AI Shorts

Pizza Bot is an open source, self-hosted inbox for AI agents built on DeepAgents and LangGraph. It combines persistent task state, MCP integrations, configurable approvals, and scheduled workflows across multiple model providers.

A shallow agent is an LLM calling tools in a loop, and on long tasks it fails in 2 ways: context overflow and goal loss. This article opens the harness layer that fixes both, with the actual thresholds shipped by LangChain Deep Agents, Claude Code, Manus, OpenAI Codex and Amazon Bedrock AgentCore, plus an interactive simulator that shows a 200K window filling up.

Cognition, the company behind the Devin coding agent, has released SWE-2, its most capable coding model to date. SWE-2 is post-trained with reinforcement learning...

The Fly Language Model (FLM) drives all 166,700 retained neurons and 25.6 million edges of the MaleCNS fruit fly connectome with token embeddings, then adds a small learned correction to a frozen LFM2.5-1.2B-Instruct backbone. Only 278,528 parameters train. The accompanying preprint reports a 0.0222 nat per token improvement over the backbone, but a parameter-matched control without the graph does slightly better in every seed. We break down the architecture, the four controls, the 0.6 per token memory bound, and what it takes to run the MIT-licensed code locally.

ByteDance Seed, SUTD, Georgia Tech, M-A-P, and TokenWave.AI introduce HarnessDev, a benchmark that scores the runnable harness a model builds rather than the answer it returns. Starting from a seed that scores 0, 6 creator LLMs construct harnesses across 5 benchmarks and 2,207 tasks, then evolve them from execution feedback. Self-built harnesses match human references on writing and ML experimentation but trail on code and search, and only 34 of 64 evolution changes move the same direction on held-out tasks.

Anthropic has published a new plugin evals workflow for Claude Code. The claude plugin eval command runs a plugin against realistic prompts, grades what...

Cohere has released North Small Translate, an open-weight Mixture-of-Experts model built for machine translation across 50 languages. It uses 25B of its 218B parameters per token and scores 83.6 on Cohere's WMT26 evaluation. Weights are free for non-commercial use, with commercial access through Cohere Model Vault or RWS Language Weaver.

Sakana AI has released Fugu Max and Fugu Ultra v2, 2 models built on the same learned orchestration architecture. Fugu Max routes tasks to lean open and specialized models, including NVIDIA Nemotron, at $2/$6 per 1M tokens. Fugu Ultra v2 targets peak capability, scoring 48.3 on Chartography and 74.3 on DeepSWE.

Google Research has released ToolGrad, an ACL 2026 Findings framework that inverts tool-use dataset generation: it builds a verified API chain first, then writes the matching user query. Guided by textual "gradients" from a 4-module propose-execute-select-update loop, ToolGrad reaches a 99.8% pass rate on ToolBench versus 63.8% for DFS search. Gemma-3-12B fine-tuned on only 500 samples scores 83.1 on BFCL, next to Gemini 2.5 Pro at 83.2. Code, dataset, and models are public under Apache-2.0.

Production LLM applications rarely receive a question nobody has asked before. Support assistants and RAG pipelines field the same intents thousands of times a...

NVIDIA has detailed BioNeMo Inference Runtime (BioIR), a Python library that accelerates biomolecular structure-prediction models on NVIDIA GPUs while staying in plain PyTorch. In a matched benchmark on 1,000 human dimer targets across 8xH100 GPUs, BioIR-accelerated Boltz-2 delivered 58.5K successfully folded residues per GPU-hour versus 20.2K for a torch-compiled open-source implementation, a 2.90x gain. The runtime optimizes at 3 layers: custom kernel selection, CUDA Graph capture, and Ray-based replica scaling that places 1 full model copy per GPU. BioIR already powered the AlphaFold Database expansion, generating about 31 million candidate protein complexes across 4,777 proteomes.

OpenAI has released the Agents API in public beta. It gives developers the same harness and infrastructure that run Codex. OpenAI hosts and maintains...

Long-horizon agents have turned LLM serving into an input-heavy workload. Repeated prefills and million-token contexts leave KV caches that strain HBM, SSD capacity, and...

LandingAI has shipped Agentic Document Extraction Gen2, a rebuild of its document stack on the DPT-3 model family. Chunks are retired in favor of a document, page and block tree. DPT-3 Pro grounds to the line, DPT-3 Verity grounds to the word with a confidence score, and Parse billing now counts output characters instead of flat pages. Gen1 code will not run against Gen2 endpoints.

Google has open-sourced Mantis, a stack-agnostic toolkit of security review skills for AI coding agents. It runs the full vulnerability lifecycle: sweep the code, filter false positives, reproduce the bug in a sandbox, patch it, re-attack the patch, then score the risk. Apache 2.0, and documented as demonstration-only.

Today, Meta has introduced Muse, a personal AI agent that takes actions rather than just answering questions. Muse can send emails, book travel, negotiate...

NVIDIA has announced CUDA Rust, a push to make Rust a first-class language for GPU kernels. Two open-source NVlabs projects cover the two CUDA programming models: cuda-oxide compiles SIMT kernels from Rust MIR through Pliron and LLVM to PTX, while cutile-rs JIT-compiles Tile kernels through CUDA Tile IR on stable Rust 1.89+.

DeepMind's AlphaGenome Atlas maps every single-letter change in the human genome with 1 impact score per variant.

We look at r-1, the document parsing model Reducto released on September 1, 2026. We walk through how it folds OCR, layout detection, tables, formatting and grounding into one full page pass, replacing the multi stage agentic pipeline it ships alongside. We break down the two numbers that matter for a migration decision: a reported 20% error reduction and a flat 1 cent per page rate against the legacy 3 to 6 cents.

OpenBMB has released MiniCPM5-2B, a dense causal language model with 2,516,756,480 parameters and a native 131,072 token context. It averages 53.9 across the 34 benchmarks in its model card, ahead of Qwen3.5-4B at 51.1, with its clearest leads in tool use, coding agents and long-context retrieval. Post-training pairs 400B tokens of deep-thinking SFT with RL teachers and on-policy distillation that merges 16 expert models into one checkpoint. The weights ship under Apache 2.0 alongside the pre-training, SFT and RL datasets and the intermediate Base, Midtrain and SFT-only checkpoints. GGUF builds start at 1.56 GB, and the standard LlamaForCausalLM architecture loads in vLLM, SGLang, llama.cpp, Ollama and MLX without a model-code fork.

Robot datasets have grown far slower than the models trained on them, mostly because collection stays locked to lab hardware. AXIS moves demonstration collection into a web browser and pushes everything expensive to backend GPUs. The result is 207 tasks and 50,129 verified Franka trajectories, and continual pretraining that lifts π0.5 from 83.9 to 88.8 on LIBERO-Plus while a volume-matched RoboCasa365 control reaches only 57.5.

Most open model launches release one checkpoint and a benchmark table. The Institute of Foundation Models (IFM) released something wider last week. IFM is...

AI research agents can propose far more experiments than they can afford to run. Meta FAIR, Oxford and UCL introduce AI Research Preference Models — frozen LLM judges that rank 15 unexecuted candidates and execute only one. On AIRS-Bench, the average normalized score rises from 0.684 to 0.729, and the baseline's 24-hour result arrives in roughly 15 hours.

Training and benchmarking a computer-use agent needs four things — agents, environments, traces, and a framework to evaluate and train them — and all four ship in incompatible formats today. CUA-Lite, from a UC Berkeley led team, puts them behind one action space and one data schema, and replaces OSWorld's per-task virtual machine with a plain Docker container at 0.9 GB instead of 4.1 GB.

Retrieval quality in an AI search product is bounded by two things: how good the embedding model is, and how cheaply you can run...

We look at Project HydraFusion, GitHub's research preview that treats workflow selection as an optimization problem rather than a model picker. We break down the three execution patterns it routes between — Single, Cascade with a quality gate, and Critique with a read-only cross-family reviewer.

Adaption Labs has released Invent a Dataset, which generates a structured, training-ready dataset from a description of the behavior you want a model to learn. There is no seed corpus, no schema design, and no labeling guide. A single datasets.invent call sets domains, row count, output format, and language expansion, and the rows download as JSONL, JSON, CSV, or Parquet. The dataset ID then passes straight into AutoScientist, closing an intent-to-trained-model loop.

Gemini now navigates video instead of ingesting it at 1 FPS, loading only the segments a prompt needs.

We look at NVIDIA Personal AI Router (PAIR), an open source virtual inference router that spreads local AI requests across the machines already on a home network. We cover how PAIR proxies existing Ollama and LM Studio endpoints so agent harnesses need no changes, and how its scheduler filters nodes on readiness, engine state, exact model presence, job load, and GPU utilization. We walk through NVIDIA's five-subagent demonstration, which averaged 18 minutes on one RTX Spark laptop against 8 minutes 48 seconds on a three-device cluster, and note why NVIDIA labels it unofficial rather than a benchmark. We also cover where PAIR will not help, including its single scheduling policy and its blindness to VRAM and model warmness.

WeatherNext 3 ingests live geostationary satellite mosaics, refreshes hourly, and outputs 5 km forecasts across Search, Gemini, Maps.

OpenAI released GPT-6 Astra on September 3, 2026, positioning it as a computer-use flagship rather than a chat model. It reports 72.6% on OSWorld V2-Offline, replaces Codex compaction with searchable notes, and ships a 1.05M-token context at $10/$50 per million tokens. It is also the first OpenAI model to cross the Critical cybersecurity threshold, which shapes who can use it and for what.

Most teams building a shopping assistant or agent rebuild the same scaffolding: an agent loop, a tool layer over the catalog, an approval gate,...

Perplexity has shipped hybrid compute for its Mac app, splitting a single Perplexity Computer task between frontier models in the cloud and a compact model running on the user's machine. Tasks start in the cloud for search, planning and reasoning, then hand sensitive steps down to the Mac without restarting or losing context. An on-device privacy gate decides what may cross: keep local, mask, refuse, or ask for consent. Perplexity also open-sourced the 0.6B classifier behind that gate, which records the highest character F1 (0.629) among twelve detectors evaluated. Hybrid compute is live now for Pro, Max and Enterprise on Apple silicon Macs with 24GB of unified memory.

Perplexity has open sourced Lily, the local inference engine behind Hybrid Compute in Perplexity Computer. Built in Rust with custom Metal kernels for one model on one chip family, it averages 1.23x MLX-LM's prefill throughput and 1.35x its decode throughput on a 40-core, 128 GB M5 Max.

We look at zg (zvec-grep), the local-first search layer that Qwen Developers just open-sourced under Apache 2.0. We explain how it puts ripgrep, BM25, and vector search behind a single interface, so an agent can move from a plain-language description to an exact line span without switching tools. We cover the deliberately small MCP surface, the on-device embedding catalog, and the authorization gate that stands between local content and remote models.

NVIDIA has released Switchyard, an Apache-2.0 Rust proxy and library for LLM traffic. It decodes requests into provider-neutral types, routes them with passthrough, random, LLM-classifier, or stage-router algorithms, and translates responses back into the client's format, so Claude Code or Codex CLI can run against vLLM, NIM, or Ollama unchanged. The project is pre-alpha and not for production.

Google released Gemini 3.8 Flash and Gemini 3.8 Flash Cyber on September 2, 2026. Both variants run on the same foundational intelligence, split by safety mitigations rather than model size. Gemini 3.8 Flash is generally available at $0.75 and $3.75 per 1M tokens, introductory through December 31, 2026. Flash Cyber reaches 47.2% pass@1 on CWE-Bench and is restricted to vetted defenders through the Fairwind Program. This article covers the benchmarks, the token-for-accuracy tradeoff, and what deploying either one actually requires.

Anthropic announced Enterprise Frontier Safeguards on September 1, 2026, an architecture that stores monitoring data in the customer's own cloud account rather than Anthropic's. Detection stays automated and Anthropic-run; custody, encryption keys, and flag review stay with the customer. Built with more than 100 enterprises and rolling out in phases this fall.

Most production voice stacks are three systems stitched together. One model transcribes, a second separates speakers, and a detector decides when the user stopped...

Agentic assistants have a structural problem: the context that makes them useful — deal documents, privileged files, client records — is exactly the context...

Anthropic has released Claude Fable 5.1 and Claude Mythos 5.1, the same model behind two different safeguard layers. Fable 5.1 is generally available on the Claude API, AWS, Google Cloud, and Microsoft Foundry; Mythos 5.1 remains restricted to vetted organizations under Project Glasswing. Fable 5.1 scores 52.6% on Terminal-Bench-Science 0.1 against 24.7% for Fable 5, ships a 1M token context window, and drops cache reads 75% to $0.25 per million tokens while base pricing holds at $10 and $50. Three API changes are breaking, including the removal of forced tool use.

Quantitative research agents that write their own experiments can corrupt the evidence they later learn from. A leaky feature that scores well gets stored...

Speed and accuracy usually pull against each other in text-to-speech. Gradium AI's new default model reports both: an 81.0% human-rated pass rate on 500 hard sentences across five languages, at 216 ms P50 time-to-first-audio on Coval. The evaluation set is open on Hugging Face under CC BY 4.0.

How do you benchmark a web search API when the thing being tested can read the answer key? A search agent has a fetch...

Google Research has released TimesFM-3, a 330 million parameter time series foundation model that forecasts multiple related series in a single forward pass. Unlike every TimesFM checkpoint through 2.5, it is pretrained natively for multivariate forecasting, accepting multiple targets, past covariates, and past-future covariates with no task-specific fine-tuning. It takes the top average rank among pretrained foundation models on GIFT-Eval, fev-bench, and the TIME leaderboard. The weights, however, ship under a non-commercial, non-production license.

The OpenClaw Foundation has released v2026.8.1, which the project calls OpenClaw 2.0: 933 contributors, 569 first-timers, and more than 16,000 pull requests, roughly half of every PR ever merged into the repo. Setup now reuses existing subscriptions, API keys and local models. The rebuilt Control UI cut test-harness startup from about 1.6 seconds to 575 ms. Shared cloud sessions add real multiplayer, though the docs are explicit that they are not a security boundary.

Google Cloud AI Research, with Washington University in St. Louis and UNC Chapel Hill, has released EnvHarness, an Apache-2.0 layer that turns a static agent benchmark into one that adapts to the policy training on it. It wraps a frozen environment through the standard reset()/step() interface, so tasks and human-built verifiers stay untouched — and an LLM designer, EnvRigger, writes those wrappers automatically against flaws diagnosed in the agent's own rollouts. Across five benchmarks, mined skills gain up to 9.0 points on held-out tasks with 9.8% fewer execution steps.

Anthropic has opened a research preview of the Model Hardware Standard (MHS), a shared driver specification that lets AI agents discover and safely operate physical devices. Instrument integration that normally takes weeks or months drops to hours: Carnegie Mellon went from raw equipment to a finished dose-response curve in eight, and QuEra's laser relock improved from 58% to 99.3% across 700 trials. MHS is model-agnostic, reachable over MCP, and enforces safety limits in the driver rather than the prompt.

Code-as-World recovers editable MuJoCo scene code from real video, then uses those verified worlds to train physical reasoning.

We look at Gemini Omni 1.1 Flash, Google's production update to its native multimodal video generation and editing model. We break down what changed: scene extension now reads up to 10 seconds of prior context instead of a single final frame, first and last frames can be pinned to control camera movement, and video clips can be passed as references for character consistency.

Pollen Robotics, the Bordeaux robotics team at Hugging Face, opened pre-orders for Microduck — a 25 cm bipedal robot where every movement is a neural policy trained in MuJoCo and exported to ONNX. At $399, it puts the full sim-to-real loop on a desk: 15 motors, camera, LiDAR, two IMUs, and an Apache-2.0 training stack you can retrain yourself.

Vercel has open-sourced vgpu, the WebGPU library it built to ship the shaders on vercel.com. It treats .wgsl files as importable TypeScript modules, runs the same shader in the browser, in headless Node.js via Dawn, and in a deterministic CI mock, and ships a fullscreen effect in 25 KB gzipped.

Google has released Gemini 3.5 Transcribe, a speech-to-text model that ships as two separate endpoints rather than one. The streaming endpoint delivers sub-second transcription but drops speaker diarization and word timestamps. The batch endpoint keeps both, at half the cost. Google reports 4.0% word error rate streaming and 2.6% non-streaming, with 70% faster finalization than Chirp 3. Here is what the split means for anyone building voice agents or transcription pipelines.

Cohere has released Parse (parse-v5.0), a 2.3B-parameter vision language model that converts PDFs, slides and images into Markdown with HTML tables, bounding boxes and image descriptions. It runs at $1.50 per 1,000 pages through the API, or on dedicated Model Vault instances from $2,500 a month. Cohere reports a ParseBench score of 79.2, ahead of Mistral OCR 4, Azure Document Intelligence and Databricks AI Parse — but that figure averages three of the benchmark's five dimensions and drops charts and visual grounding entirely.

Google Research and UNSW Sydney released GlucoFM, a self-supervised foundation model that splits a CGM trace into a slow physiological stream and a transient event stream instead of encoding it as one sequence. At 0.72M parameters it reached 58.8 task-averaged PR-AUC across 14 cohort–task evaluations, beating a 135M GluFormer and a 385M MOMENT. It remains a research prototype with no regulatory clearance.

Z.ai has released GLM-5.3-Flash, the first natively multimodal model in the GLM-5 series — a 320B-total / 18B-active MoE with a 1,048,576-token context window, MIT-licensed weights on Hugging Face, and API pricing at $0.15/M input and $0.50/M output. It scores 84.3 on Terminal-Bench 2.1 and 63.4 on DeepSWE v1.1, using hybrid KDA linear plus NoPE sparse MLA attention to cut attention compute ~3× and KV cache 4.4× versus GLM-5.3.

We look at Qwen3.8-Flash-Next, Alibaba's open-weight multimodal Mixture-of-Experts model and an early preview of the Qwen4 architecture. We break down where the 180B parameters actually sit: a 125B backbone, a 51B N-gram embedding table, and a 4B multi-token prediction module, with only 6B active per token. We walk through the four architectural changes — the Gated DeltaNet and Qwen Sparse Attention hybrid, Gated Residual, N-gram Embedding, and the Muon optimizer. We also cover the benchmark results, the reported 1/9 training cost against Qwen3.7-Plus, and what self-hosting a 172.78 GiB FP8 checkpoint really demands.

Four falsifiable conditions for agentic coding replacing juniors, tested against METR, OpenAI, DORA and Stanford primary source evidence

Model cards report quality under server-class, full-precision conditions. Those numbers rarely predict how the same model behaves on a phone. This week, Liquid AI...

Perplexity releases Portable Computer, packaging local models, harness, sandbox, and connectors into one system running on NVIDIA DGX Spark.

Training and serving frontier models is now a networking problem as much as a compute problem. Collective operations like all-reduce and all-to-all synchronize thousands...

Fastino released GLiNER2.5, replacing span enumeration with boundary prediction so entity width no longer costs compute. Three Apache 2.0 checkpoints ship at 74M, 194M, and 287M parameters, all CPU-runnable. The release adds joint entity-relation decoding, constrained classification, span attributes, and 4,096-word context. Overall macro F1 reaches 56.17 on 16 zero-shot benchmarks.

Generalist AI has released GEN-1.5, a robot foundation model that learns a new physical task from a single demonstration. Drop 3–12 seconds of sensorimotor...

framework that folds aggregate human movement into text-based place embeddings. Language models describe what a place is; they miss how it is used. ME-POIs encodes each visit as a contextualized vector and aligns it with one learnable prototype per POI through contrastive learning, then transfers visit distributions from data-rich anchors to the long tail across three spatial scales. Across five map-enrichment tasks on Los Angeles and Houston mobility data, adding ME-POIs improved 34 of 35 model-task pairings in Los Angeles — up to 81.9% relative F1 on visit intent and a 24.7% MAE reduction on busyness. A mobility-only variant beat Gemini text embeddings on price-level classification.

Harvey's first post-trained model nearly doubles LAB task completion, but only one benchmark number survives independent verification today

FreeToken splits MoE cache misses between PCIe fills and CPU execution using measured bandwidths, unlocking frontier models locally

Most teams treat 'which model' as the important decision. The harness engineering literature keeps pointing somewhere else. In LangChain's Terminal-Bench experiment, changing only the...

S1-mini is a 462 MB open-weights normalizer that sits after ASR, removing fillers and resolving self-corrections locally.

PDFs are easy to read and hard to change. AI can now summarize a 90-page contract in seconds, but it still won't rewrite the source file cleanly. UPDF is built for that second half: direct editing, 14-format conversion, 38-language OCR, and ten AI agents shipped in version 2.5.

Three ~300M drafters bring speculative decoding to LFM2.5, delivering up to 3.18x faster decoding with identical greedy output.

NVIDIA has released TensorRT Model Connect (TRTMC) in public preview, an Apache-2.0 project that takes a supported Hugging Face or local checkpoint to end-to-end TensorRT inference in two commands, with no intermediate ONNX export. The build emits a versioned .bundle artifact that runs through native C++ task APIs, so inference executes without PyTorch in the runtime path. NVIDIA's July 29, 2026 GB300 snapshot covers 105 release profiles across 76 model families.

Cartesia has released Sonic-3.6, a streaming text-to-speech model built on state space models rather than transformers. It now ranks #1 on both Artificial Analysis speech leaderboards — 1,283 Elo on Provider Voice and 1,123 on Controlled Voice, the board that clones every model onto the same eight reference voices to isolate the synthesis engine. Cartesia states sub-90ms time-to-first-audio. The model is available in beta on Cartesia's own API

Nous Research has shipped Bot Mode for Hermes Agent, its MIT-licensed open source agent. Bot Mode replaces the single-agent session list with a roster of named bots. Each bot is a real Hermes profile, with its own chat, memory, skills, and pinned model. It is now bundled and default-on in Hermes Desktop.

ByteDance Seed and Tsinghua AIR have released CUDA Agent, an agentic reinforcement learning system that trains a large language model to write GPU kernels...

MiniMax released MiniMax-Music3, an open-weights text-to-music model. Given lyrics with section tags and a structured caption, it generates a complete song of up to five minutes in a single pass, as 32 kHz, 16-bit stereo WAV. Here is the architecture, the three serving paths, and the license conditions that matter before you ship.

DeepSeek Harness v0.1 is an MIT-licensed agent harness where every capability is a Cordis plugin. Four runtime modes, append-only session logs, and provider-agnostic model routing.

Z.ai released GLM-5.3 on August 14, 2026. The model reuses the 743B GLM-5.2 base unchanged. Every reported gain comes from scaled post-training: more long-horizon task environments, more environment types, longer training. Terminal-Bench 3.0 moves from 4.6 to 28.3, and DeepSWE v1.1 from 46.2 to 66.9. Cybersecurity moved further than Z.ai says it planned, with CyberGym at 84.5% and ExploitBench more than doubling to 54.4%. Weights arrive in about two weeks.

Cactus Compute released Needle 2, an open 45M-parameter model for tool calling, device use, and structured extraction. The full model is a single 14MB binary that runs a session in about 28MB of RAM. It leads both Seal-Tools splits while targeting hardware with no GPU and no NPU.

Google has released Gemini 3.7 Flash, a refinement of Gemini 3.6 Flash with algorithmic improvements to its reasoning core. It handles text, images, audio, and video across a 1M-token context window with 64K-token output, and supports customizable thinking configurations. Coding results move notably: 43.6% on FrontierCode 1.1 Main versus 34.4%, 65.3% on DeepSWE v1.1, and 1588 Elo on WebDev Arena. Document and workflow evals gain more — GDP.pdf goes from 22.0% to 34.0%, AutomationBench from 17.0% to 30.4%. GPT-5.6 Terra still leads on terminal and computer-use agents. The model is API and enterprise only, with no open weights, at an introductory $0.75/$3.75 per 1M tokens through December 31, 2026.

Liquid AI released LFM2.5-VL-3B, a 3.1B-parameter vision-language model built for on-device deployment. It averages 80.7 on ScreenSpot-v2 and lifts RefCOCO grounding from 57.1 to 87.9. Function calling is new to the VL line, with ToolSandbox moving from 26.4 to 59.5. The model fits in roughly 3 GB and decodes 228 tokens/s on an Apple M5 Max.

Dyna Robotics has released Dyna-2, a world-action model pre-trained on more than one million hours of egocentric human video. The technical report establishes three results: a scaling law on human data to 1M hours, the first transfer of that law to unseen robot data, and evidence that video co-training drives cross-embodiment generalization.

SpaceXAI released Grok 4.6 on August 12, 2026 — a post-training upgrade over Grok 4.5, not a larger base model. It ties GPT-5.6 Sol Max at 61 on the Artificial Analysis Intelligence Index, ships 500K context and a new xhigh reasoning level, and holds pricing at $2/$6 per million tokens. The coding benchmarks are where it still loses.

NVIDIA's open 30B MoE targets the agent execution layer, with Switchyard routing each step to the cheapest capable model.

Object removal models have improved faster than the metrics used to judge them. Diffusion erasers now reconstruct shadows, reflections and occluded structure convincingly, yet...

LTX-2.5 brings frontier video generation to local NVIDIA hardware: 6.8-second clips, native multishot, day-one ComfyUI, open weights.

webAI has released TwIL-LM, a family of formal-logic models at 1.7B and 3B parameters that translate English into first-order logic and check whether conclusions follow from premises. The 3B runs on CPU or 4GB of VRAM; the 1.7B downloads at 1.06GB. Both ship under a non-commercial license. The model card also shows that the headline benchmark scores belong to an unreleased checkpoint, not the weights on the Hub.

Meta's Muse Glimmer is a 30B open-weights agentic model under Apache 2.0. It fits 24 GB VRAM and decodes 3.1x faster with DFlash speculation.

ByteDance's Seed team has introduced SeedRealtime, a native audio-visual full-duplex LLM. The model fuses audio, video and text in a single unified architecture. It...

NVIDIA releases NemotronLabs VoiceChat 11B, an open full-duplex speech-to-speech model with 448 ms latency and live tool calling.

Long agent runs accumulate state that no transcript records — edited files, a live dev server, installed packages, a warm prompt cache. When an agent misreads a traceback at step 10 and rewrites a correct file, patching forward burns tokens and restarting re-pays every call. Researchers at Northeastern University and Stanford University released Shepherd, an MIT-licensed Python runtime substrate that records every agent-environment interaction as a typed event in a Git-like execution trace. Each commit covers the agent process and filesystem together, copy-on-write, so a rewind restores live state instead of just files. The paper reports 5× faster forks than Docker, over 95% prompt-cache reuse on replay, and a live supervisor raising CooperBench pair-coding pass rates from 28.8% to 54.7%.

Pokee AI released Pokee-Isaac 28B, a 28B text-only foundation model with a 10M-token context window built to run inside the customer boundary. It scores 93.3% on RULER at 10M tokens, where every baseline in its comparison panel returns 0.0 beyond 2M, and leads BFCL v4 at 70.94 while placing second on Terminal-Bench 2.1. Prefill reaches 137,200 tokens/s at full context on a single B200, with decode flat near 335 tokens/s. Weights are not published; deployment is licensed into VPC, on-premises, or on-device, with list pricing at $0.15/$1.00 per million tokens.

Mistral AI has released Shieldstral 1.0 3B, an open-weights, policy-adaptive multimodal safety classifier that frames content moderation as a single yes/no question instead of a fixed harm taxonomy. Operators supply the policy as a plain-language query at inference time and get back a calibrated safety score from one forward pass — no retraining required to re-target the model. Built on Ministral-3-3B-Base-2512 with a Pixtral vision encoder and trained on roughly 54.1M samples, it reports 84.9% average F1 on text safety (matching GPT-OSS-Safeguard-20B), 83.8% on multimodal safety, and 91.3% on Mistral's adaptability benchmark — while fitting in 16GB of VRAM under an Apache 2.0 license.

Tencent Cloud has open-sourced TencentDB Agent Memory v2.0, a team-level memory hub that turns conversations, documents and code into four governed, reusable assets — Chat Memory, Skill, LLM-Wiki and Code-Graph. It is MIT-licensed, self-hosted via Docker, and integrates with Claude Code, OpenClaw, Hermes and CodeBuddy. The differentiator is not retrieval but governance: ACL-based visibility decides which agent gets which asset, and which version is valid.

Microsoft has open sourced code-testing-generator, a polyglot unit-test agent shipping in the MIT-licensed dotnet/skills repository. It reads a repository before writing anything — detecting the language, test framework, existing conventions, and the real build and test commands — then plans, writes, runs and validates the tests it produces. On Microsoft's internal 152-task benchmark it completed 140 tasks against 120 for stock GitHub Copilot on the same model, with the gain concentrated almost entirely in vague prompts and diff-targeted requests.

Liquid AI released LFM2.5-2.6B, an agentic model that plans, calls tools, and completes multi-step tasks entirely on-device. The 2.69B parameter model pairs 22 double-gated short convolution blocks with 8 GQA blocks across 30 layers, handles 131,072 tokens of context, and decodes at 220 tokens/s on an M5 Max in under 2.5 GB. Open weights ship in GGUF, MLX, and ONNX.

Cloudflare has released Kitesurf, a stateless web browser built specifically for AI agents that runs entirely in V8 isolates on Cloudflare Workers, with no Chromium underneath. The browser drops human-facing features like tabs and extensions in favor of what agents need: machine-readable content, scalability, and isolation. Built in 12 weeks using Rust components like Blitz, Stylo, and Boa JS, it already passes 215,000+ Web Platform Tests. Benchmarks show 3.1–3.8× less CPU and 4.7–7.0× less memory than Chromium on screenshots and HTML extraction. Existing Puppeteer, Playwright, and MCP clients work by adding a single browser=kitesurf parameter, free while in beta.

Prime Intellect has open-sourced Prime Agent, a coding and research harness built on two abstractions: the Recursive Language Model, which turns sub-agent calls into functions inside a persistent IPython kernel, and the Continual Harness, which lets the agent edit its own prompts, skills, memory, and sub-agent specs mid-run. With Opus 5 it reports 95.5% RHAE Best@1 on ARC-AGI-3, above the reported human expert baseline of 95.4%.

Meta Superintelligence Labs has released Muse Code, a terminal coding agent in beta, powered by the new Muse Spark 1.2 model. Muse Code plans changes, writes code, and validates results across large repositories. Async background agents stay active for the whole session instead of spawning per task. A local append-only event log makes the runtime replay-exact and restart-safe after a crash. Muse Spark 1.2 was co-trained with the harness and trained on long-horizon, repository-scale work.

NVIDIA released Alpamayo 2 Super, a 34B vision-language-action model for autonomous driving, under OpenMDW-1.1 — a permissive license covering fine-tuning, derivatives and commercial redistribution. It pairs a 32B Cosmos 3 Super Reasoner backbone with a 2.3B diffusion action decoder, scores 79.2 on LingoQA, and emits trajectories, Chain-of-Causation traces, meta-actions, auto-labels and grounded VQA from a single pass.

CopilotKit has published the Channels SDK, an MIT licensed library that runs an existing AG-UI agent inside Slack and Microsoft Teams. Version 0.5.0 ships five platform adapters and a documented runtime contract. This breakdown covers the verified deployment paths, the baseline requirements, and the one dependency that is easy to miss