News from the AI & ML world
DeeperML - #opensourceai
|
@github.com
//
Google's Magenta project has unveiled Magenta RealTime (Magenta RT), an open-weights live music model designed for interactive music creation, control, and performance. This innovative model builds upon Google DeepMind's research in real-time generative music, providing opportunities for unprecedented live music exploration. Magenta RT is a significant advancement in AI-driven music technology, offering capabilities for both skill-gap accessibility and enhancement of existing musical practices. As an open-weights model, Magenta RT is targeted towards eventually running locally on consumer hardware, showcasing Google's commitment to democratizing AI music creation tools.
Magenta RT, an 800 million parameter autoregressive transformer model, was trained on approximately 190,000 hours of instrumental stock music. It leverages SpectroStream for high-fidelity audio (48kHz stereo) and a newly developed MusicCoCa embedding model, inspired by MuLan and CoCa. This combination allows users to dynamically shape and morph music styles in real-time by manipulating style embeddings, effectively blending various musical styles, instruments, and attributes. The model code is available on Github and the weights are available on Google Cloud Storage and Hugging Face under permissive licenses with some additional bespoke terms. Magenta RT operates by generating music in sequential chunks, conditioned on both previous audio output and style embeddings. This approach enables the creation of interactive soundscapes for performances and virtual spaces. Impressively, the model achieves a real-time factor of 1.6 on a Colab free-tier TPU (v2-8 TPU), generating two seconds of audio in just 1.25 seconds. This technology unlocks the potential to explore entirely new musical landscapes, experiment with never-before-heard instrument combinations, and craft unique sonic textures, ultimately fostering innovative forms of musical expression and performance. Recommended read:
References :
@www.analyticsvidhya.com
//
MiniMaxAI, a Chinese AI company, has launched MiniMax-M1, a large-scale open-source reasoning model, marking a significant step in the open-source AI landscape. Released on the first day of the "MiniMaxWeek" event, MiniMax-M1 is designed to compete with leading models like OpenAI's o3, Claude 4, and DeepSeek-R1. Alongside the model, MiniMax has released a beta version of an agent capable of running code, building applications, and creating presentations. MiniMax-M1 presents a flexible option for organizations looking to experiment with or scale up advanced AI capabilities while managing costs.
MiniMax-M1 boasts a 1 million token context window and utilizes a new, highly efficient reinforcement learning technique. The model comes in two variants, MiniMax-M1-40k and MiniMax-M1-80k. Built on a Mixture-of-Experts (MoE) architecture, the model is trained on 456 billion parameters. MiniMax has introduced Lightning Attention for its M1 model, dramatically reducing inference costs and only consumes 25% of the floating point operations (FLOPs) required by DeepSeek R1 at a generation length of 100,000 tokens. Available on AI code sharing communities like Hugging Face and GitHub, MiniMax-M1 is released under the Apache 2.0 license, enabling businesses to freely use, modify, and implement it for commercial applications without restrictions or payment. MiniMax-M1 features a web search functionality and can handle multimodal input like text, images, and presentations. The expansive context window allows the model to exchange information equivalent to a small collection or book series, far exceeding OpenAI's GPT-4o, which has a context window of 128,000 tokens. Recommended read:
References :
Carl Franzen@AI News | VentureBeat
//
Mistral AI has launched its first reasoning model, Magistral, signaling a commitment to open-source AI development. The Magistral family features two models: Magistral Small, a 24-billion parameter model available with open weights under the Apache 2.0 license, and Magistral Medium, a proprietary model accessible through an API. This dual release strategy aims to cater to both enterprise clients seeking advanced reasoning capabilities and the broader AI community interested in open-source innovation.
Mistral's decision to release Magistral Small under the permissive Apache 2.0 license marks a significant return to its open-source roots. The license allows for the free use, modification, and distribution of the model's source code, even for commercial purposes. This empowers startups and established companies to build and deploy their own applications on top of Mistral’s latest reasoning architecture, without the burdens of licensing fees or vendor lock-in. The release serves as a powerful counter-narrative, reaffirming Mistral’s dedication to arming the open community with cutting-edge tools. Magistral Medium demonstrates competitive performance in the reasoning arena, according to internal benchmarks released by Mistral. The model was tested against its predecessor, Mistral-Medium 3, and models from Deepseek. Furthermore, Mistral's Agents API's Handoffs feature facilitates smart, multi-agent workflows, allowing different agents to collaborate on complex tasks. This enables modular and efficient problem-solving, as demonstrated in systems where agents collaborate to answer inflation-related questions. Recommended read:
References :
Carl Franzen@AI News | VentureBeat
//
Mistral AI has launched Magistral, its inaugural reasoning large language model (LLM), available in two distinct versions. Magistral Small, a 24 billion parameter model, is offered with open weights under the Apache 2.0 license, enabling developers to freely use, modify, and distribute the code for commercial or non-commercial purposes. This model can be run locally using tools like Ollama. The other version, Magistral Medium, is accessible exclusively via Mistral’s API and is tailored for enterprise clients, providing traceable reasoning capabilities crucial for compliance in highly regulated sectors such as legal, financial, healthcare, and government.
Mistral is positioning Magistral as a powerful tool for both professional and creative applications. The company highlights Magistral's ability to perform "transparent, multilingual reasoning," making it suitable for tasks involving complex calculations, programming logic, decision trees, and rule-based systems. Additionally, Mistral is promoting Magistral for creative writing, touting its capacity to generate coherent or, if desired, uniquely eccentric content. Users can experiment with Magistral Medium through the "Thinking" mode within Mistral's Le Chat platform, with options for "Pure Thinking" and a high-speed "10x speed" mode powered by Cerebras. Benchmark tests reveal that Magistral Medium is competitive in the reasoning arena. On the AIME-24 mathematics benchmark, the model achieved an impressive 73.6% accuracy, comparable to its predecessor, Mistral Medium 3, and outperforming Deepseek's models. Mistral's strategic release of Magistral Small under the Apache 2.0 license is seen as a reaffirmation of its commitment to open source principles. This move contrasts with the company's previous release of Medium 3 as a proprietary offering, which had raised concerns about a shift towards a more closed ecosystem. Recommended read:
References :
anket.sah@lambda.ai (Anket@lambdalabs.com
//
References:
lambda.ai
, thezvi.wordpress.com
,
DeepSeek's latest model, R1-0528, is now available on Lambda’s Inference API, marking an upgrade to the original R1 model released in January 2025. The new model, built upon the deepseek_v3 architecture, boasts a blend of mathematical capabilities, code generation finesse, and reasoning depth, aiming to challenge the dominance of OpenAI’s o3 and Google’s Gemini 2.5 Pro. DeepSeek-R1-0528 employs FP8 quantization, enhancing its ability to handle complex computations efficiently and features a mixture-of-experts (MoE) model with multi-headed latent attention (MLA) and multi-token prediction (MTP), enabling efficient handling of complex reasoning tasks.
DeepSeek-R1-0528, while a solid upgrade, didn't generate the same excitement as the initial R1 release. When R1 was released in January 2025, it was seen as a watershed moment for the company. This time around, it's considered a solid model for its price and status as an open model, and is best suited for tasks that align with its specific strengths. The initial DeepSeek release created a "DeepSeek moment", leading to market reactions and comparisons to other models. The first R1 model was released with a free app featuring a clear design and visible chain-of-thought, which forced other labs to follow suit. While DeepSeek R1-0528 offers advantages, experts warn of potential risks associated with open-source AI models. Cisco issued a report shortly after R1 began dominating headlines which claimed DeepSeek failed to block a single harmful prompt when tested against 50 random prompts taken from the HarmBench dataset. These risks include potential misuse for cyber threats, spread of misinformation, and reinforcement of biases. There are concerns regarding data poisoning, where compromised training data could lead to biased or disinformation. Furthermore, adversaries could modify the models to bypass controls, generate harmful content, or embed backdoors for exploitation. Recommended read:
References :
@medium.com
//
References:
, TheSequence
DeepSeek's latest AI model, R1-0528, is making waves in the AI community due to its impressive performance in math and reasoning tasks. This new model, despite having a similar name to its predecessor, boasts a completely different architecture and performance profile, marking a significant leap forward. DeepSeek R1-0528 has demonstrated "unprecedented levels of demand" shooting to the top of the App Store past closed model rivals and overloading their API with unprecedented levels of demand to the point that they actually had to stop accepting payments.
The most notable improvement in DeepSeek R1-0528 is its mathematical reasoning capabilities. On the AIME 2025 test, the model's accuracy increased from 70% to 87.5%, surpassing Gemini 2.5 Pro and putting it in close competition with OpenAI's o3. This improvement is attributed to "enhanced thinking depth," with the model using significantly more tokens per question, engaging in more thorough chains of reasoning. This means the model can check its own work, recognize errors, and course-correct during problem-solving. DeepSeek's success is challenging established closed models and driving competition in the AI landscape. DeepSeek-R1-0528 continues to utilize a Mixture-of-Experts (MoE) architecture, now scaled up to an enormous size. This sparse activation allows for powerful specialized expertise in different coding domains while maintaining efficiency. The context also continues to remain at 128k (with RoPE scaling or other improvements capable of extending it further.) The rise of DeepSeek is underscored by its performance benchmarks, which show it outperforming some of the industry’s leading models, including OpenAI’s ChatGPT. Furthermore, the release of a distilled variant, R1-0528-Qwen3-8B, ensures broad accessibility of this powerful technology. Recommended read:
References :
@www.marktechpost.com
//
DeepSeek, a Chinese AI startup, has launched an updated version of its R1 reasoning AI model, named DeepSeek-R1-0528. This new iteration brings the open-source model near parity with proprietary paid models like OpenAI’s o3 and Google’s Gemini 2.5 Pro in terms of reasoning capabilities. The model is released under the permissive MIT License, enabling commercial use and customization, marking a commitment to open-source AI development. The model's weights and documentation are available on Hugging Face, facilitating local deployment and API integration.
The DeepSeek-R1-0528 update introduces substantial enhancements in the model's ability to handle complex reasoning tasks across various domains, including mathematics, science, business, and programming. DeepSeek attributes these improvements to leveraging increased computational resources and applying algorithmic optimizations in post-training. Notably, the accuracy on the AIME 2025 test has surged from 70% to 87.5%, demonstrating deeper reasoning processes with an average of 23,000 tokens per question, compared to the previous version's 12,000 tokens. Alongside enhanced reasoning, the updated R1 model boasts a reduced hallucination rate, which contributes to more reliable and consistent output. Code generation performance has also seen a boost, positioning it as a strong contender in the open-source AI landscape. DeepSeek provides instructions on its GitHub repository for those interested in running the model locally and encourages community feedback and questions. The company aims to provide accessible AI solutions, underscored by the availability of a distilled version of R1-0528, DeepSeek-R1-0528-Qwen3-8B, designed for efficient single-GPU operation. Recommended read:
References :
@www.marktechpost.com
//
DeepSeek has released a major update to its R1 reasoning model, dubbed DeepSeek-R1-0528, marking a significant step forward in open-source AI. The update boasts enhanced performance in complex reasoning, mathematics, and coding, positioning it as a strong competitor to leading commercial models like OpenAI's o3 and Google's Gemini 2.5 Pro. The model's weights, training recipes, and comprehensive documentation are openly available under the MIT license, fostering transparency and community-driven innovation. This release allows researchers, developers, and businesses to access cutting-edge AI capabilities without the constraints of closed ecosystems or expensive subscriptions.
The DeepSeek-R1-0528 update brings several core improvements. The model's parameter count has increased from 671 billion to 685 billion, enabling it to process and store more intricate patterns. Enhanced chain-of-thought layers deepen the model's reasoning capabilities, making it more reliable in handling multi-step logic problems. Post-training optimizations have also been applied to reduce hallucinations and improve output stability. In practical terms, the update introduces JSON outputs, native function calling, and simplified system prompts, all designed to streamline real-world deployment and enhance the developer experience. Specifically, DeepSeek R1-0528 demonstrates a remarkable leap in mathematical reasoning. On the AIME 2025 test, its accuracy improved from 70% to an impressive 87.5%, rivaling OpenAI's o3. This improvement is attributed to "enhanced thinking depth," with the model now utilizing significantly more tokens per question, indicating more thorough and systematic logical analysis. The open-source nature of DeepSeek-R1-0528 empowers users to fine-tune and adapt the model to their specific needs, fostering further innovation and advancements within the AI community. Recommended read:
References :
Kevin Okemwa@windowscentral.com
//
References:
engineering.fb.com
, www.windowscentral.com
Microsoft is strategically prioritizing AI model accessibility through Azure, with CEO Satya Nadella emphasizing making AI solutions available to customers for maximum profit. This approach involves internal restructuring, including job cuts, to facilitate increased investment in AI and streamline operations. The goal is to build a robust, subscription-based AI operating system that leverages advancements like ChatGPT, ensuring that Microsoft remains competitive in the rapidly evolving AI landscape.
Microsoft is actively working on improving integrations with external data sources using the Model Context Protocol (MCP). This initiative has led to a collaboration with Twilio to enhance conversational AI capabilities for enterprise customer communication. Twilio's technology helps deliver the "last mile" of AI conversations, enabling businesses to integrate Microsoft's conversational intelligence capabilities into their existing communication channels. This partnership gives Twilio greater visibility among Microsoft's enterprise customers, exposing its developer tools to large firms looking to build extensible custom communication solutions. In addition to these strategic partnerships, Microsoft is also contributing to the open-source community by releasing Pyrefly, a faster Python type checker written in Rust. Developed initially at Meta for Instagram's codebase, Pyrefly is now available for the broader Python community to use, helping developers catch errors before runtime. The release of Pyrefly signifies Microsoft's commitment to fostering innovation and supporting the development of AI-related tools and technologies. Recommended read:
References :
@www.unite.ai
//
Anaconda Inc. has launched the Anaconda AI Platform, the first unified AI development platform tailored for open source. This platform is designed to streamline and secure the entire AI lifecycle, enabling enterprises to move from experimentation to production more efficiently. The Anaconda AI Platform aims to simplify the open-source Python stack by offering a standardized user experience that enhances data governance and streamlines AI workflows. The goal is to unify the experience across various Anaconda products, making it easier for administrators to have a comprehensive view of open source within their organizations.
The Anaconda AI Platform addresses the challenges enterprises face when deploying open-source tools like TensorFlow, PyTorch, and scikit-learn at scale. Issues such as security vulnerabilities, dependency conflicts, compliance risks, and governance limitations often hinder enterprise adoption. The platform provides essential guardrails that enable responsible innovation, delivering documented ROI and enterprise-grade governance capabilities. By combining trusted distribution, simplified workflows, real-time insights, and governance controls, the Anaconda AI Platform delivers secure and production-ready enterprise Python. Peter Wang, Chief AI and Innovation Officer and Co-founder of Anaconda, stated that until now, there hasn’t been a single destination for AI development with open source. He emphasized that the Anaconda AI Platform not only offers streamlined workflows, enhanced security, and substantial time savings but also provides choice, allowing enterprises to customize their AI journey with the first unified AI platform for open source, accelerating AI innovation and real-time insights. The platform empowers organizations to leverage open source as a strategic business asset, building reliable and innovative AI systems without sacrificing speed, value, or flexibility. Recommended read:
References :
@felloai.com
//
Alibaba has launched Qwen3, a new family of large language models (LLMs), posing a significant challenge to Silicon Valley's AI dominance. Qwen3 is not just an incremental update but a leap forward, demonstrating capabilities that rival leading models from OpenAI, Google, and Meta. This advancement signals China’s growing prowess in AI and its potential to redefine the global tech landscape. Qwen3's strengths lie in reasoning, coding, and multilingual understanding, marking a pivotal moment in China's AI development.
The Qwen3 family includes models of varying sizes to cater to diverse applications. Key features include complex reasoning, mathematical problem-solving, and code generation. The models support 119 languages and are trained on a massive dataset of over 36 trillion tokens. Another innovation is Qwen3’s “hybrid reasoning” approach, enabling models to switch between "fast thinking" for quick responses and "slow thinking" for deeper analysis, enhancing versatility and efficiency. Alibaba has also emphasized the open-source nature of some Qwen3 models, fostering wider adoption and collaborative development in China's AI ecosystem. Alibaba also introduced ZeroSearch, a method that uses reinforcement learning and simulated documents to teach LLMs retrieval without real-time search. It addresses the challenge of LLMs relying on static datasets, which can become outdated. By training the models to retrieve and incorporate external information, ZeroSearch aims to improve the reliability of LLMs in real-world applications like news, research, and product reviews. This method mitigates the high costs associated with large-scale interactions with live APIs, making it more accessible for academic research and commercial deployment. Recommended read:
References :
@venturebeat.com
//
Nvidia has launched Parakeet-TDT-0.6B-V2, a fully open-source transcription AI model, on Hugging Face. This represents a new standard for Automatic Speech Recognition (ASR). The model, boasting 600 million parameters, has quickly topped the Hugging Face Open ASR Leaderboard with a word error rate of just 6.05%. This level of accuracy positions it near proprietary transcription models, such as OpenAI’s GPT-4o-transcribe and ElevenLabs Scribe, making it a significant advancement in open-source speech AI. Parakeet operates under a commercially permissive CC-BY-4.0 license.
The speed of Parakeet-TDT-0.6B-V2 is a standout feature. According to Hugging Face’s Vaibhav Srivastav, it can "transcribe 60 minutes of audio in 1 second." Nvidia reports this is achieved with a real-time factor of 3386, meaning it processes audio 3386 times faster than real-time when running on Nvidia's GPU-accelerated hardware. This speed is attributed to its transformer-based architecture, fine-tuned with high-quality transcription data and optimized for inference on NVIDIA hardware using TensorRT and FP8 quantization. The model also supports punctuation, capitalization, and detailed word-level timestamping. Parakeet-TDT-0.6B-V2 is aimed at developers, researchers, and industry teams building various applications. This includes transcription services, voice assistants, subtitle generators, and conversational AI platforms. Its accessibility and performance make it an attractive option for commercial enterprises and indie developers looking to build speech recognition and transcription services into their applications. With its release on May 1, 2025, Parakeet is set to make a considerable impact on the field of speech AI. Recommended read:
References :
@syncedreview.com
//
DeepSeek AI has unveiled DeepSeek-Prover-V2, a new open-source large language model (LLM) designed for formal theorem proving within the Lean 4 environment. This model advances the field of neural theorem proving by utilizing a recursive theorem-proving pipeline and leverages DeepSeek-V3 to generate high-quality initialization data. DeepSeek-Prover-V2 has achieved top results on the MiniF2F benchmark, showcasing its state-of-the-art performance in mathematical reasoning. The release includes ProverBench, a new benchmark for evaluating mathematical reasoning capabilities.
DeepSeek-Prover-V2 features a unique cold-start training procedure. The process begins by using the DeepSeek-V3 model to decompose complex mathematical theorems into a series of more manageable subgoals. Simultaneously, DeepSeek-V3 formalizes these high-level proof steps in Lean 4, creating a structured sequence of sub-problems. To handle the computationally intensive proof search for each subgoal, the researchers employed a smaller 7B parameter model. Once all the decomposed steps of a challenging problem are successfully proven, the complete step-by-step formal proof is paired with DeepSeek-V3’s corresponding chain-of-thought reasoning. This allows the model to learn from a synthesized dataset that integrates both informal, high-level mathematical reasoning and rigorous formal proofs, providing a strong cold start for subsequent reinforcement learning. Building upon the synthetic cold-start data, the DeepSeek team curated a selection of challenging problems that the 7B prover model couldn’t solve end-to-end, but for which all subgoals had been successfully addressed. By combining the formal proofs of these subgoals, a complete proof for the original problem is constructed. This formal proof is then linked with DeepSeek-V3’s chain-of-thought outlining the lemma decomposition, creating a unified training example of informal reasoning followed by formalization. DeepSeek is also challenging the long-held belief of tech CEOs who've argued that exponential AI improvements require ever-increasing computing power. DeepSeek claims to have produced models comparable to OpenAI, but with significantly less compute and cost, questioning the necessity of massive scale for AI advancement. Recommended read:
References :
Alexey Shabanov@TestingCatalog
//
Alibaba's Qwen team has launched Qwen3, a new family of open-source large language models (LLMs) designed to compete with leading AI systems. The Qwen3 series includes eight models ranging from 0.6B to 235B parameters, with the larger models employing a Mixture-of-Experts (MoE) architecture for enhanced performance. This comprehensive suite offers options for developers with varied computational resources and application requirements. All the models are released under the Apache 2.0 license, making them suitable for commercial use.
The Qwen3 models boast improved agentic capabilities for tool use and support for 119 languages. The models also feature a unique "hybrid thinking mode" that allows users to dynamically adjust the balance between deep reasoning and faster responses. This is particularly valuable for developers as it facilitates efficient use of computational resources based on task complexity. Training involved a large dataset of 36 trillion tokens and was optimized for reasoning, similar to the Deepseek R1 model. Benchmarks indicate that Qwen3 rivals top competitors like Deepseek R1 and Gemini Pro in areas like coding, mathematics, and general knowledge. Notably, the smaller Qwen3–30B-A3B MoE model achieves performance comparable to the Qwen3–32B dense model while activating significantly fewer parameters. These models are available on platforms like Hugging Face, ModelScope, and Kaggle, along with support for deployment through frameworks like SGLang and vLLM, and local execution via tools like Ollama and llama.cpp. Recommended read:
References :
Alexey Shabanov@TestingCatalog
//
Alibaba Cloud has unveiled Qwen 3, a new generation of large language models (LLMs) boasting 235 billion parameters, poised to challenge the dominance of US-based models. This open-weight family of models includes both dense and Mixture-of-Experts (MoE) architectures, offering developers a range of choices to suit their specific application needs and hardware constraints. The flagship model, Qwen3-235B-A22B, achieves competitive results in benchmark evaluations of coding, math, and general knowledge, positioning it as one of the most powerful publicly available models.
Qwen 3 introduces a unique "thinking mode" that can be toggled for step-by-step reasoning or rapid direct answers. This hybrid reasoning approach, similar to OpenAI's "o" series, allows users to engage a more intensive process for complex queries in fields like science, math, and engineering. The models are trained on a massive dataset of 36 trillion tokens spanning 119 languages, twice the corpus of Qwen 2.5 and enriched with synthetic math and code data. This extensive training equips Qwen 3 with enhanced reasoning, multilingual proficiency, and computational efficiency. The release of Qwen 3 includes two MoE models and six dense variants, all licensed under Apache-2.0 and downloadable from platforms like Hugging Face, ModelScope, and Kaggle. Deployment guidance points to vLLM and SGLang for servers and to Ollama or llama.cpp for local setups, signaling support for both cloud and edge developers. Community feedback has been positive, with analysts noting that earlier Qwen announcements briefly lifted Alibaba shares, underscoring the strategic weight the company places on open models. Recommended read:
References :
|
BenchmarksBlogsResearch Tools |