MiniMax Launches M3.1 Flash, Meta Details Prometheus, and DHS Publishes DeepSeek Removal Order
MiniMax launched M3.1 Flash with new figures for inference speed, agent task costs, and API pricing. OpenAI released MultiLangBench, covering 20 programming languages. Meta detailed training operations at its Prometheus supercluster and an architecture that coordinates 14 specialist agents for data-pipeline work.
This edition also covers Mercury 3, Ming-3.0-tts, FLUX 3 Max, Trainium development tools, and Claude's workplace features. n8n documents an offline invoice-processing workflow, the U.S. Federal Register publishes a DeepSeek exclusion and removal order, and NorForge reports website attacks conducted through Cairn agents.
01
Model Releases & Multimodal Generation
5 stories
2026-10-19MiniMax
MiniMax Launches M3.1 Flash, Reporting 70% Lower Agent Task Costs Than Its Predecessor
MiniMax released M3.1 Flash, whose early version was tested under the Space Bunny name. The company reports online inference nearly three times as fast as its predecessor, with peak output above 200 tokens per second and a median above 170. It says overall agent task costs are 70% lower. API prices per million tokens are $0.15 for input, $0.60 for output, and $0.015 for cached input.
MiniMax's published results include 73.7% on DeepSWE v1.1, which tests code understanding, feature development, bug fixes, and testing, and 61.7 points on OSWorld V2 for computer operation and cross-application tasks. The model is available through the MiniMax API, M Plan, and MiniMax Code, offering API and desktop access. Model weights are planned for a later release.
2026-10-19OpenAI
OpenAI Releases MultiLangBench and Publishes a COBOL Code-Migration Evaluation
OpenAI open-sourced MultiLangBench, which asks models to build a SQLite-compatible database engine in each of 20 languages. Shared database files and SQL queries let the evaluation check results across implementations. At maximum reasoning effort, GPT-6 Astra averaged an 87.4% test pass rate, compared with GPT-5.6 Sol's 26.8%. COBOL improved from 10.9% to 72.1%.
A separate evaluation migrates COBOL programs to C++, Java, Kotlin, Python, or Rust. Success requires passing every hidden test with byte-for-byte matching outputs. GPT-6.1 Sol reached 86.6% at low reasoning effort, versus 5.4% for GPT-5.6 Sol at the same setting, and 97.9% at higher effort. One behavior under examination is fixed-point arithmetic: COBOL normally truncates excess decimal digits, so a rewrite that rounds instead can run successfully while returning different results.
2026-10-19Inception / OpenRouter
Inception Releases Mercury 3 With a Million-Token Context Window and Reported Output of 1,068 Tokens per Second
Inception released Mercury 3, a diffusion language model that generates and refines tokens in parallel. The company reports 1,068 output tokens per second on NVIDIA GPUs, with context expanded from Mercury 2.5's 260,000 tokens to one million. Standard prices are $0.20 per million input tokens and $0.70 per million output tokens, discounted at launch to $0.10 and $0.35.
Mercury 3 is available through Inception's API, OpenRouter, Baseten, and Vercel, with OpenAI API compatibility. In a test cited in the announcement, TinyFish used its production web-agent framework to run 50 live Mind2Web navigation tasks; Mercury 3 completed 50%.
2026-10-19Ant Group inclusionAI
inclusionAI Open-Sources Ming-3.0-tts With Seven Languages, Voice Cloning, and Multi-Speaker Dialogue
Ant Group's inclusionAI released Ming-3.0-tts. The code uses the MIT license, and weights are available on Hugging Face and ModelScope. The family includes a 0.5B dense model and a tiny variant with 7.9B total parameters and 1.3B active per step. It supports Chinese, English, German, Spanish, French, Korean, and Japanese, plus Sichuanese and Cantonese. Users can clone a reference voice, describe a voice in text, or choose a built-in voice.
Natural-language controls adjust speed, pitch, and emotion, while 14 tags add expressions such as laughter and sighs in Chinese or English. The tiny variant supports scripted multi-speaker dialogue; controls use separate task modes. The project's October leaderboard snapshot lists a first-place pronunciation robustness score of 93.2%. That evaluation uses human judgments of highlighted English passages covering numbers, shorthand, identifiers, and ambiguous words.
2026-10-19fal / Black Forest Labs
fal Offers FLUX 3 Max With End-Frame Guidance, Audio, and Promotional 720p Pricing of $0.0425 per Second
FLUX 3 Max, developed by fal with Black Forest Labs, offers text-to-video and image-to-video endpoints. fal's product page reports about 12.45 seconds to generate a 10-second 720p clip in its test. Promotional 720p pricing is $0.0425 per output second, or $0.425 for a 10-second clip, through November 7. The subsequent rate is $0.17 per second.
Image-to-video uses a starting image to establish the scene and a prompt to direct subjects, camera movement, and environmental changes. An optional ending image guides the final frame. The current endpoint supports 5–20-second clips, optional generated audio, 720p output, and optional 1080p upscaling. fal also published 50 examples with their original prompts, showing how starting images combine with motion, lighting, and sound descriptions.
02
Compute Clusters, Training Backends & Production Agent Engineering
5 stories
2026-10-19Meta Engineering
Meta Details Prometheus: Phase One Trained Muse, With One Million GPUs a Future Expansion Target
Meta detailed its gigawatt-scale Prometheus supercluster in New Albany, Ohio: its first phase is complete and has trained the Muse family. Hundreds of thousands of GPUs across multiple buildings are connected by fiber, with a future target of roughly one million GPUs. The cluster uses GB200 and GB300 platforms. Meta reports training goodput above 85% for Muse Wonder pretraining and above 90% for reinforcement learning.
To reduce disruption from hardware failures, Meta assigns 64 GPUs in each 72-GPU domain to training and keeps eight as spares. A spare replaces a failed GPU, allowing the job to restart from its latest checkpoint. When a failure can be contained within one data-parallel replica, that replica can go offline while the rest continue training. Expansion plans include Vera Rubin hardware using the existing network.
2026-10-19Meta Engineering
Meta's Thin Orchestrator Coordinates 14 Agents and Cuts Data-Column Configuration Prep to Under an Hour
Meta described a Thin Orchestrator for recommendation-system data pipelines: roughly 1,300 lines of code route work, collect required information, and run checks, while 14 specialist agents handle configuration, code generation, and validation. The work spans three repositories, three languages, and up to six pipeline stages. An earlier single-agent approach succeeded only 10%–20% of the time without manually tailored prompts for each task.
The new system follows fixed workflows, with each agent loading only its own domain knowledge. Adding a data column proceeds through schema definition, validation, population, and testing; pipeline migration reuses the existing agents in a different sequence. Meta says configuration preparation fell from three or four days manually to under an hour, and about two-thirds of the system's changes land without substantial human edits.
2026-10-19PyTorch / AWS
AWS Announces TorchNeuron to Bring Trainium Into Native PyTorch Workflows
AWS announced TorchNeuron, a native PyTorch backend for Trainium2 and Trainium3, scheduled for public beta under Apache 2.0 on October 20. Developers can select
neuronas the model's device. Code written against PyTorch's device-agnostic interface can run unchanged, while compiled execution requires selecting the Neuron backend.TorchNeuron provides eager execution for step-by-step debugging and a compiled mode using
torch.compilefor performance optimization. Developers needing custom low-level operations can use the Neuron Kernel Interface. It also connects PyTorch's profiler to Neuron Explorer, linking hardware execution events to source locations so engineers can identify time-consuming steps. The announcement lists TorchTitan and Hugging Face Transformers v5 among supported projects.2026-10-08Anthropic
Claude Dashboards and Motion Offer Live Data Views and Editable Animation in Plan-Specific Betas
Anthropic introduced Claude Dashboards and Motion in its October 8 announcement. Dashboards is in beta on paid plans, connecting to sources such as BigQuery, Databricks, Snowflake, and Salesforce to create dashboards from natural-language requests and update them as data changes. Clicking a number reveals its query, and each chart shows its latest refresh time. Users can continue analysis in tools including Grafana, Hex, and Mixpanel.
Motion is in beta on Team and Enterprise, turning reports or charts into animated explainers exported as MP4 files. Its code-based animations allow edits to text, numbers, images, and timing. Enterprise administrators must enable both features. The same announcement moved Docs, Slides, and Design out of beta across all plans, including Free, and set December 14 as the standalone Design site's closing date.
2026-10-19n8n
n8n Demonstrates Offline Invoice Processing With Seven Models on One DGX Spark
n8n's Jamie Madden demonstrated seven commercially usable open-weight models on one NVIDIA DGX Spark with 128GB of unified memory. Qwen3.6 handles chat, coding, vision, and extraction; the other models provide embeddings, reranking, speech recognition, OCR, image generation, and speech synthesis. LiteLLM exposes them through one OpenAI-compatible endpoint, letting n8n workflows use local models throughout.
The guide's invoice workflow accepts an image, reads it with GLM-OCR, then asks Qwen3.6 to extract supplier, invoice number, date, total, and currency into structured form results. On the author's clean, single-page sample, OCR took 1.8 seconds; with reasoning disabled, extraction returned all five fields correctly in 0.59 seconds. The OCR output also dropped a zero from the sample's tax identifier. An importable workflow template accompanies the guide.
03
Supply-Chain Mandates & Autonomous Agent Security
2 stories
2026-10-20U.S. Department of Homeland Security / Federal Register
Federal Register Publishes DeepSeek Exclusion and Removal Order Covering Civilian Agencies and Related Third-Party Products
The October 20 Federal Register notice publishes DHS Exclusion and Removal Order 2026-01; the attached order lists July 24 as its original release date. It directs federal civilian executive agencies to stop procuring DeepSeek and affiliate products and remove covered products from their information systems. Its scope includes third-party products and services developed using DeepSeek implementations, excluding national security and sensitive compartmented information systems.
The order requires procurement exclusion and removal immediately after issuance. Agencies may submit written exception requests within 60 days of issuance, for CISA to review and decide in writing. Requests must identify the products, justification, and mitigating measures. Until removal is complete, DHS encourages traffic monitoring, data-exfiltration logging, and anomaly detection. The filing cites a security audit of DeepSeek's mobile application and its data practices as grounds for concern.
2026-10-19NorForge / The Verge
NorForge Documents Cairn Agent Attacks, With Reporting Identifying Websites of More Than 100 Companies
Security firm NorForge found automated website attacks through more than 20 publicly exposed Cairn control panels it assessed as likely malicious. The most active cluster recorded over 1,100 attack projects. Researchers estimate an average success rate of about 10% across observed activity. Among successful tasks in that cluster, the average time from start to goal was just under 23 hours, with agents recording findings and subsequent actions along the way.
The Verge, citing Reuters, reports affected websites belonging to more than 100 companies, including John Deere, SharkNinja, and Cummins. John Deere said there was no indication sensitive information was accessed. SharkNinja said the affected site was a standalone product-support portal and that the threat was contained. Cummins told The Verge it was investigating a potential attack on a third-party vendor's website.