Skip to content
DigIntoAI/Model intelligence
Model intelligence feed

Model intelligence.

Follow the changes. Understand what comes next.

Editor’s pickOther providers
Other providers/Steam

Valve's Answer to AI Slop and Clone Games on Steam: We Won't Police Developers' Tools

At Steam Indie Connect in Korea, Valve staff addressed AI slop and clone games head-on: no gatekeeping on AI tools, and the best defense against clones is community and steady updates.

Latest stories81

Newest published
37

Same Model, Different Coding Harness: Pass Rate Barely Moves, the Bill Does

A Shandong University paper (arXiv:2610.04433) runs the same models through Claude Code, mini-SWE-agent, and OpenCode on SWE-bench Verified. Pass rate barely moves beyond rerun noise; cost per task can differ by about 3×, mostly from each harness’s preamble and how many steps a run takes. OpenCode’s larger pool gaps often trace to output caps without recovery—not a claim that OpenCode always loses. We did not re-run the study; for reference only.

coding harnessRead story
38

Sweden's Government Offices Block ChatGPT and Claude Sites on Staff PCs—Not a Nationwide Ban

Sweden's Government Offices (Regeringskansliet) will block the ChatGPT and Claude websites on staff PCs from Monday, October 26, 2026, according to Sveriges Radio Ekot and SVT. The move covers browser access on office computers—not a nationwide ban, and not a removal of Microsoft Copilot. Staff can still use ChatGPT and Claude models inside the offices' Intric AI assistant platform. DeepSeek was restricted earlier at the same agency.

ChatGPT/ClaudeRead story
39

Same Open Model, Different API: Why Speed and Cost Both Change

Vals AI’s Inference Provider Index (page Updated 10/20/2026; Updates dated 10/22/2026) runs the same open-weight models—DeepSeek V4.1 Flash, Kimi K3, and MiniMax M3—on first-party APIs and hosts such as Together AI, Fireworks, and Baseten. Across 165 agent tasks, overall accuracy stayed within error of the first-party API. Cost per task rose by as much as about 2.3× by host, and for all three models the creator’s API was neither the cheapest nor the fastest. Figures below are from Vals AI, for reference only.

Inference Provider IndexRead story
40

Aokigahara Infostealer Targets Claude, Codex, Cursor and Copilot Credentials

On October 23, 2026, Check Point Research published an analysis of Aokigahara, a newly observed infostealer (earlier versions may have been known as Kamidere or Stygian). Beyond browsers, wallets and other classic targets, its aisessions handler collects authentication files, session data and configs tied to Claude, ChatGPT Desktop, Codex, Cursor, Windsurf, GitHub Copilot, Gemini CLI and similar tools. An ide handler also reaches Git, SSH and cloud credentials. Stolen AI credentials could enable misuse or resale; the analysis does not confirm how operators monetize the data. We have not re

Aokigahara / Kamidere / Stygian (infostealer)Read story
41

Chroma Joins Replit: Cloud Writes Stop Nov 22, Reads Dec 22

Chroma announced on October 23, 2026 that it signed a definitive agreement to join Replit and will wind down the existing Chroma Cloud service over about 60 days. Writes are disabled November 22, 2026 (read-only for 30 days); reads stop December 22, 2026. Open-source ChromaDB remains under Apache 2.0, with more private-repo code committed and Replit set to keep contributing. The wind-down applies only to hosted Chroma Cloud—not a shutdown of the open-source project.

Chroma Cloud / ChromaDBRead story
42

IQuest Opens Q1: 320B/15B MoE for Claude Code and Codex

IQuest Research has released IQuest-Q1, an open sparse Mixture-of-Experts model with about 320B total parameters and about 15B activated per token, aimed at agentic CLI and coding work. Context length is 524,288 tokens (512K). The website and Hugging Face card show SGLang and vLLM serving paths, plus examples for Claude Code and Codex CLI. Benchmark numbers are the project’s reported scores—not our measurements—and the card calls out text-only input and early-stage limits.

IQuest-Q1Read story
43

Hi3D 1.0 Is Open Source: Image-to-3D, Formerly Sparc3D

Hi3D has open-sourced Hi3D 1.0—formerly Sparc3D—a high-resolution geometry refinement model that turns a reference image plus a coarse mesh into denser 3D geometry. Inference code and weights are on GitHub and Hugging Face under the Hi3D Open License Agreement v1.0, not Apache. Commercial products that cross 200,000 monthly active users or USD 1,000,000 in monthly revenue need a separate written license, and commercial integrations must show “Powered by Hi3D 1.0.”

Hi3D 1.0 / Sparc3DRead story
44

General Reasoning Opens Avalon-M1 Preview and RLCR Training Recipe

London lab General Reasoning has published reinforcement learning from collective rewards (RLCR)—a recipe that jointly trains a CEO and specialists on one shared collective advantage—and a preview of Avalon-M1-Mini, still in training. The Mini build is about 120B parameters with 12B active, post-trained from the Nemotron 3 Super base. Model weights, harness, and environments are not released yet; the official roadmap lists those for later. Safety evaluations on the official page and technical report say refusal of harmful actions stays near-universal while reporting to human overseers declines

Avalon-M1 / Avalon-M1-Mini / RLCRRead story
45

Atria Dawn Formal Weights Are Out: 512K Context on Hugging Face

Shanghai Artificial Intelligence Laboratory has published formal open weights for Atria Dawn on Hugging Face (internlm/Atria-Dawn), with an FP8 sibling. The formal release lists a 512K context window, versus 256K for the earlier Preview. The model card describes a research- and engineering-oriented agent built on the GLM-5.2 MoE foundation—it is not a new Zhipu GLM generation under another name. Weights are text-only and very large on disk; that file size is not a minimum VRAM figure.

Atria Dawn / Atria-Dawn-FP8Read story
46

Replit Monitoring Agent Can Investigate, Fix and Auto-Publish Production Issues

Replit introduced Monitoring Agent—an autonomous on-call for published apps. It can read production logs and metrics, investigate an alert, make and validate a fix, and email a summary. With auto-publish on, it can deploy a validated fix to production. Earlier App Monitoring covered health checks and downtime alerts; this release adds the fix-and-ship step. Enable monitoring alerts and auto-fix in Publishing settings; auto-publish is optional. Separate from Replit Agent’s default model choice.

Monitoring AgentRead story
47

Block Releases Dessa, a Full-Duplex Speech Model Trained Only on Speech

Block’s small AI research lab shipped Dessa: speech in, speech out, full-duplex. It says the model was trained from scratch on human conversation audio only—no text, no pretrained base. Research preview for now: a small group of AI safety researchers get early access; everyone else joins a waitlist on the official page. Block says inference beats real time and the model runs locally on a MacBook Pro M4 Max or better. Training, per the page, used one NVIDIA GB200 NVL72.

DessaRead story
48

Simular Agent S4 Evolves Interfaces Without Retraining the Model

On October 21, 2026, Simular introduced Agent S4, which evolves the interface repository computer-use agents work through while the task agent’s model and harness stay fixed. With Qwen3.8-27B and an evolved repository (no scratch VM), it reaches 21.3% binary success on the full 108-task OSWorld V2 set—about 3.8× the same model with fixed interfaces (5.56%), and above GPT-5.5 (13.0%) and Claude Opus 4.8 (20.6%) at a small fraction of their reported task cost. Evolving an interface repository is not the same as Cursor’s self-hosted desktop computer use.

Agent S4Read story