ByteDance Releases Protenix-v1: A New Open-Source Model Achieving AF3-Level Performance in Biomolecular Structure Prediction

ByteDance releases Protenix-v1, an AF3-class all-atom biomolecular structure prediction model with open code and weights under Apache 2.0, targeting proteins, DNA, RNA and ligands while explicitly matching AlphaFold3’s training data cutoff, model scale class and inference budget for fair comparison. Benchmarks are run with PXMeter v1.0.0 on more than 6k curated complexes with time-split and domain-specific subsets, showing Protenix-v1 outperforming AF3 and exhibiting clean, log-linear inference-time scaling as the number of sampled candidates increases. The ecosystem includes Protenix-v1-20250630 for applied use, compact Protenix-Mini variants for efficient inference, PXDesign for high-hit-rate binder design and Protenix-Dock for docking, giving researchers and devs an AF3-style reference implementation plus a reproducible evaluation stack they can integrate, profile and extend in real-world pipelines.....

Full analysis: https://www.marktechpost.com/2026/02/08/bytedance-releases-protenix-v1-a-new-open-source-model-achieving-af3-level-performance-in-biomolecular-structure-prediction/

Repo: https://github.com/bytedance/Protenix

Server to try it: https://protenix-server.com/login

Alibaba Open-Sources Zvec: An Embedded Vector Database Bringing SQLite-like Simplicity and High-Performance On-Device RAG to Edge Applications

Zvec is an open source, embedded, in-process vector database that targets edge and on-device RAG workloads by acting like the SQLite of vector databases. Built on Alibaba’s production grade Proxima engine and released under Apache 2.0, it runs as a simple Python library and delivers more than 8,000 QPS on VectorDBBench with the Cohere 10M dataset, over 2× the previous leaderboard #1 ZillizCloud, while also reducing index build time. Zvec exposes explicit memory and CPU controls through streaming writes, mmap mode, optional memory limits, and thread configuration, which makes it practical for mobile, desktop, and other constrained environments. It is RAG ready with full CRUD, schema evolution, multi vector retrieval, built in weighted fusion and RRF reranking, and scalar vector hybrid search......

Full analysis: https://www.marktechpost.com/2026/02/10/alibaba-open-sources-zvec-an-embedded-vector-database-bringing-sqlite-like-simplicity-and-high-performance-on-device-rag-to-edge-applications/

Repo: https://github.com/alibaba/zvec

Technical details: https://zvec.org/en/blog/introduction/

OpenAI Releases a Research Preview of GPT‑5.3-Codex-Spark: A 15x Faster AI Coding Model Delivering Over 1000 Tokens Per Second on Cerebras Hardware

OpenAI has launched GPT-5.3 Codex-Spark, a research preview optimized for near-instant coding by delivering over 1000 tokens per second—a 15x speed increase over the flagship model. This massive performance jump is powered by the Cerebras Wafer-Scale Engine 3 (WSE-3), which eliminates traditional GPU bottlenecks by keeping all compute on a single silicon wafer, paired with a new persistent WebSocket connection that reduces networking overhead by 80%.....

Full analysis: https://www.marktechpost.com/2026/02/12/openai-releases-a-research-preview-of-gpt-5-3-codex-spark-a-15x-faster-ai-coding-model-delivering-over-1000-tokens-per-second-on-cerebras-hardware/

Technical details: https://openai.com/index/introducing-gpt-5-3-codex-spark/

Exa AI Introduces Exa Instant: A Sub-200ms Neural Search Engine Designed to Eliminate Bottlenecks for Real-Time Agentic Workflows

Exa has launched Exa Instant, a proprietary neural search engine designed to solve the latency bottleneck in AI agent workflows. By bypassing traditional search engine wrappers and using a custom transformer-based stack, Exa Instant delivers web results in under 200ms with network speeds as low as 50ms. This 15x speed improvement allows engineers to treat search as a real-time primitive in RAG pipelines rather than a slow, external dependency. Priced at $5 per 1,000 requests, the model prioritizes semantic intent over keywords, effectively turning the live web into a high-speed context extension for LLMs.....

Full analysis: https://www.marktechpost.com/2026/02/13/exa-ai-introduces-exa-instant-a-sub-200ms-neural-search-engine-designed-to-eliminate-bottlenecks-for-real-time-agentic-workflows/

Technical details: https://exa.ai/blog/exa-instant

product on ainews platform: https://ainews.sh/functions/socialShare?id=698f91e3c30ec9e1a6b27895&type=product

Alibaba Qwen Team Releases Qwen3.5-397B MoE Model with 17B Active Parameters and 1M Token Context for AI agents

Alibaba's Qwen3.5 release marks a major breakthrough in open-source AI, introducing the 397B-A17B flagship model that utilizes a sparse Mixture-of-Experts (MoE) architecture and a unique Gated Delta Network hybrid design. This technical synergy allows the model to offer 400B-class reasoning with the inference speed of a 17B model, achieving a massive 8.6x to 19.0x increase in decoding throughput. As a native vision-language model trained through Early Fusion, it excels at agentic tasks and visual reasoning across 201 languages, supported by a staggering 1M token context window in the Qwen3.5-Plus version. Released under the Apache 2.0 license, it provides devs and data scientists a high-performance, cost-efficient foundation for building the next generation of multimodal autonomous agents....

Full analysis: https://lnkd.in/g4AaHdEt

Model weights: https://lnkd.in/gNtCyKR6

Repo: https://lnkd.in/gqj3wVX3

Cohere Releases Tiny Aya: A 3B-Parameter Small Language Model that Supports 70 Languages and Runs Locally Even on a Phone

Tiny Aya is a new family of small multilingual language models (SLMs) from Cohere Labs that delivers state-of-the-art performance across 70 languages with only 3.35B parameters. By prioritizing balanced linguistic coverage over brute-force scaling, the model family—which includes a global model and three region-specific variants—outperforms larger competitors like Gemma3-4B in translation quality for 46 of 61 languages and mathematical reasoning in underrepresented regions like Africa. The models utilize a dense decoder-only architecture and were refined through a sophisticated synthetic data pipeline called Fusion-of-N, which distills high-quality signals from frontier models while preserving regional nuances. Designed for accessibility and practical deployment, Tiny Aya is optimized for edge devices, achieving 10 to 32 tokens per second on iPhones while maintaining high generation quality through efficient 4-bit quantization.....

Full analysis: https://www.marktechpost.com/2026/02/17/cohere-releases-tiny-aya-a-3b-parameter-small-language-model-that-supports-70-languages-and-runs-locally-even-on-a-phone/

Paper: https://github.com/Cohere-Labs/tiny-aya-tech-report/blob/main/tiny_aya_tech_report.pdf

Model weights: https://huggingface.co/collections/CohereLabs/tiny-aya?

Try it here: https://huggingface.co/spaces/CohereLabs/tiny-aya?ref=cohere.com%2Fblog

Google DeepMind Releases Lyria 3: An Advanced Music Generation AI Model that Turns Photos and Text into Custom Tracks with Included Lyrics and Vocals

Lyria 3 is Google's new multimodal generative AI model integrated into the Gemini app that converts text prompts and photos into high-fidelity, 30-second music tracks. Designed for both creators and engineers, the model achieves superior long-range coherence and 48kHz audio quality, generating full arrangements complete with vocals and lyrics. For technical safety, Google implements SynthID, an inaudible digital watermarking technology that ensures AI-generated content remains detectable even after heavy editing. This release, paired with the Music AI Sandbox, transitions generative audio from simple MIDI loops to professional-grade, "human-in-the-loop" synthesis, setting a new standard for the 2026 AI music landscape......

Full analysis: https://www.marktechpost.com/2026/02/18/google-deepmind-releases-lyria-3-an-advanced-music-generation-ai-model-that-turns-photos-and-text-into-custom-tracks-with-included-lyrics-and-vocals/

Technical details: https://deepmind.google/models/lyria/

Google AI Releases Gemini 3.1 Pro with 1 Million Token Context and 77.1 Percent ARC-AGI-2 Reasoning for AI Agents

Key upgrades include a massive 1,048,576 token input context paired with a new 65,536 token output window, and a breakthrough 77.1% score on the ARC-AGI-2 benchmark—more than double the reasoning power of its predecessor. Developers gain a specialized customtools endpoint for prioritized terminal and bash execution. With expanded 100MB file limits and direct YouTube URL support, Gemini 3.1 Pro positions itself as the high-efficiency, reasoning-first engine for the next generation of software engineering and scientific research agents......

Full analysis: https://www.marktechpost.com/2026/02/19/google-ai-releases-gemini-3-1-pro-with-1-million-token-context-and-77-1-percent-arc-agi-2-reasoning-for-ai-agents/

Technical details: https://www.marktechpost.com/2026/02/19/google-ai-releases-gemini-3-1-pro-with-1-million-token-context-and-77-1-percent-arc-agi-2-reasoning-for-ai-agents/

Try it here: https://aistudio.google.com/prompts/new_chat?model=gemini-3.1-pro-preview

NVIDIA Releases DreamDojo: An Open-Source Robot World Model Trained on 44,711 Hours of Real-World Human Video Data

NVIDIA has introduced DreamDojo, an open-source, generalizable foundation world model designed to simulate complex robotics tasks by 'dreaming' future outcomes directly in pixels. By pretraining on 44,711 hours of egocentric human videos—the largest dataset of its kind—the model acquires a deep understanding of real-world physics and interaction dynamics. To overcome the lack of motor labels in human data, the NVIDIA team implemented continuous latent actions as a hardware-agnostic proxy, allowing the model to transfer knowledge across different robot embodiments. Optimized through a Self Forcing distillation pipeline, DreamDojo achieves real-time speeds of 10.81 FPS, unlocking advanced applications such as live teleoperation, model-based planning, and highly accurate policy evaluation with a 0.995 Pearson correlation to real-world performance....

Read the full analysis: https://www.marktechpost.com/2026/02/20/nvidia-releases-dreamdojo-an-open-source-robot-world-model-trained-on-44711-hours-of-real-world-human-video-data/

Paper: https://arxiv.org/pdf/2602.06949

Repo: https://github.com/NVIDIA/DreamDojo

Is There a Community Edition of Palantir? Meet OpenPlanter: An Open Source Recursive AI Agent for Your Micro Surveillance Use Cases

OpenPlanter is a recursive-language-model investigation agent designed to automate civic oversight and forensic data analysis. The system ingests disparate structured and unstructured datasets to perform entity resolution and detect probabilistic anomalies across public records. It utilizes a recursive sub-agent delegation strategy with a max-depth of 4 to parallelize complex evidence-chain construction. The technical stack includes gpt-5.2 and claude-opus-4-6, supported by 19 tools for shell execution, file I/O, and web search. By acting as an open-source alternative to proprietary surveillance platforms.....

Full analysis: https://www.marktechpost.com/2026/02/21/is-there-a-community-edition-of-palantir-meet-openplanter-an-open-source-recursive-ai-agent-for-your-micro-surveillance-use-cases/

Repo: https://github.com/ShinMegamiBoson/OpenPlanter?tab=readme-ov-file

Forget Keyword Imitation: ByteDance AI Maps Molecular Bonds in AI Reasoning to Stabilize Long Chain-of-Thought Performance and Reinforcement Learning (RL) Training

ByteDance researchers have introduced a 'molecular' framework to explain Long Chain-of-Thought (Long CoT) reasoning, positing that effective trajectories are held together by 3 distinct behavioral bonds: Deep Reasoning (covalent-like) forming the logical backbone, Self-Reflection (hydrogen-bond-like) providing stability through 'logical folding,' and Self-Exploration (van der Waals-like) bridging distant concepts. The research team proves that models internalize these structural behaviors rather than just surface-level keywords, and that mixing incompatible Semantic Isomers—trajectories with similar concepts but different behavior distributions—can lead to structural chaos and performance loss.

To address this, they developed MOLE-SYN, a distribution-transfer-graph method that synthesizes these stable reasoning structures from scratch using instruction-tuned LLMs, achieving performance near-distillation levels and enhancing Reinforcement Learning (RL) stability across 6 benchmarks. Ultimately, this framework suggests that Long CoT mimics protein folding, where the arrangement of these logical bonds determines the model's ability to converge toward stable, optimized solutions in semantic space.....

Full analysis: https://www.marktechpost.com/2026/02/22/forget-keyword-imitation-bytedance-ai-maps-molecular-bonds-in-ai-reasoning-to-stabilize-long-chain-of-thought-performance-and-reinforcement-learning-rl-training/

Paper: https://arxiv.org/pdf/2601.06002

Alibaba Qwen Team Releases Qwen 3.5 Medium Model Series: A Production Powerhouse Proving that Smaller AI Models are Smarter

Alibaba’s Qwen 3.5 Medium Model Series signals a decisive pivot from "brute-force" scaling to architectural efficiency, proving that superior data quality and Reinforcement Learning (RL) can outperform traditional parameter density. The series starts by Qwen3.5-35B-A3B, a Mixture-of-Experts (MoE) model that utilizes just 3 billion active parameters to surpass the older 235B giant, effectively slashing inference costs while maintaining frontier-level reasoning.

With Qwen3.5-Flash offering a default 1M context window and native tool support, this release provides a high-throughput, agent-ready infrastructure that narrows the gap between open-weight versatility and the industry's most massive proprietary models.....

Full analysis: https://www.marktechpost.com/2026/02/24/alibaba-qwen-team-releases-qwen-3-5-medium-model-series-a-production-powerhouse-proving-that-smaller-ai-models-are-smarter/

Model Weights: https://huggingface.co/collections/Qwen/qwen35

API: https://modelstudio.console.alibabacloud.com/ap-southeast-1/?tab=doc#/doc/?type=model&url=2840914_2&modelId=group-qwen3.5-flash

Tailscale and LM Studio Introduce ‘LM Link’ to Provide Encrypted Point-to-Point Access to Your Private GPU Hardware Assets

LM Link is a game-changing product from LM Studio and Tailscale that effectively turns your high-end GPU workstation into a private, encrypted AI cloud. By integrating tsnet directly into the architecture, the tool creates a secure, identity-based tunnel that allows you to run massive models on remote hardware as if they were plugged into your local machine—no public endpoints, no firewall tinkering, and zero "API key sprawl." The workflow is dead simple: you query localhost:1234 on your laptop, and LM Link handles the heavy lifting via a point-to-point WireGuard® connection to your "Big Rig" at home. It’s the ultimate fix for the "Big Model, Small Laptop" dilemma, providing data-center performance with local-level privacy.....

Full analysis: https://www.marktechpost.com/2026/02/25/tailscale-and-lm-studio-introduce-lm-link-to-provide-encrypted-point-to-point-access-to-your-private-gpu-hardware-assets/

Technical details: https://lmstudio.ai/link

Nous Research Releases ‘Hermes Agent’ to Fix AI Forgetfulness with Multi-Level Memory and Dedicated Remote Terminal Access Support

Tired of AI with "goldfish memory"? Nous Research just launched Hermes Agent, an open-source system designed to evolve from a simple agent into a persistent digital colleague agent that actually gets smarter the more you use it. By leveraging a multi-level memory system—including searchable Skill Documents—and providing persistent machine access via Docker, SSH, and local backends, Hermes Agent doesn't just write code; it lives in your environment and retains its state across sessions. Powered by the highly steerable Hermes-3 (Llama 3.1) and the Atropos RL framework, it bridges the gap between reasoning and execution, offering engineers a sovereign, self-improving assistant that stays entirely within their own infrastructure while communicating via familiar tools like Telegram and Slack.....

Full analysis: https://www.marktechpost.com/2026/02/26/nous-research-releases-hermes-agent-to-fix-ai-forgetfulness-with-multi-level-memory-and-dedicated-remote-terminal-access-support/

Technical details: https://nousresearch.com/hermes-agent/

GitHub Repo: https://github.com/NousResearch/hermes-agent

Perplexity Just Released pplx-embed: New SOTA Qwen3 Bidirectional Embedding Models for Web-Scale Retrieval Tasks

pplx-embed is a suite of state-of-the-art multilingual embedding models (0.6B and 4B) built on the Qwen3 architecture and released under a permissive MIT License. Unlike standard causal models, pplx-embed utilizes bidirectional attention and diffusion-based pretraining to extract clean semantic signals from noisy, web-scale data. Optimized for Retrieval-Augmented Generation (RAG), the collection includes specialized versions—pplx-embed-v1 for queries and pplx-embed-context-v1 for document chunks—while supporting native INT8 quantization and Matryoshka Representation Learning for high-efficiency production deployment across Hugging Face, Sentence Transformers, and Transformers.js.....

Full analysis: https://www.marktechpost.com/2026/02/26/perplexity-just-released-pplx-embed-new-sota-qwen3-bidirectional-embedding-models-for-web-scale-retrieval-tasks/

Paper: https://arxiv.org/pdf/2602.11151

Model weights: https://huggingface.co/collections/perplexity-ai/pplx-embed

Technical details: https://research.perplexity.ai/articles/pplx-embed-state-of-the-art-embedding-models-for-web-scale-retrieval

Sakana AI Introduces Doc-to-LoRA and Text-to-LoRA: Hypernetworks that Instantly Internalize Long Contexts and Adapt LLMs via Zero-Shot Natural Language

Doc-to-LoRA (D2L) and Text-to-LoRA (T2L) are two innovative methods that utilize lightweight hypernetworks to instantly customize Large Language Models (LLMs) through a single forward pass. T2L enables zero-shot task adaptation based solely on natural language descriptions, matching the performance of specifically tuned adapters while significantly reducing adaptation costs compared to traditional in-context learning. D2L addresses the "long context" bottleneck by internalizing documents directly into model parameters through a Perceiver-based architecture and a chunking mechanism. This allows models to answer queries without re-consuming original context, maintaining near-perfect accuracy on information retrieval tasks at lengths exceeding the model's native window by more than four times while reducing KV-cache memory usage from gigabytes to less than 50 megabytes. Both systems operate with sub-second latency, effectively amortizing training costs and opening possibilities for rapid, on-device personalization. Remarkably, D2L also demonstrates cross-modal capability, transferring visual information from Vision-Language Models into text-only LLMs zero-shot to enable image classification purely through internalized weights.....

Full analysis: https://www.marktechpost.com/2026/02/27/sakana-ai-introduces-doc-to-lora-and-text-to-lora-hypernetworks-that-instantly-internalize-long-contexts-and-adapt-llms-via-zero-shot-natural-language/

Updates: https://pub.sakana.ai/doc-to-lora/

Doc-to-LoRA
Paper: https://arxiv.org/pdf/2602.15902
Code: https://github.com/SakanaAI/Doc-to-LoRA

Text-to-LoRA
Paper: https://arxiv.org/pdf/2506.06105
Code: https://github.com/SakanaAI/Text-to-LoRA

Alibaba Releases OpenSandbox to Provide Software Developers with a Unified, Secure, and Scalable API for Autonomous AI Agent Execution

Alibaba has open-sourced OpenSandbox, an Apache 2.0-licensed execution environment designed to provide AI agents with secure, isolated spaces for code execution, web browsing, and model training. Built on a modular four-layer architecture—comprising SDKs, Specs, Runtime, and Sandbox Instances—the tool utilizes a FastAPI-based control plane and a Go-based execd daemon to manage workloads across Docker or Kubernetes runtimes. By integrating with Jupyter kernels for stateful code execution and supporting tools like Playwright and VNC desktops, OpenSandbox offers a unified, vendor-free API that eliminates the per-minute billing and fragmentation common in proprietary sandbox services......

Full analysis: https://www.marktechpost.com/2026/03/03/alibaba-releases-opensandbox-to-provide-software-developers-with-a-unified-secure-and-scalable-api-for-autonomous-ai-agent-execution/

Repo: https://github.com/alibaba/OpenSandbox?tab=readme-ov-file

Docs: https://open-sandbox.ai/

Examples: https://open-sandbox.ai/examples/readme

Google Drops Gemini 3.1 Flash-Lite: A Cost-efficient Powerhouse with Adjustable Thinking Levels Designed for High-Scale Production AI

Google’s new Gemini 3.1 Flash-Lite is a tactical play for the "intelligence at scale" era, offering a faster, cheaper alternative to the Gemini 2.5 Flash baseline. By introducing "thinking levels," Google is giving a literal dial to balance reasoning depth against latency, allowing for $0.25/1M input token efficiency without sacrificing the logic needed for complex UI generation or simulations. It’s essentially a high-throughput workhorse that proves you don’t need a frontier-sized budget to ship production-grade reasoning—all while clocking in at 2.5x faster startup times......

Full analysis: https://www.marktechpost.com/2026/03/03/google-drops-gemini-3-1-flash-lite-a-cost-efficient-powerhouse-with-adjustable-thinking-levels-designed-for-high-scale-production-ai/

Technical details: https://blog.google/innovation-and-ai/models-and-research/gemini-models/gemini-3-1-flash-lite/?

Public Preview via the Gemini API (Google AI Studio): https://aistudio.google.com/prompts/new_chat?model=gemini-3.1-flash-lite-preview

Physical Intelligence Team Unveils MEM for Robots: A Multi-Scale Memory System Giving Gemma 3-4B VLAs 15-Minute Context for Complex Tasks

Multi-Scale Embodied Memory (MEM) is a dual-track architecture that allows Vision-Language-Action (VLA) models—specifically π0.6 initialized from Gemma 3-4B—to solve complex, long-horizon robotic tasks spanning up to 15 minutes. The system factorizes memory into two modalities: a short-term video encoder that uses space-time separable attention to process dense visual history (up to ~1 minute) without exceeding the critical ~380ms real-time inference barrier, and a long-term language-based memory where a high-level policy maintains a compressed semantic summary of past events. By reducing computational complexity to O(Kn^2+nK^2), MEM enables robots to handle partial observability and perform in-context adaptation—such as automatically switching door-opening directions after a failure (a +62% success rate improvement)—while matching the dexterous performance of state-of-the-art memoryless policies.....

Full analysis: https://www.marktechpost.com/2026/03/03/physical-intelligence-team-unveils-mem-for-robots-a-multi-scale-memory-system-giving-gemma-3-4b-vlas-15-minute-context-for-complex-tasks/

Paper: https://www.pi.website/download/Mem.pdf

Technical details: https://www.pi.website/research/memory

LangWatch Open Sourced the Missing Evaluation Layer for AI Agents to Enable End-to-End Tracing, Simulation, and Systematic Testing

LangWatch has open-sourced an evaluation and tracing platform designed to bring engineering rigor to non-deterministic AI agents. By integrating OpenTelemetry-native tracing with end-to-end simulations—featuring automated user simulators and LLM-based judges—it allows developers to pinpoint logic failures before production. The platform collapses tool sprawl through an 'Optimization Studio' that creates a closed loop between traces, datasets, and prompt iteration. Framework-agnostic and model-independent, LangWatch supports major stacks like LangGraph, CrewAI, and Vercel AI SDK while offering a self-hosted, ISO 27001-certified environment for enterprise-grade security and GitOps-aligned prompt versioning......

Full analysis:https://www.marktechpost.com/2026/03/04/langwatch-open-sources-the-missing-evaluation-layer-for-ai-agents-to-enable-end-to-end-tracing-simulation-and-systematic-testing/

GitHub Repo: https://github.com/langwatch/langwatch