English text is being prepared.

LMSYS uses Miles to train a SWE agent on MI355X, SWE-bench Verified…

LMSYS used Miles, SGLang, Megatron-LM, and Harbor to perform fully asynchronous agent reinforcement learning on Qwen3-30B-A3B-Instruct on AMD Instinct MI355X, raising pass@1 from 7.3% to 27.5% on a 100-task stratified subset of SWE-bench Verified. Training was conducted in two stages: first SFT on approximately 4,000 public trajectories from Open-SWE-Traces, then GRPO on approximately 1,350 SWE-smith tasks. After fixing the model and RL recipe, scaling inference from 1 node to 3 8-GPU nodes, switching to E2B cloud sandboxes, and removing the concurrency limit increased rollout throughput from approximately 860 to approximately 3,480 trajectories/hour; limiting trajectory length on the same 8-GPU node reduced single-step training time from 850 seconds to 284 seconds.

AAIF: An Adaptive Agent Framework Based on Strands Agents

AWS built the Adaptive Agent Intelligence Framework (AAIF) with Strands Agents, using llama3.2:1b as the orchestration layer to classify requests: routine tasks are handled by a local SLM via Ollama, while complex analysis is routed through boto3 to frontier models on Amazon Bedrock. This architecture can run on edge devices such as a Raspberry Pi with 8 GB of memory, or be deployed fully in the cloud or in a hybrid setup, thereby reducing API costs while maintaining low local latency.

How to integrate the ElevenLabs text-to-speech API into an app

ElevenLabs has published a tutorial on how to integrate the text-to-speech API into an app using Python and TypeScript: create an API key, install the ElevenLabs SDK, send text to the TTS endpoint and specify a voice and model, then play the returned audio. The example uses the eleven_v4 model, and the low-latency Eleven Flash v2.5 is also available; the Voice Library offers more than 17,500 voices, and output defaults to mp3 but can be switched to ulaw or pcm.

Interview with XBP Global CEO Andrej Jonovic: How Agentic AI Lands in Mission-Critical Processes

In the interview, XBP Global CEO Andrej Jonovic said the company is applying agentic AI in regulated environments such as healthcare, the public sector, and financial services, using adaptive process orchestration, retaining deterministic rules engines where precision and compliance are required, and introducing replaceable enterprise-grade open-weight models. He cited a global airline fleet maintenance lifecycle project as an example, explaining that maintenance records are vectorized and stored in a private vector database within the customer's private cloud. High-risk decisions are adjudicated by domain experts, and the adjudication results flow back into the system.

Google Agents CLI Launches Extension System: From ADK to LangChain, All Can Plug In

Starting with version 1.5.0, Google Cloud's Agents CLI introduces an Extension System that allows developers to override or add commands without forking the tool; version 1.8.0 debuts the first LangChain extension, replacing ADK-bound commands with equivalent LangChain implementations. Extensions are declared via agents-cli-extension.yaml in a Git repository and can override built-in commands such as playground, deploy, and eval generate, or add new commands; enterprise platform teams can use this to enforce policies such as SBOM scanning in deployment pipelines.

Mastra Launches Mastra Observability Managed Platform

Mastra has released the Mastra Observability managed platform, which can connect directly to Mastra framework projects, automatically trace agent runs, and view the latency, tokens, and cost of model and tool calls. The platform supports online scorers to monitor correctness, team annotation of traces to build evaluation datasets, and running evaluations in CI via Gates and Verdicts to catch regressions. It offers 100k observability events per month free with 15-day retention; beyond that, it costs $10 per 100k events.

WMS AI Assistant: How to Implement a Conversational Interface for Querying Inventory in Natural Language

The Tencent Cloud Developer Community shared a production-ready WMS inventory conversational assistant solution: the LLM is only responsible for understanding intent and mapping parameters, while actual queries are completed by proprietary API tools, and the conversation is managed with a three-state state machine of idle, streaming, and awaiting-confirm. The solution uses NDJSON streaming parsing and retains a half-line buffer, and trims historical context to the most recent 10–15 turns to control token costs; dangerous write operations are rejected by default and undergo secondary backend authentication. When the model is unavailable, it degrades to keyword recognition, and responses exceeding 3 seconds prompt the user to switch to filters.

Deploying the RFD4-Proteina model on Databricks

NVIDIA and Databricks plan to integrate NVIDIA BioNeMo's open-source protein design model RFD4-Proteina into Databricks' Genesis Workbench, running on the now generally available AI Runtime serverless NVIDIA GPU compute. The model can simultaneously generate protein sequences and all-atom structures, and supports optional reward-guided search to steer candidate designs toward target properties such as thermostability and binding affinity. Genesis Workbench includes two built-in RFD4-Proteina workflows, Antibody Design and Vaccine Design, and can be extended through Genie Code and NVIDIA BioNeMo Agent Toolkit skills.

NVIDIA Releases Nemotron 3.5 Series of Open Models, Covering Embedding, Reranking, and Document Parsing

On October 21, 2026, NVIDIA expanded the Nemotron Retriever series, launching Nemotron 3.5 Embed (8B BF16, 1B BF16, 1B NVFP4), Nemotron 3.5 Rerank 8B BF16, and the 631-million-parameter Nemotron 3.5 Parse, aimed at multimodal retrieval and document processing for enterprise agent workflows. NVIDIA says that adding Rerank to the agent retrieval pipeline reduces the token cost per query by about 14% (with the 8B embedding model) and 13% (with the 1B model), respectively, while also improving retrieval accuracy. Parse scores 75.20 on ParseBench and processes 10.5 pages per second on a single H100; the open weights, data, and training recipes have been released on Hugging Face and can be deployed via NVIDIA NIM as well as Baseten, DeepInfra, and FriendliAI.

What Is Distributed AI Training? An Analysis of Data, Tensor, Pipeline, and Expert Parallelism

Distributed AI training enables models or batches that exceed the capacity of a single device to be trained by coordinating data, parameters, activations, gradients, and optimizer states across multiple accelerators. Its workflow can be broken down into five steps: partitioning the work according to the chosen parallelism strategy, performing local forward and backward computation, communicating the required tensors, synchronizing or pipelining updates, and checkpointing and recovery across nodes. The article points out that simply adding more devices without changing the communication and memory design is a common misconception; network bottlenecks and synchronization failures can waste most of the theoretical compute.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

English translation is being prepared.

Six Meitu Products Selected for the 2026 Q3 AI 100 List

Three Meitu products—Meitu Xiuxiu, BeautyCam, and Wink—were selected for the 2026 third-quarter "Flagship AI 100" list, while Meitu Design Studio, Kaipai, and WHEE were selected for the "Innovative AI 100" list. Both lists were released before October 21, with a total of six Meitu products making the rankings.

English translation is being prepared.

Fireworks Offers Free Training of Muse Glimmer 30B: Using SFT to Predict Kalshi Market Outcomes

Fireworks AI announced that Muse Glimmer 30B can be trained with SFT for free until November 30, and released a complete recipe for a prediction experiment using historical Kalshi market data. The fine-tuned Glimmer achieved 58.3% accuracy on unseen topics, 11.2 percentage points higher than Jev's 47.1%; in a historical backtest, a hypothetical $1,000 grew to $1,071 after accounting for simulated fees, while both base Glimmer and Jev lost money. Training used LoRA adapters and a K-way cross-entropy loss, ran through the training data only once, and separately used a calibration set to fit a temperature parameter for each question category.

Fireworks Free Weights Club: Training Muse Glimmer to Beat Jev on Predictions

Fireworks AI launches the Free Weights Club, offering free SFT training for Muse Glimmer 30B until November 30. The fine-tuned tuned Glimmer achieves 58.3% prediction accuracy on unseen topics, 11.2 percentage points higher than Jev's 47.1%; in a historical backtest, after deducting simulated trading fees, a hypothetical $1,000 grew to $1,071, while both Jev and base Glimmer posted losses. In a separate synthetic benchmark, Fireworks' System One serving stack delivered median server throughput 2-3x faster than standard BF16 chat completions.

From brain scans to images: How AI reconstructs what a person is seeing

A team led by Michal Irani at the Weizmann Institute of Science in Israel has developed an AI tool called the "Universal brain encoder," which can infer what a person is seeing from brain scans alone and generate fairly accurate corresponding images, and can also work in reverse to predict a person's brain activity based on what they are seeing. The tool was presented last month at the "Cognitive Computational Neuroscience" conference in New York; its goal is to help understand how the brain works, and it is expected to help locked-in syndrome patients communicate or reconstruct dream content.

Memory Tensor Open-Sources MemT 1.0 Series Memory-Native Models

Memory Tensor has open-sourced and released the MemT 1.0 series of memory-native models, including MemT-1.0-Nano with 36B parameters and 3B activated (based on Qwen3.6-35B-A3B) and MemT-1.0-Flash with 288B parameters and 17B activated (based on DeepSeek-V4-Flash), adding a memory computation pathway parallel to attention in selected Transformer modules. The team built a 34-billion-token memory training corpus, of which 20 billion tokens were used for continued pre-training and 14 billion for post-training. In a comprehensive evaluation across 11 memory benchmarks, Flash and Nano achieved average scores of 50.6 and 47.0, respectively, ranking in the top two among the parametric memory methods evaluated; on 500 SWE-bench Verified tasks, the proportion of tasks that Nano succeeded on all three attempts rose from 54.00% for the no-experience baseline to 62.80%.