marktechpost.com·platform·589 items·last fetched
StepFun has launched Step 5 Preview, a 600B-total, 27B-active parameter Mixture-of-Experts (MoE) model designed for long-horizon agentic workloads with a 1M-token context window.
Alibaba’s Qwen team released Qwen-Image-2.1, a unified 7B text-to-image and image-editing model with multi-reference editing and transparent RGBA output. It supports Diffusers, ComfyUI, vLLM-Omni, SGLang, and LightX2V…
Flet 1.0 has been released for production app development, featuring new AI-powered coding tools and an MCP server for AI assistants.
MarkTechPost reports that Google confirmed Gemini breached three companies during AI security tests, highlighting the growing capability—and risk—of agentic AI in offensive cybersecurity scenarios.
OpenClaw Releases 2026.9.5 With Atomic Updates, Plugin Hot Reload, Conversation Sharing, and Expanded GPT Live MarkTechPost
GGUF vs GPTQ vs AWQ vs EXL2: LLM Model Formats Explained (2026) MarkTechPost
MarkTechPost reports that Linkup Research released SPARSEUP, an open 149-million-parameter AI model or system; the search result does not expose enough detail to summarize its capabilities reliably.
OpenClaw, an open-source personal AI agent, released version 2026.9.5 featuring atomic updates, plugin hot reloading, conversation sharing, and expanded GPT Live capabilities.
OpenClaw’s 2026.9.5 release adds atomic updates with rollback, plugin installation or reload without Gateway restarts, read-only conversation sharing, and expanded GPT Live support in meetings and calls.
Meta has launched Muse for Mac, a new personal AI agent designed for macOS.
TypeSafe AI has released Jev, a transformer-based system one model designed to return typed, calibrated decisions instead of generating text.
SpaceXAI released Grok Voice Transcribe 2.0, a speech-to-text API claiming double the accuracy of its predecessor at the same price point.
MarkTechPost reports that Alibaba’s Qwen team released Qwen3.8-LiveTranslate, a real-time translation release aimed at live multilingual communication.
A MarkTechPost guide surveying open-source agent harnesses for running or coordinating local language models in 2026.
MarkTechPost reports a Qwen3.8-Omni-Flash release from Alibaba, adding another recent multimodal-model announcement to the site’s September 18 coverage.
PrismML released Ternary Bonsai 2 27B, an Apache 2.0 ternary-weight model reported at 5.93 GB versus 53.80 GB in FP16, while retaining 98.2% of the parent model’s average across 20 benchmarks. The article says it can…
Jina AI released jina-ocr-v1, a 3.4-billion-parameter mixture-of-experts document parser that includes speculative decoding and targets deployment on lower-budget GPUs.
The article examines Salesforce Agentforce as an enterprise agent platform focused on production orchestration, evaluation, regression testing, guardrails, and enterprise data integration. It highlights Southwest…
MarkTechPost examines Salesforce Agentforce as an enterprise agent platform focused on evaluation, regression testing, guardrails, data integration, and production orchestration. The article highlights Southwest…
Microsoft has open-sourced TauGrid, a Kubernetes-native stack designed for managing GPU AI workloads.
Anthropic launched a beta of Claude Code Projects, featuring parallel cloud sessions that continue running independently even after closing the user's laptop.
OpenAI has introduced a new framework for disclosing model misalignment, including three review tracks and six initial incident reports based on reinforcement learning training.
Knowledgator released GLiFormer, a schema-conditioned encoder for tasks like NER and JSON extraction, achieving high performance without token generation.
Stanford researchers introduced Paper2Agent, which converts research papers into Model Context Protocol (MCP) servers, allowing AI agents to run methods directly from PDFs.
Nums AI released Causilo, a pretrained tabular foundation model that leads the TabArena benchmark for single models in both classification and regression tasks.
Google introduced Gemini 3.8 Live and Gemini 3.8 Live Extended Thinking, which are native speech-to-speech models designed for real-time voice agents.
Reward AI introduced OM-1, a general-purpose robotics manipulation policy trained exclusively on human demonstrations without teleoperation or on-robot data.
NVIDIA has open-sourced OSMO, a Kubernetes-native workflow orchestrator designed to unify AI training, simulation, and robot testing using a single YAML configuration.
Sakana AI researchers introduced Augmented Lagrangian Predictive Coding (PC-ALM), a layer-local alternative to backpropagation capable of training 1000-layer networks.
Anthropic CEO Dario Amodei released an essay titled 'We Must Pace the Frontier' advocating for slowing down AI development, gaining endorsements from leaders at OpenAI, xAI, and Microsoft.
A Princeton researcher proposed the Recurrent Looped Transformer (RLT), an architecture designed to carry decoder states across tokens to provide unbounded temporal depth.
AWS introduced Pizza Bot, an open-source inbox designed for background AI agents.
Fly Language Model (FLM) Wires the Full Fruit Fly Connectome Into a Frozen 1.2B LLM, and Its Own Controls Show the Wiring Does Not Help MarkTechPost
Cognition released SWE-2, a coding model based on Kimi K3, which aims to match performance metrics of high-end models at a significantly lower cost.
Cognition released SWE-2, a Kimi K3 post-trained coding model designed to match Fable 5.1 performance at a significantly lower cost.
Researchers introduced the Fly Language Model (FLM), which integrates the fruit fly connectome into a 1.2B LLM, though control experiments suggest the specific wiring does not improve performance.
Anthropic Adds Plugin Evals to Claude Code: 6 Grader Types, a No-Plugin Baseline, and a CI Gate for Skills MarkTechPost
Google Research Releases ToolGrad: Answer-First Framework Hits 99.8% Pass Rate for Tool-Use Data Generation MarkTechPost
Anthropic has added plugin evaluations to Claude Code, including new grading types and a CI gate for skills.
Google Research introduced ToolGrad, an answer-first framework for tool-use data generation that significantly improves pass rates for LLMs.
Researchers from ByteDance Seed and other institutions evaluated the ability of LLMs to build their own agent harnesses, finding that generated improvements were often noisy and frequently failed to generalize to…
Sakana AI released Fugu Max and Fugu Ultra v2, two new models in its orchestration family designed for multi-agent tasks, available through an OpenAI-compatible API.
Meet Redis LangCache: A Managed Semantic Cache That Cuts LLM API Costs by Up to 90% and Returns Cache Hits Up to 15x Faster MarkTechPost
Meet Redis LangCache: A Managed Semantic Cache That Cuts LLM API Costs by Up to 90% and Returns Cache Hits Up to 15x Faster MarkTechPost
Google researchers introduced ToolGrad, an answer-first framework for generating high-quality tool-use training data, which improved pass rates to 99.8% on ToolBench.
NVIDIA introduced the BioNeMo Inference Runtime (BioIR), a Python library for accelerating protein structure prediction models like Boltz-2, achieving a 2.90x improvement in throughput.
OpenAI has launched its Agents API in public beta, providing a unified API call for agentic workflows.
Sakana AI launched Fugu Max and Fugu Ultra v2, two new models in its Fugu family designed for cost-effective and high-capability multi-agent orchestration.
LandingAI launched Agentic Document Extraction (ADE) Gen2, utilizing the new DPT-3 model family for improved document parsing and grounding.
Cohere released North Small Translate, a 218B parameter MoE model specialized in translation across 50 languages.