...

VentureBeat ↗ · June 3 briefing

JetBrains releases Mellum2 under Apache 2.0 — a 12B MoE model trained on 10.6 trillion tokens for AI workflows. The post JetBrains Releases Mellum2: A 12B MoE Model for Fast, Specialized Tasks in Multi-Model AI Pipelines appeared first on MarkTechPost .

Planet AI (Aggregator) ↗ · June 2 briefing · The frontier model race

Canadian AI lab Cohere made waves recently by announcing a merger with German AI startup Aleph Alpha, but now it has even more in store for enterprise builders around the globe: today, the firm co-founded by former Googler and "Attention Is All You Need" co-author Aidan Gomez unveiled Command A+, a highly optimized, 218-billion-parameter language model engineered specifically for complex reasoning, multimodal document processing, and agentic workflows. The most significant aspect of the...

VentureBeat ↗ · May 20 briefing

An eventful month with one flagship release after another

Interconnects ↗ · May 17 briefing

Pinterest uses a multimodal generative AI strategy to lower computing costs. Their approach includes OpenAI's and Alibaba's large language models.

Business Insider Tech ↗ · April 30 briefing

Moonshot open-sources Kimi K2.6 with claims of stronger long-horizon coding under a modified MIT-style license.

TechMeme ↗ · April 20 briefing · The frontier model race · AI that writes code

Iain Martin /Forbes:Mistral, which once aimed for top open models, now leans on being an alternative to Chinese and US labs, says it's on track for $80M in monthly revenue by Dec.— Paris-based Mistral wanted to develop a top-tier AI model to rival OpenAI and Anthropic. That didn't work out.

TechMeme ↗ · April 19 briefing

The new offering is designed for advanced reasoning and multimodal capabilities.

AI Business ↗ · April 5 briefing · The frontier model race

Mistral's sovereign AI stack, investment in data center capacity and open-weight frontier models provide CIOs with an alternative to U. S. proprietary AI models.

AI Business ↗ · March 18 briefing

Nvidia is planning to spend $26 billion on open-weight AI models, according to filings. The move could position the AI infrastructure powerhouse to quickly compete with OpenAI, Anthropic, and DeepSeek.

Wired ↗ · March 11 briefing

NVIDIA Nemotron 3 Super is a 120-billion-parameter open model with 12 billion active parameters designed to run complex agentic AI systems at scale. Perplexity offers users access to the model.

NVIDIA AI Blog ↗ · March 11 briefing

Microsoft has released a 15 billion parameter open-weight multimodal model designed for image and text tasks requiring perception and selective reasoning. The model excels at scientific and mathematical reasoning and GUI understanding while balancing compute efficiency.

AI Research News ↗ · March 8 briefing · The frontier model race

Indian AI startup Sarvam has released a new family of open-source models specifically optimized for edge devices, including cars, smart glasses, and even low-cost feature phones, targeting the next billion AI users.

TechCrunch AI ↗ · February 19 briefing

Z.ai released GLM-5, claiming best-in-class performance for reasoning, coding, and long-horizon agentic tasks among open-source models.

MarkTechPost ↗ · February 11 briefing · The frontier model race

ByteDance has introduced Protenix-v1, a comprehensive AlphaFold3 (AF3) reproduction for biomolecular structure prediction, released with code and model parameters under Apache 2.0.

AI Research News ↗ · February 8 briefing · The frontier model race

The Qwen team released Qwen3-Coder-Next, an 80B parameter sparse MoE model designed specifically for coding agents and local development, matching top-tier proprietary performance.

AI Research News ↗ · February 5 briefing · The frontier model race · AI that writes code

Amid investor frenzy over AI, the ‘anti Silicon Valley’ group chose independence and desire to promote open models

Financial Times Tech ↗ · January 28 briefing

The company said that the model was trained on 15 trillion mixed visual and text tokens.

TechCrunch AI ↗ · January 27 briefing · The frontier model race · AI that writes code

Z.ai:Chinese AI startup Z.ai releases GLM-4.7, an open-weight model that Z.ai says delivers significant improvements in coding performance compared to GLM-4.6— - Core Coding: GLM-4.7 brings clear gains, compared to its predecessor GLM-4.6, in multilingual agentic coding and terminal-based tasks …

TechMeme ↗ · December 22 briefing · The frontier model race · AI that writes code

NVIDIA has released the Nemotron 3 family of open models as part of a full stack for agentic AI, including model weights, datasets and reinforcement learning tools. The family has three sizes, Nano, Super and Ultra, and targets multi agent systems that need long context reasoning with tight control over inference cost. Nano has about […]The postNVIDIA AI Releases Nemotron 3: A Hybrid Mamba Transformer MoE Stack for Long Context Agentic AIappeared first onMarkTechPost.

AI Research News ↗ · December 21 briefing

The company's Starcloud-1 satellite is running Gemma, an open model from Google, marking the first time in history that an LLM has been trained in outer space.

CNBC Technology ↗ · December 10 briefing · The compute buildout

The partners will use Nvidia's platforms to optimize Mistral's open-source model family.

AI Business ↗ · December 7 briefing

Essential AI, whose CEO co-wrote Google's Attention Is All You Need paper, unveils Rnj-1, an 8B-parameter open model with SWE-bench performance close to GPT-4o. The long-term advancement and equitable diffusion of AI technologies crucially depend on their development in the Open.

TechMeme ↗ · December 7 briefing · The frontier model race

Researchers worldwide rely on open-source technologies as the foundation of their work. To equip the community with the latest advancements in digital and physical AI, NVIDIA is further expanding its collection of open AI models, datasets and tools — with potential applications in virtually every research field. At NeurIPS, one of the world's top AI conferences, NVIDIA announced major open-source AI developments.

NVIDIA AI Blog ↗ · December 1 briefing · Robots and physical AI

DeepSeek AI has released DeepSeekMath-V2, an open weights large language model that scored 118 out of 120 on the Putnam 2024 competition, optimized for natural language theorem proving with self verification.

AI Research News ↗ · November 30 briefing

OpenAI released gpt-oss-120b and gpt-oss-20b, its first open-weight language models in more than five years, under the Apache 2.0 license. The models deliver strong reasoning and tool-use capabilities and can be deployed privately on consumer hardware without sending data to the cloud.

VentureBeat ↗ · August 6 briefing