Parcha Labs has open-sourced AgentRun, an agent platform that layers self-improvement, durable execution, and a console on top of Claude Code, Codex, OpenCode, and pi, under the Apache-2.0 license.
Why it matters: AgentRun layers self-improvement, durable execution, and a console on top of existing harnesses such as Claude Code and Codex, and provides a concrete case where code review costs dropped from $6,451 per month to 18 cents per run.
Edge0 open-sourced Edge0-145B and Edge0-120B along with its unified inference framework and deployment tools, claiming Edge0-145B is the first open-source 145B-parameter on-device model. The company reports 29GB peak memory, 1,100+ tokens/s prefill, 80+ tokens/s decode and roughly 120W typical power on a compatible laptop, with 128K–256K context. It cites internal scores of 64.6 on DeepSWE 1.1, 86.52 on Terminal-Bench 2.1, 90.4 on GPQA-Diamond and 80.6 on IFBench, and says weights are on Hugging Face with day-one support for macOS, Windows and Linux.
Why it matters: Edge0 open-sources a 145B on-device model that reports 29GB peak memory and 80+ tokens/s decode, offering a concrete data point on how far local inference can scale.
IBM released two experimental preview models, granite-4.1-3b-LLMON-preview and granite-4.1-8b-LLMON-preview, derived from Granite 4.1 base checkpoints and trained to consume and generate LLMON, an LLM-native markup language that marks instructions, data and requirements with reserved special-token delimiters. On the OPI benchmark the models avoid 47% more indirect prompt injection attacks than the best frontier model tested and 95% more than the best open-source model, while performing on par with Granite 4.1 Instruct on standard benchmarks. The release includes a GitHub toolkit with lossless JSON-to-LLMON converters, and the authors note the preview is for exploration, not production.
Why it matters: IBM trained a preview model on Granite 4.1 that can read and write LLMON tags, using explicit instruction-data separation to reduce the indirect prompt injection success rate on the OPI benchmark to near zero while maintaining general capabilities.
AWS has open-sourced UAEF (Universal Agent Evaluation Framework), a framework-agnostic system that uses adapters to normalize agent traces from LangGraph, LangChain, Strands Agents, Amazon Bedrock AgentCore and Langfuse into one canonical AgentTrace, then scores them with 36 built-in metrics across 7 dimensions. Results are stored per experiment in Amazon DynamoDB and Amazon S3, so teams can compare agent versions against a baseline, flag regressions with a default 5 percent threshold, and generate statistical reports plus LLM-written insights. The post says customers across 12 organizations in telecom, healthcare and financial services used UAEF before release, and the Apache 2.0 code, Python library, CDK deployment and per-framework notebooks are in the awslabs/uaef GitHub repository.
Why it matters: The AWS team has open-sourced UAEF, which uses adapters to unify agent trajectories from different frameworks into a single format, then scores them across 7 dimensions using 36 metrics, providing a reusable evaluation layer for comparing agent versions across frameworks.
LangChain details how its open source Deep Agents harness uses middleware to fight context rot on long agent runs. Tool results over roughly 20,000 tokens and user messages over 50,000 tokens are offloaded to files with only a preview shown, while summarization condenses older messages once usage hits 85% of the context window. Skills, subagents and a markdown long-term memory file keep the main agent's window small and let memory carry over across models and providers.
Why it matters: LangChain explains how its open source Deep Agents harness uses middleware to offload, summarize and isolate context, offering a reusable pattern for agents on long tasks.
MiniMax released H3-Image, an open-weight image generation model built on H3 that supports text-to-image, reference-based generation and image editing, and can produce a 4096 × 4096 image in under 10 seconds. The company says it handles up to 16 reference images, claims state-of-the-art prompt adherence among open-weight models, and reports speeds up to 5× mainstream alternatives. H3-Image is available via the MiniMax API Platform, MiniMax Code and Hailuo AI, priced at 50–70% of comparable offerings, with 50% off 4K-and-above requests during launch week.
Why it matters: MiniMax releases open-weight H3-Image, claiming sub-10-second 4K generation and pricing at 50–70% of comparable mainstream offerings, giving developers a new open option for high-resolution image work.
Google released A2UI 1.0, the first stable version of its open protocol that lets AI agents answer with interactive screens rendered by each app's own native components. Version 1.0 adds bidirectional function calls, accessibility support and extension metadata, plus open-source renderers for React, Angular, Lit, Flutter, SwiftUI and Jetpack Compose (the Compose renderer is at 0.9.1 with 1.0 coming soon) and agent SDKs for Python, TypeScript and Dart. Atlassian is bringing A2UI to Jira, Alation renders its feeds with it in Claude, and ByteDance shipped a Lynx renderer in June 2026 that supports 0.9 today.
Why it matters: Google's stable A2UI 1.0 protocol lets agents return native interactive screens instead of text walls, with renderers for web, Android and iOS and early adopters including Atlassian and ByteDance.
Amazon introduced Mitra-v2, a second-generation tabular foundation model for classification and regression, released with fine-tuned models and inference and fine-tuning code under the Apache-2.0 license on Hugging Face. It is pretrained on synthetic tables with 10 times as many rows and roughly three times as many columns as Mitra-v1, using a new hybrid SCM prior and hybrid optimizer while keeping the Transformer architecture unchanged. On TabArena, a public benchmark of 51 real-world datasets, Mitra-v2 leads the single-model field, beats Mitra-v1 on 32 of the 33 datasets both run natively, and finishes all 51 within the one-hour budget; fine-tuned Mitra-v2 also ranks first on ScoringBench.
Why it matters: Mitra-v2 shows that widening synthetic pretraining data, rather than deepening the network, lifts a tabular foundation model to the top of TabArena's single-model field.
Google introduced on-device Automatic Speech Recognition and Text-to-Speech in Google AI Edge Gallery, letting apps run multilingual ASR and expressive TTS fully offline alongside on-device SLMs like Gemini Nano and Gemma 4. On Android, ML Kit's new speech models transcribe 100+ languages at 10x real-time speed with an 8.0% to 9.0% Word Error Rate on FLEURS, while the few-step diffusion TTS starts speaking in about 100 ms and generates 24 kHz audio at up to 50x real-time speed. Cross-platform support comes through MediaPipe Tasks and LiteRT with open-weight bundles including Parakeet-TDT 0.6B, PocketTTS 100M, Kokoro-82M and Phonon 31M, demonstrated in the XR Language Teacher showcase and voice-enabled AI Chat.
Why it matters: Google adds fully offline ASR and TTS to AI Edge Gallery, with ML Kit reporting 100+ language transcription at 10x real time and a diffusion TTS that starts speaking in about 100 ms.
NVIDIA introduced GR00T 2, an open world action model built on Cosmos and trained on 100,000 hours of human egocentric data, alongside Isaac Lab 3.0 and new robotics benchmarks at CoRL. GR00T 2 completed 40 of 50 evaluated tasks at least once without task-specific demonstrations and reached 71% average success across 50 tasks with one demonstration each, rising to 95% with five. Isaac Lab 3.0 adds Newton physics integration for deformable objects and reports up to 2.5x the training throughput of Isaac Lab 2.3 with PhysX in a vision-based dexterous manipulation benchmark.
Why it matters: NVIDIA's CoRL release shows how simulation plus world action models now target delicate assembly work, with GR00T 2 and Isaac Lab 3.0 details worth comparing against prior robotics stacks.
Zyphra released ZONOS3-Mini, a sparse MoE text-to-speech model with 100M active and 800M total parameters under Apache 2.0, trained on 112 NVIDIA H100s in 2.5 days. It reaches under 570 ms time-to-first-audio in the browser on a Pixel 9 and under 150 ms on an M3 MacBook Air, with runtimes for WebGPU, SGLang, vLLM and GGML. The model adds control over identity, accent, quality, mood and delivery, including word-level steering, and Zyphra reports it ranks first among open models on its new ZTTS2-Eval benchmark.
Why it matters: Zyphra's compact MoE TTS model pairs zero-shot voice cloning with word-level delivery control while running in real time on phones and laptops.
Google Cloud released long-context training support in MaxText v0.2.4, its open-source JAX training framework, adding four context-parallel strategies on Cloud TPU: All-gather, A2A (DeepSpeed Ulysses), P2P (Ring Attention) and the A2A + P2P hybrid (USP). Benchmarks with Llama 3.1 8B reached 70% MFU on TPU 5p at sequence lengths up to 4 million tokens, and context extended to 14 million tokens on both TPU 5p and TPU 7, with TPU 7 holding 50% MFU below 6 million tokens. The release is available on PyPI, in Docker images and on GitHub main, with sample recipes for TPU 7 and TPU 5p.
Why it matters: MaxText's four context-parallel strategies on Cloud TPU, with benchmark MFU figures for Llama 3.1 8B, give teams a concrete reference for choosing how to train at multi-million-token sequence lengths.
NVIDIA Research and Inferact's REINFORCE Zero paper shows that aligning rollout and trainer kernels lets unclipped token-level importance sampling correct the remaining weight lag, training stably for 20,000 updates at policy lag 64 and reaching 2.40x the training-step throughput of synchronous training with aligned kernels. With standard kernels, the median rollout-to-learner KL rises 83.33x across the run and online AIME24 pass@1 falls from 34.27% to 13.96%, while a sequence gate stabilizes training but rejects 24.48% of sequences by the last 1,000 updates. The zero-mismatch setup now covers DeepSeek-V4.1-Flash, Kimi K3.1 on release day, and DeepSeek-V4-Flash and Kimi K3.1 on AMD MI355X, with code in Inferact/zero-mismatch-rl and NVIDIA-NeMo/labs-molt.
Why it matters: NVIDIA Research and Inferact split the async RL rollout-learner gap into weight lag and kernel mismatch, showing that aligning kernels lets unclipped importance sampling hold stable training at policy lag 64.
Preferred Networks released PLaMo 3 Translate, a translation-specialized LLM used in its PLaMo Translation service, as an open-weight model on Hugging Face on November 5. 8B, 2B, 8B and 31B) and supports 53 languages including Japanese and English, running on smartphones, laptops and internal servers without a cloud connection.
Why it matters: PFN releases a translation-specialized LLM in four sizes from 0.8B to 31B, letting companies and individuals run 53-language translation on-device without cloud access.
Mistral AI released the open weights of Mistral Large 4 (Le Chonk), a 1-trillion-parameter natively multimodal model that entered public preview earlier this month and was trained from scratch in Europe. The company says three additional weeks of reinforcement learning lifted it to the top of open-weight coding, with 41% on Terminal-Bench 4.0 and 48% on the Artificial Analysis Coding Agent Index while using 47% fewer tokens per task than GLM-5.3. It adds Low, Medium and High reasoning modes, scores 51% on the Artificial Analysis Cyber Index, and is priced at $0.70 input, $0.07 cached input and $2.10 output per million tokens under a modified MIT license.
Why it matters: Mistral opens the weights of its 1-trillion-parameter flagship, pairing coding and defensive-security benchmark gains with three reasoning modes and per-token pricing.
Modal released Spindle, an open-source, Tinker-compatible training server that runs in a customer's Modal workspace, letting teams customize the stack and pay for GPU time instead of tokens. Spindle supports full-parameter training and multi-LoRA across open models, and Modal reports that sharing one trainer across 8 clients cut cost per 1M output tokens by up to 80% versus Tinker on a DAPO-math GRPO job. Modal says Spindle is still in beta and positions it between its Dojo training library and raw cluster primitives.
Why it matters: Modal's open-source Spindle server offers teams a Tinker-compatible alternative that trades per-token pricing for GPU time, with multi-LoRA sharing reported to cut costs by up to 80%.
MiniMax launched Mini App in the MiniMax Code desktop app, a plugin-format application with a frontend and a Node.js backend that keeps the Agent's work on a page that stays live instead of scattered across conversations. The release includes a runtime with a persistent data directory, hosted MCP, the Agent API, Host Connectors and in-page tool approvals, plus /miniapp-creator for going from a one-sentence request to a working page and official Mini Apps in the plugin marketplace. MiniMax Bot is itself a Mini App, and its source code along with community-built Mini Apps now lives in the open-source MiniMax-Code-MiniApps repository; Mini Apps are available in local sessions on macOS and Windows.
Why it matters: MiniMax Code 把 Agent 的工作从对话流搬进常驻页面，并公开了 Mini App 的运行时与开源仓库，为 Agent 工具形态提供了一种可参考的架构。
Mistral released Timestral, its first family of time-series foundation models, in four sizes from 5M to 209M parameters. The largest model ranks first on all three main public benchmarks, GIFT-Eval, TIME and fev-bench, among pre-trained zero-shot models from a single checkpoint without ensembling, using two-thirds of the parameters of the previous state of the art. Timestral produces probabilistic multivariate forecasts with 21 quantiles, supports past-only and known-future covariates, runs on CPU without a GPU, and is available through Mistral Forge for fine-tuning, with a GitHub repository and technical report published under a modified MIT licence.
Why it matters: Mistral's first time-series foundation model family claims top zero-shot rankings on GIFT-Eval, TIME and fev-bench with a 209M-parameter checkpoint, offering a size-accuracy tradeoff reference for forecasting work.
inclusionAI released Ming-3.0-audio-flash, a 126B-parameter Mixture-of-Experts audio model with 6.5B activated parameters, alongside the compact Ming-3.0-audio-tiny at 9.3B parameters with 2.7B activated. The flash model covers Chinese and English speech recognition across dialects and noisy multi-speaker settings, long-audio understanding, non-speech sound and music perception, audio reasoning, and end-to-end spoken dialogue with optional synthesized speech output. It ships with vLLM-Omni, SGLang, and Transformers inference paths under the MIT License, with default Flash profiles using tensor parallelism across four 80 GB H800 GPUs at 128K context.
Why it matters: The release details a 126B-parameter MoE audio model with 6.5B activated parameters, plus a 9.3B tiny variant, giving a concrete reference for open audio-language model sizing and deployment.
Goodfire researcher argues that the canonical evidence for polysemanticity and superposition in neural networks was misread, and that what looked like arbitrary packing is mostly a motif he calls recursive generalization. Retraining InceptionV1 with ordinary architectural improvements and no change to model size or training data made 99-100% of neurons in its last layer easily interpretable, and preliminary audits of Qwen3-32B estimate roughly 90% of neuron poles have one main job. The essay concludes networks use one learning principle rather than two, so interpretability can proceed by studying neurons and circuits directly instead of first training decomposition models such as sparse autoencoders.
Why it matters: A first-hand reinterpretation of the field's canonical polysemanticity evidence, arguing that networks generalize recursively rather than pack unrelated concepts, which would make direct mechanistic study viable.