AI News from MarkTechPost

Latest coverage from MarkTechPost, summarized and scored for signal.

• NVIDIA Releases TensorRT Model Connect in Public Preview: Hugging Face Checkpoint to Native C++ Inference in Two Commands — NVIDIA released TensorRT Model Connect (TRTMC) in public preview — an Apache 2.0-licensed tool that converts Hugging Face checkpoints to native C++ TensorRT inference in two commands, skipping the ONNX export step and producing versioned .bundle artifacts for C++ services and embedded systems.
• Meet SAM (Sovereign Agent Mesh): A Zero-Config, Zero-Trust P2P Network for AI Agents — Google's open-source SAM (Sovereign Agent Mesh) lets AI agents running across clouds, datacenters, and edge devices share MCP tools via a zero-trust P2P overlay without exposing internal APIs to the internet — Apache 2.0 licensed and ships with Kubernetes support today.
• Cartesia Ships Sonic-3.6: A Streaming TTS Model That Now Leads Both Artificial Analysis Speech Arenas — Cartesia's Sonic-3.6 takes #1 on both Artificial Analysis speech leaderboards — 1,283 Elo on the Provider Voice board and 1,123 on the stricter Controlled Voice board — delivering sub-90ms time-to-first-audio at $49/1M characters, exactly half the price of ElevenLabs Eleven v3.
• Nous Research Ships Bot Mode for Hermes Agent, Turning Agent Profiles Into a Roster of Named Bots — Nous Research's Hermes Agent v0.20.3 ships Bot Mode by default — converting agent profiles into a roster of named bots with isolated memory and skills, communicating via @mention handoffs through a persistent Agent Inbox.
• ByteDance Seed and Tsinghua AIR Introduces CUDA Agent: A Large-Scale Agentic RL System for CUDA Kernel Generation — ByteDance Seed and Tsinghua AIR's CUDA Agent trains an LLM via reinforcement learning to write GPU kernels that outperform torch.compile on 96.8% of KernelBench tasks at 2.11× geomean speedup — the first RL system to close the correctness-vs-speed gap in CUDA generation.
• MiniMax Releases MiniMax-Music3: An Open-Weights Music Model Generating Complete Five-Minute Songs From Lyrics and a Structured Caption — MiniMax released MiniMax-Music3, an open-weights text-to-music model that generates full five-minute songs from lyrics and a structured caption, using a hybrid 8B+0.6B LLM with flow-matching synthesis — weights, inference code, and three serving paths shipped on day one under a commercial-friendly license.
• Developing an End-to-End Document Intelligence Pipeline with docTR for OCR, Layout Analysis, KIE, Benchmarking, and Searchable PDFs — A step-by-step tutorial for building production-ready document intelligence pipelines using docTR — covering GPU-aware OCR, layout analysis, key information extraction from invoices, and exporting results as searchable PDFs, JSON, and hOCR.
• DeepSeek AI Releases DeepSeek Harness in Developer Preview: An MIT-Licensed Agent Harness Where Everything is a Plugin — DeepSeek open-sourced Harness v0.1 under MIT — a Cordis-powered agent runtime where models, tools, loops, sandboxes, and UI are all swappable plugins, letting teams assemble custom agent pipelines without forking the core codebase.
• Fine-Tuning Tool-Calling LLMs: A Complete Guide Using XYZ-Aquila-SFT and Qwen3 — MarkTechPost walks through a complete supervised fine-tuning pipeline for Qwen3-0.6B using the XYZ-Aquila-SFT dataset and LoRA, covering tool-call parsing, ChatML formatting, and pre/post-training evaluation — a practical starting point for teams building tool-use agents.
• Z.ai Ships GLM-5.3 Without Retraining the Base Model: Better at Complex Coding and Long-Horizon Tasks — Z.ai released GLM-5.3 on August 14, 2026 — same 743B base model as GLM-5.2, all gains from post-training alone — with Terminal-Bench 3.0 jumping from 4.6 to 28.3 and CyberGym cybersecurity scoring reaching 84.5%, edging past GPT-5.6 Sol and Mythos 5 on that benchmark.
• Meet Needle 2: An Open 45M-Parameter Tool-Calling Model That Ships as a 14MB Binary and Runs a Full Session in 28MB of RAM — Cactus Compute releases Needle 2, an open 45M-parameter tool-calling model that ships as a 14MB binary and runs a full session in 28MB RAM, delivering 500 tokens/sec on a Raspberry Pi 5 -- making structured extraction and voice-to-action viable on zero-GPU, zero-NPU hardware.
• Create a Reasoning-Focused LLM: A Practical Guide to Streaming, Curating, and Fine-Tuning the SupraLabs Reasoning Corpus — MarkTechPost tutorial walks through building a compact reasoning-focused LLM by streaming 8,000 samples from the SupraLabs 5M-record reasoning corpus, applying quality filters, converting to chat format with explicit think tags, and fine-tuning SmolLM2-135M-Instruct using LoRA via TRL's SFTTrainer in Google Colab.
• Google AI Just Released Gemini 3.7 Flash: A Coding and Agent Model at $0.75/1M Input Tokens — MarkTechPost's technical breakdown of Gemini 3.7 Flash highlights $0.75/1M input pricing, a 1M context window, and coding gains (FrontierCode 43.6%, DeepSWE 65.3%, AutomationBench 30.4%) while noting GPT-5.6 Terra still leads on DeepSWE at 69.6%.
• Liquid AI Releases LFM2.5-VL-3B: A 3B Vision-Language Model That Reads Screens, Grounds Objects, and Calls Tools On-Device — Liquid AI's LFM2.5-VL-3B is a 3.1B-parameter on-device vision model that achieves 69.4 across 28 vision benchmarks — matching 4.7B models — while running at 228 tokens/s on an Apple M5 Max in 3GB of memory, with function calling and a 30-point grounding accuracy gain.
• Dyna Robotics Introduces Dyna-2: A World-Action Model Pre-Trained on 1 Million Hours of Human Video — Dyna Robotics has released Dyna-2, a world-action model pre-trained on over 1 million hours of egocentric human video — roughly 170 years of continuous footage — establishing the first scaling laws showing that more human video directly improves robot manipulation performance without robot-specific training data.
• SpaceXAI Releases Grok 4.6: A 500K-Context Frontier Model Tuned for Long-Running Agents, Coding, and Knowledge Work — xAI's Grok 4.6 arrives as a post-training upgrade over Grok 4.5, extending context to 500K tokens, scoring 61 on the AI Intelligence Index (tied with GPT-5.6 Sol Max), and targeting agentic workloads across coding, CAD design, and knowledge work — available now via xAI API, Cursor, and OpenRouter at $2/$6 per million tokens.
• AllenAI Open Instruct Tulu 3 Post-Training with SFT, DPO, RLVR, GRPO, and Verifier-Based Evaluation — MarkTechPost walks through a complete LLM post-training pipeline using AllenAI's Open Instruct framework — covering SFT, DPO, and GRPO with LoRA adapters on GSM8K math data — adapted to run on a single 16GB GPU by replacing distributed components like vLLM and DeepSpeed.
• NVIDIA AI Releases Nemotron 3.5 Lightning: A 30B Open MoE with 3B Active Parameters, and NeMo Switchyard Model Router — NVIDIA releases Nemotron 3.5 Lightning — an open 30B mixture-of-experts model with only 3B active parameters, 1M-token context window, and 4x faster output than comparable models — deployable on a single H100 under the permissive OpenMDW-1.1 license, alongside NeMo Switchyard, an open-source agent routing library.
• Xiaomi’s MiLM Plus Releases PROVE: Perception-Aligned Object Removal Metrics RC-S and RC-T With a Real-World Video Benchmark — Xiaomi's MiLM Plus team releases PROVE — two reference-free video object removal metrics (RC-S for spatial coherence, RC-T for temporal consistency) using DINOv2 features, plus a real-world benchmark, accepted at ACM MM 2026. No paired ground truth required; RC-S runs at 134ms per frame on an RTX 4090.
• The Video Production Stack Now Fits on One Desk: LTX-2.5 Launches as NVIDIA-Accelerated Open Weights World Model — LTX-2.5 launches as an open weights video world model for NVIDIA RTX local inference, generating 10-second clips in 6.8 seconds on 2×GB200 hardware — 7-10x faster than closed alternatives like Veo 3.1 (52-70s) and Kling 3.0 Pro (390s+), with no per-generation fees.
• Building and Validating a Quantitative Trading Strategy with OctoBot, Walk-Forward Backtesting, Parameter Optimization, and Interactive Analysis — A hands-on tutorial walks through building a complete quantitative backtesting pipeline using OctoBot 2.1.1 and OctoBot-Script in Google Colab — combining RSI oversold signals, EMA trend filters (50/200), and ATR-based adaptive stops — then validates the strategy on a separate 2023–2025 out-of-sample period to guard against overfitting.
• webAI Releases TwIL-LM: A 1.7B and 3B Formal-Logic Model Family for Autoformalization on Local Hardware — webAI releases TwIL-LM, a 1.7B and 3B formal logic model family for autoformalization and entailment checking that runs under 2GB on local hardware — but ships under a non-commercial license and is outperformed by Qwen3-8B on key benchmarks despite using a third of the parameters.
• Implementing a MiniMax-H3 Multimodal Video and Audio Generation Pipeline with ComfyUI APIs — A hands-on tutorial from MarkTechPost demonstrates building a fully programmatic MiniMax-H3 video and audio generation pipeline using ComfyUI's HTTP and WebSocket APIs in Python, covering text-to-video, frame-conditioned, and reference-image generation with automated hardware profiling.
• Meta AI Releases Muse Glimmer: A 30B Open-Weights Agentic Model That Runs on One Consumer GPU — Meta's Muse Glimmer squeezes a 30B agentic model into 24GB VRAM via 4-bit quantization, with DFlash speculative decoding delivering 3.1x faster inference than baseline — making it the first production-viable local agent at this capability tier under Apache 2.0.
• ByteDance Seed Introduces SeedRealtime: a Native Audio-Visual Full-Duplex LLM That Watches, Listens and Speaks in One Model — ByteDance's Seed team released SeedRealtime, a native audio-visual full-duplex LLM that processes continuous audio, video, and text in a single end-to-end architecture — with turn-taking handled inside the model itself rather than by an external voice-activity detector.