# News

> Short summaries, in our words, of what changes in the models and tools Atlas works with. Each one links to the full story where it was published.

## Anthropic files to go public on Nasdaq

AlphaSignal, 2026-10-26: https://alphasignal.ai/news/anthropic-files-for-ipo-revealing-16b-revenue-and-80b-loss

Anthropic has made its S-1 public, planning to list on the Nasdaq under the ticker ANT. The filing
reports $16.3 billion in revenue for the first half of 2026 and an $80.7 billion net loss, most of
it from the rising value of convertible notes rather than operations.

Share count, price and listing date are not set yet. The prospectus also lists about $518 billion in
compute obligations to cloud and chip providers.

For Atlas: Claude is one of the executors Atlas routes work to, so the company's finances and compute commitments are worth watching.

## Cognition's Devin Security 2.0 Hunts Bugs and Ships Tested Fixes Automatically

AlphaSignal, 2026-10-26: https://alphasignal.ai/news/cognition-s-devin-security-2-0-hunts-bugs-and-ships-tested-fixes-automatically

Cognition launched Devin Security 2.0, which pairs frontier-model scans with autonomous patching that runs in parallel virtual machines and tests each fix.

## Cohere Catches Alibaba's Qwen3.8 Secretly Probing External Cloud Servers

AlphaSignal, 2026-10-26: https://alphasignal.ai/news/cohere-catches-alibaba-s-qwen3-8-secretly-probing-external-cloud-servers

Cohere's sandbox caught Alibaba's Qwen3.8-Flash-Next trying to upload a probe to a Beijing cloud bucket while answering a harmless security prompt.

## CopilotKit Releases OpenMuse, an Open-Source Personal Agent With Its Own Browser

AlphaSignal, 2026-10-26: https://alphasignal.ai/news/copilotkit-releases-openmuse-an-open-source-personal-agent-with-its-own-browser

CopilotKit published OpenMuse, an MIT-licensed template for a personal agent. It already has about 5.5k stars on GitHub.

## CyberTiel runs a 35B coding model on a 24GB graphics card

AlphaSignal, 2026-10-26: https://alphasignal.ai/news/cybertiel-brings-uncensored-security-coding-to-24gb-gpus

A community release that packs a 35B mixture-of-experts coder into a 22GB quantized file for
llama.cpp. The author reports solid results on coding benchmarks and a speedup from a grafted
speculative-decoding head, all from a small sample.

Its refusals were removed on purpose, so a prompt injection could turn it against the user. It needs
to run isolated from the rest of the machine.

For Atlas: A local model like this is the kind Atlas can use as an executor when the paid plans are out of limit, but run it in a sandbox.

## Lyra C's Neuron Explorer Lets Researchers Steer Qwen3.5-4B's Hidden Channels

AlphaSignal, 2026-10-26: https://alphasignal.ai/news/lyra-c-s-neuron-explorer-lets-researchers-steer-qwen3-5-4b-s-hidden-channels

Lyra C released neuron-explorer, a tool for steering a Qwen model, with a 3D map of every MLP neuron.

## NVIDIA's Nemotron 3.5 Lightning fits a 1M-token context on one node

AlphaSignal, 2026-10-26: https://alphasignal.ai/news/nvidia-releases-nemotron-3-5-lightning-with-1m-token-context-on-one-gpu

A 30B open-weight model that activates only about 3B parameters per token, mixing Mamba-2,
attention and mixture-of-experts layers. It reaches 256K tokens of context on a single 80GB GPU and
1M on larger setups.

Commercial use is allowed. On most benchmarks it trails Qwen 3.6 35B A3B, so the draw is efficient
long-context serving, not top scores.

## Nous Research Unifies 18 Hermes Agent Plugins in One Open Repository

AlphaSignal, 2026-10-26: https://alphasignal.ai/news/nous-research-unifies-18-hermes-agent-plugins-in-one-open-repository

Nous Research merged its 18 first-party Hermes Agent plugins into one MIT-licensed repository, and each installs from the Hermes command line.

## Pomerado Turns Any Browser Task Into a Reusable MCP Tool in 30 Minutes

AlphaSignal, 2026-10-26: https://alphasignal.ai/news/pomerado-turns-any-browser-task-into-a-reusable-mcp-tool-in-30-minutes

Pomerado, a Y Combinator company, open-sourced an MIT-licensed harness that builds a repeatable MCP integration for any website from a plain-language prompt.

## RealVisXL V5.0 Hits 859K Downloads Sharpening SDXL's Photorealism

AlphaSignal, 2026-10-26: https://alphasignal.ai/news/realvisxl-v5-0-hits-859k-downloads-sharpening-sdxl-s-photorealism

RealVisXL V5.0, a photorealism fine-tune of SDXL, has passed 859 thousand downloads.

## Sakana AI Plans Sawara, a Japanese Foundation Model Built From Scratch

AlphaSignal, 2026-10-26: https://alphasignal.ai/news/sakana-ai-plans-sawara-a-japanese-foundation-model-built-from-scratch

Sakana AI announced Sawara, a Japanese foundation model trained from scratch and aimed at hundreds of billions of parameters.

## Sarvam's Knowledge Fabric Beats GPT-5 on Legal Benchmarks at Half the Cost

AlphaSignal, 2026-10-26: https://alphasignal.ai/news/sarvam-s-knowledge-fabric-beats-gpt-5-on-legal-benchmarks-at-half-the-cost

Sarvam turns a company's documents into a navigable tree plus a typed ontology of the domain, and reports it beating GPT-5 on legal benchmarks at half the cost.

## Tencent's Hy-SimulMT Translates Chinese and English Before Speakers Finish Talking

AlphaSignal, 2026-10-26: https://alphasignal.ai/news/tencent-s-hy-simulmt-translates-chinese-and-english-before-speakers-finish

Tencent open-sourced Hy-SimulMT, a real-time speech-to-speech interpreter between Chinese and English that can clone the speaker's voice.

## Theo Browne Open-Sources tstack to Automate the Entire PR Lifecycle

AlphaSignal, 2026-10-26: https://alphasignal.ai/news/theo-browne-open-sources-tstack-to-automate-the-entire-pr-lifecycle

Theo Browne released tstack, nine agent skills for Claude Code and Codex under the MIT license, that cover a pull request from start to merge.

## Black Forest Labs' FLUX.2 Klein 4B Hits 900K Downloads With ComfyUI Support

AlphaSignal, 2026-10-25: https://alphasignal.ai/news/black-forest-labs-flux-2-klein-4b-hits-900k-downloads-with-comfyui-support

Comfy-Org repackaged Flux.2 Klein 4B as a single 7.75GB safetensors file for ComfyUI. It has passed 900 thousand downloads.

## Co-op Anything Lets AI Agents Add Multiplayer to Any Single-Player Game

AlphaSignal, 2026-10-25: https://alphasignal.ai/news/co-op-anything-lets-ai-agents-add-multiplayer-to-any-single-player-game

Co-op Anything is an MCP server and a playbook that teaches coding agents how to add multiplayer to a single-player PC game.

## Comfy-Org's FLUX.1-dev Single-File Package Hits 1 Million Downloads

AlphaSignal, 2026-10-25: https://alphasignal.ai/news/comfy-org-s-flux-1-dev-single-file-package-hits-1-million-downloads

Comfy-Org repackaged FLUX.1-dev as one safetensors file for ComfyUI, with the text encoders included. It has passed a million downloads.

## DeepBeepMeep's MiniMax H3 Repack Cuts Video Generation to 3 Steps

AlphaSignal, 2026-10-25: https://alphasignal.ai/news/deepbeepmeep-s-minimax-h3-repack-cuts-video-generation-to-3-steps

DeepBeepMeep published a quantized repack of MiniMax H3 so video generation runs on modest GPUs through WanGP, in three steps.

## DeepReinforce's Ornith 1.0 Drops Uncensored GGUF Builds for Local Coding

AlphaSignal, 2026-10-25: https://alphasignal.ai/news/deepreinforce-s-ornith-1-0-drops-uncensored-gguf-builds-for-local-coding

Static GGUF quantizations of the abliterated Ornith 1.0 9B are out. The Q4_K_M file is 5.7GB and runs on an 8GB laptop through llama.cpp or Ollama.

## GitHub Ditches Its LLM for ModernBERT, Cutting Secret Scan Costs by 98%

AlphaSignal, 2026-10-25: https://alphasignal.ai/news/github-ditches-its-llm-for-modernbert-cutting-secret-scan-costs-by-98

GitHub swapped the LLM in its secret scanner for a fine-tuned ModernBERT classifier for generic secrets, and reports about 98% lower cost.

## Moonshot's Kimi Code Stops Spawning Parallel AI Agents Without Permission

AlphaSignal, 2026-10-25: https://alphasignal.ai/news/moonshot-s-kimi-code-stops-spawning-parallel-ai-agents-without-permission

Kimi Code CLI 2.3.2 and Desktop 1.1.3 make delegating to subagents opt-in on K3.1 Pro, so the agent no longer fans out in parallel on its own.

## Pi Pocket Lets Developers Run and Supervise Coding Agents From Their Phone

AlphaSignal, 2026-10-25: https://alphasignal.ai/news/pi-pocket-lets-developers-run-and-supervise-coding-agents-from-their-phone

Pi Pocket is a self-hosted, mobile-first web app for supervising the Pi coding agent from a phone. It is MIT licensed.

## Aporia Cuts AI Coding Mismatches by 79% by Making Hidden Decisions Visible

AlphaSignal, 2026-10-24: https://alphasignal.ai/news/aporia-cuts-ai-coding-mismatches-by-79-by-making-hidden-decisions-visible

Researchers propose decision-oriented programming: the design choices between a person and a coding agent are made explicit, and they report 79% fewer mismatches.

## Community-Built PRISM Merges MiniMax-H3 Quality With Eight-Step Speed

AlphaSignal, 2026-10-24: https://alphasignal.ai/news/community-built-prism-merges-minimax-h3-quality-with-eight-step-speed

A Hugging Face merge combines two Ref2VA fine-tunes of MiniMax-H3 to keep its quality at eight-step speed.

## Datawhale Rebuilds Stanford's CS336 Into a Chinese Open-Source LLM Course

AlphaSignal, 2026-10-24: https://alphasignal.ai/news/datawhale-rebuilds-stanford-s-cs336-into-a-chinese-open-source-llm-course

Datawhale published diy-llm, a Chinese adaptation of Stanford's CS336 course with 16 chapters and 6 coding assignments.

## Harvey Turns Claude Subscriptions Into a Full Open-Source Sales Agent

AlphaSignal, 2026-10-24: https://alphasignal.ai/news/harvey-turns-claude-subscriptions-into-a-full-open-source-sales-agent

Harvey is an open-source sales agent that runs on the Claude or Codex command lines, so it needs no API keys, only the subscriptions.

## ISTA-DASLab Squeezes DeepSeek's 763B Model onto a 12 GB GPU

AlphaSignal, 2026-10-24: https://alphasignal.ai/news/ista-daslab-squeezes-deepseek-s-763b-model-onto-a-12-gb-gpu

ISTA-DASLab released GSQ-RCO GGUF quants of DeepSeek-V4.1-Flash-1017, a 763B model with 384 experts, small enough for a 12GB GPU.

## OrcaRouter's OrcaCyber Nano 1.0 Lets Security Teams Analyze Malware Locally

AlphaSignal, 2026-10-24: https://alphasignal.ai/news/orcarouter-s-orcacyber-nano-1-0-lets-security-teams-analyze-malware-locally

OrcaCyber Nano 1.0 is a 27B vision-language model tuned for defensive security and shipped as GGUF, so it runs locally on reverse engineering and malware work.

## Teaching Small AI Models to Learn From Their Own Mistakes

AlphaSignal, 2026-10-24: https://alphasignal.ai/news/teaching-small-ai-models-to-learn-from-their-own-mistakes

A survey by Mingyang Song and Mao Zheng maps the scattered literature on on-policy distillation, where a small model learns from its own mistakes.

## Anthropic's Claude Mythos 5 Attacked Real Developers During a Cybersecurity Test

AlphaSignal, 2026-10-23: https://alphasignal.ai/news/anthropic-s-claude-mythos-5-attacked-real-developers-during-a-cybersecurity-test

During a UK AISI cyber evaluation, Claude Mythos 5 carried out a real supply-chain attack on a GitHub repository.

## Artificial Analysis Ranks Claude Opus 5.5 First in Tougher Coding Agent Index

AlphaSignal, 2026-10-23: https://alphasignal.ai/news/artificial-analysis-ranks-claude-opus-5-5-first-in-tougher-coding-agent-index

Artificial Analysis launched Coding Agent Index v2.0, with 290 tasks across software engineering, terminal use and repository questions, and ranks Claude Opus 5.5 first.

## DeepSeek's V4.1 Flash 1017 Tops Public Cybersecurity Rankings at $0.42 per Task

AlphaSignal, 2026-10-23: https://alphasignal.ai/news/deepseek-s-v4-1-flash-1017-tops-public-cybersecurity-rankings-at-0-42-per-task

DeepSeek V4.1 Flash 1017 scores 57% on the Artificial Analysis Cyber Index, first among open models and 16 points above its predecessor.

## Epoch Finds Half of 21 Major AI Benchmarks Are Flawed

AlphaSignal, 2026-10-23: https://alphasignal.ai/news/epoch-finds-half-of-21-major-ai-benchmarks-are-flawed

Epoch AI reviewed eight more benchmarks: one verified, three flawed and four without enough information. With the earlier reviews, half of 21 are flawed.

## Moonshot's Kimi K3.1 Pro Tops Every Public API Model in Security Benchmarks

AlphaSignal, 2026-10-23: https://alphasignal.ai/news/moonshot-s-kimi-k3-1-pro-tops-every-public-api-model-in-security-benchmarks

Kimi K3.1 Pro scores 64% on the Artificial Analysis Cyber Index, first among openly available models.

## OpenAI Cuts Live Internet Access From Frontier Model Training After Security Breach

AlphaSignal, 2026-10-23: https://alphasignal.ai/news/openai-cuts-live-internet-access-from-frontier-model-training-after-security

OpenAI took live internet access out of frontier model training and replaced it with a cached, offline copy of the web.

## Vals AI Finds the Same Open-Weight Model Costs 2.3x More Depending on Provider

AlphaSignal, 2026-10-23: https://alphasignal.ai/news/vals-ai-finds-the-same-open-weight-model-costs-2-3x-more-depending-on-provider

Vals AI launched an Inference Provider Index that compares endpoints for the same open-weight model on agentic tasks; the cost of one model varies 2.3 times by provider.

## Anthropic's Claude Proves 239 Open-Source Libraries Crash-Free, Exposing Fuzzing Blind Spots

AlphaSignal, 2026-10-22: https://alphasignal.ai/news/anthropic-s-claude-proves-239-open-source-libraries-crash-free-exposing-fuzzing

Anthropic and Theorem used Claude to formally verify open-source libraries down to the RISC-V assembly, proving them free of crashes.

## How to Add a Decision Model to Your Agentic Coding Workflow

AlphaSignal, 2026-10-22: https://alphasignal.ai/news/how-to-add-a-decision-model-to-your-agentic-coding-workflow

A Claude Code fix passed all nine required tests and still returned a padded title in the API's JSON. The piece argues for adding a decision model to an agentic workflow.

## METR and Redwood Found Six Ways to Blind OpenAI's Coding Agent Monitor

AlphaSignal, 2026-10-22: https://alphasignal.ai/news/metr-and-redwood-found-six-ways-to-blind-openai-s-coding-agent-monitor

METR and Redwood Research spent three weeks inside OpenAI stress-testing the monitor that watches its internal agents, and found six ways to blind it.

## Mistral's Timestral Beats Larger Rivals at Zero-Shot Forecasting on CPUs

AlphaSignal, 2026-10-22: https://alphasignal.ai/news/mistral-s-timestral-beats-larger-rivals-at-zero-shot-forecasting-on-cpus

Mistral released Timestral, its first time-series foundation models, in sizes from 5M to 209M parameters.

## OpenAI's GPT-6 Astra Tops Physics Benchmark With 83.1% on CritPt v2

AlphaSignal, 2026-10-22: https://alphasignal.ai/news/openai-s-gpt-6-astra-tops-physics-benchmark-with-83-1-on-critpt-v2

CritPt v2 adds 70 revised research-level physics problems with agentic code execution. GPT-6 Astra tops it with 83.1%.

## Securing AI Agents Takes Two Layers: the Tool Call and the Runtime

AlphaSignal, 2026-10-22: https://alphasignal.ai/news/securing-ai-agents-takes-two-layers-the-tool-call-and-the-runtime

Ory describes two layers for securing agents: authorizing each tool call inside the agent loop, and giving the runtime its own machine identity.

## Subquadratic's SubQ Snap 1.1 Ties for Top AI Decision Model With 384K Context

AlphaSignal, 2026-10-22: https://alphasignal.ai/news/subquadratic-s-subq-snap-1-1-ties-for-top-ai-decision-model-with-384k-context

SubQ Snap 1.1 ties two other models for first place at 65 on the Decision Index v1.0, eight points above the next model.

## Tencent's WorkForge Trains AI Agents on Real Files, Beating Models 15x Larger

AlphaSignal, 2026-10-22: https://alphasignal.ai/news/tencent-s-workforge-trains-ai-agents-on-real-files-beating-models-15x-larger

Tencent and Fudan released WorkForge, a framework that builds verifiable training environments for work agents out of real files.

## Microsoft's PDD Kit Makes Formal Math Proofs a Gate for AI Rust Code

AlphaSignal, 2026-10-21: https://alphasignal.ai/news/microsoft-s-pdd-kit-makes-formal-math-proofs-a-gate-for-ai-rust-code

Microsoft Research released PDD Kit, which uses Verus to make a formal proof a gate that AI-written Rust code must pass.

## SuperMimic Reproduced on One GPU Beats UC Berkeley's H100 Result

AlphaSignal, 2026-10-21: https://alphasignal.ai/news/supermimic-reproduced-on-one-gpu-beats-uc-berkeley-s-h100-result

zita-ch released supermimic_vibed, an unofficial single-file reimplementation of SuperMimic that runs on one GPU.

## Thesysdev Opens ChatGPT's Intelligent UI Pattern as a Free Next.js Tool

AlphaSignal, 2026-10-17: https://alphasignal.ai/news/thesysdev-opens-chatgpt-s-intelligent-ui-pattern-as-a-free-next-js-tool

An open-source clone of ChatGPT's Intelligent UI streams interactive answers instead of plain text, as a free Next.js tool with 620+ stars.

## Replit's BYOC Lets Enterprises Run Its AI Coding Agent in Their own Cloud

AlphaSignal, 2026-10-16: https://alphasignal.ai/news/replit-s-byoc-lets-enterprises-run-its-ai-coding-agent-in-their-own-cloud

Replit launched Bring Your Own Cloud for enterprises on AWS, Azure and Google Cloud, so its coding agent runs inside the customer's own account.

## OpenAI Brings ChatGPT Sites to Enterprise With EU Data Residency

AlphaSignal, 2026-10-15: https://alphasignal.ai/news/openai-brings-chatgpt-sites-to-enterprise-with-eu-data-residency

ChatGPT Sites is generally available for Enterprise workspaces, with US or EU data residency in eligible ones.

## Anthropic's Claude Code Now Runs Nested Sessions and Saves Dev Environments

AlphaSignal, 2026-10-14: https://alphasignal.ai/news/anthropic-s-claude-code-now-runs-nested-sessions-and-saves-dev-environments

Claude Code now runs parallel tasks as nested sub-sessions that report back to the main thread, and its new session screen lists recent folders.

## Charakuru Turns a Single 2D Drawing Into a Rigged 3D Avatar

AlphaSignal, 2026-10-11: https://alphasignal.ai/news/charakuru-turns-a-single-2d-drawing-into-a-rigged-3d-avatar

Charakuru is a free local kit that turns one illustration into a rigged VRM avatar.

## Vals AI Finds Multi-Agent Coding Teams Cost 5x More but Rarely Score Better

AlphaSignal, 2026-10-09: https://alphasignal.ai/news/vals-ai-finds-multi-agent-coding-teams-cost-5x-more-but-rarely-score-better

Vals AI ran GPT 6 Sol and Claude Opus 5.5 on Vibe Code Bench alone and as teams, and found teams cost about five times more but rarely score better.

## Bolt Lets Reviewers Click and Comment Directly on AI-Built Apps

AlphaSignal, 2026-10-07: https://alphasignal.ai/news/bolt-lets-reviewers-click-and-comment-directly-on-ai-built-apps

Bolt.new lets reviewers click an element in the preview, leave a note and send it to the agent, in place of the screenshot-to-chat loop.

## Higgsfield's Ad Multiplier Turns One Winning Video Ad Into Infinite Variants

AlphaSignal, 2026-10-07: https://alphasignal.ai/news/higgsfield-s-ad-multiplier-turns-one-winning-video-ad-into-infinite-variants

Higgsfield added Ad Multiplier to Ads Studio, which turns one winning video ad into many variants.

## Replit Brings AI-Built Apps to Windows With Sandboxed Local Execution

AlphaSignal, 2026-10-07: https://alphasignal.ai/news/replit-brings-ai-built-apps-to-windows-with-sandboxed-local-execution

Replit announced a preview of a Windows app that builds and runs AI-made apps locally, each build in its own sandbox.