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
25

DiffuSpace Opens Clover-35B: Open-Weight MoE Diffusion LLM Around 35B

DiffuSpace (October 28, 2026) open-sources Clover-35B-A3B-Preview—an MoE diffusion LLM (~35B total / ~3B active), thinking mode, 128K context, Apache-2.0. Weights on Hugging Face, ModelScope, and GitHub; SGLang and llama.cpp supported. Lineage: HKU NLP Dream; HF base_model Qwen3.6-35B-A3B. Vendor-reported agent scores include SWE-bench Verified 73.6, Terminal-Bench 2.1 51.7, τ³ Overall 65.3; ~680 tok/s on one GB300 BF16 agent workload—for reference only. A decision model comes later. No OpenRouter clover/diffuspace ID yet; try-out is apply/chat, not an ungated public API.

Clover-35B/Clover-35B-A3B-PreviewRead story
26

Opera WebAI Early Preview: What Local Inference Saves, Which Devices Can't Run It

Opera (27 Oct 2026) opens WebAI early preview in Opera One, GX, and Neon: sites and extensions call local Prompt, Decision, and Semantic Embeddings APIs in-browser. No separate AI app, local server, or API key; page content used by local AI need not go to a cloud provider. Opt-in: Settings → Opera AI → Enable Web AI API. Supported: Win64 Intel/AMD and Apple silicon Macs; not Intel Mac, Linux, or Windows on Arm. Six demo extensions are samples, not built-in. Proposed web standard—not a finished universal API. Speed and CJK quality not tested here. For reference only.

Opera One/Opera GX/Opera Neon/WebAIRead story
27

Liquid AI Releases LFM3: Open-Weight Multimodal Family Around 2B

Liquid AI (October 27, 2026) announces LFM3, an open-weight multimodal family around 2B on one shared backbone—text, vision, audio, omni, and decision (d1). Weights are on Hugging Face, Liquid Playground, and the Liquid AI API. LFM3-1.7B is about 1.73B parameters with 256K context; Liquid reports a tri-hybrid architecture and about 22T pre-training tokens. OpenRouter is a named day-0 partner, but public OpenRouter LFM3 model IDs were not listed yet—today's entry points are HF / Playground / Liquid API. Blog benchmarks are vendor-reported. For reference only.

LFM3-1.7B/LFM3-VL/LFM3-Audio/LFM3-Omni/d1Read story
28

Artificial Analysis Intelligence Index Moves to v5 — Private Benchmarks Now Dominate

On October 28, 2026 at 10:11 Asia/Shanghai, Artificial Analysis moved Intelligence Index to v5. About 60% of the score rests on private test sets and held-out answers. Nine evaluations remain; Coding is 25%—FrontierCode 2.0 (15%) plus Terminal-Bench 4.0 (10%). On max settings, Claude Opus 5.5 leads at 58; GPT-6 Astra is 57 and GPT-6.1 Sol is 55. FrontierCode 2.0 covers 132 real open-source repo tasks scored by merge readiness. Progress Index v1 is separate; frontier rankings still use Intelligence Index. Not re-run on this site; scores follow Artificial Analysis. For reference only.

Intelligence Index v5 / FrontierCode 2.0(旁 Progress Index v1)Read story
29

Claude Code Is Multiplayer Now: Share a Session with Delta (Beta)

On October 27, 2026, Delta’s blog (Franciska Dethlefsen) said Claude Code is now multiplayer in beta via Delta CLI and an open-source Claude Code plugin. /delta share yields a link; teammates use /delta join in their own Claude Code session. The plugin shares the conversation and unfinished code changes / worktree; the CLI syncs files across machines. Each person keeps their own Claude session, with the shared conversation as context. More harness support is coming, starting with Pi. This is a Delta / Zed-side capability, not an Anthropic official Claude Code update. For reference only.

Claude Code/DeltaRead story
30

What Is Runway App Studio? How It Differs from Apps and Workflows, Plus Credits and Sharing Limits

On October 23, 2026, Runway’s Paid Plans changelog listed App Studio: describe a creative app; it builds with Runway’s models, brand, and files, then hosts it. Help: ~700 credits balance to start a build session; ~50 per block while building—a start-balance threshold, not a fixed price per app. Classic Apps via Switch to classic Apps; Workflows stay node pipelines. Sharing docs conflict—Who can open and the FAQ deny an outside public link, while Publishing lists Anyone with a link—so outside sharing stays pending verification. Credit burn and time savings have not been measured on this site.

App StudioRead story
31

Why AGENTS.md in Downloads Is Dangerous: Working-Directory Pitfalls for Codex and Claude Code

Forever Security (22 Oct 2026) found malicious AGENTS.md in a real customer's ~/Downloads; delivery undetermined. Researchers reproduced Codex rewriting ~/.codex/config.toml and pulling function-color (self-test >90%), plus a Claude Code variant writing ANTHROPIC_BASE_URL into .claude/settings.json (~50%). BleepingComputer (25 Oct) verified a PoC can land AGENTS.md in Downloads in Chrome/Edge/Firefox without save-as; Brave asks first—yet the chain still needs the agent started there under approval and sandbox. Working-directory checklist; not a claim every visit or version is compromised.

Codex / Claude CodeRead story
32

Entire Trails vs Pull Requests: What AI Review Scores, Tests, and Human Approval Each Control

On October 26, 2026, Entire launched an independent Git forge for agent–human collaboration. Trails put requirements, session traces, code changes, and scored agent evals in one HUD; every push runs evals and findings. Gates can block merge until findings clear, CI passes, and a human approves. CI: Buildkite and Depot. Private repos today in US, EU, India, Australia; public repos and self-hosted Git nodes on the roadmap. Vendor speed and token-savings claims are not measurements run on this site; review scores are not a security guarantee; pricing unverified—no purchase or migration advice.

TrailsRead story
33

Code2Video-AA: How Models Design Videos in Code — and What It Costs

On October 26, 2026, Artificial Analysis published Code2Video-AA—its independent run of HeyGen’s Code2Video Bench. Twenty-seven models each got 114 short product/brand briefs, one call, no tools: HyperFrames HTML (layout + GSAP) rendered to video. Gemini 4 Argon scores pairwise Elo on five axes (anchor GPT-5.5 xhigh = 1000). Claude Opus 5.5 (max with fallback) leads at 1536 Elo (~$3.08/task); GPT-6.1 Sol (max) is fourth at 1369 (~$0.31). This ranks code-designed motion demos—not Kling/Sora-style pixel generators. Dollars are AA evaluation cost per task, not a finished-film quote.

Code2Video-AARead story
34

CodeRabbit Covers Up to $5,000 in Deep Scan Usage Through November 8

Through November 8, 2026, CodeRabbit says it will cover up to $5,000 in Deep Scan usage for any organization with two or more members, against a pool of up to $1 million that is first come, first served until it runs out. Eligible orgs include open-source projects, current or former customers, and teams that have never used CodeRabbit. CodeRabbit Secure uses GPT 6.1 Sol and GPT Cyber; Deep Scan examines committed code across a repository, independently verifies exploitability, and can draft fixes for eligible findings—developers still review and ship. As of October 27, 2026.

CodeRabbit, CodeRabbit Secure, Deep Scan, GPT-6.1 SolRead story
35

Nanbeige4.3-33B Goes Open Source—Instruct, Base, and 512K Context

Nanbeige—the LLM lab at BOSS Zhipin (Kanzhun Limited)—has open-sourced Nanbeige4.3-33B Instruct and its Base sibling on Hugging Face under Apache 2.0. Both are about 33B parameters with a context window up to 512K (524,288 tokens). Nanbeige positions the Instruct build as the flagship teacher for the 4.3 series and plans to distill smaller models starting with Nanbeige4.3-3B, which is not released yet. Official benchmark claims are Nanbeige’s; for reference only.

Nanbeige4.3-33BRead story
36

Strata Puts DeepSeek V4.1 Flash on a 12GB GPU. The Catch Is 46–64GB RAM

Strata is a third-party local runner—not a new DeepSeek official model. GitHub v0.2.5 (about 18:06 Beijing time on 2026-10-25) puts DeepSeek V4.1-Flash in the Windows/Linux desktop app: three sizes, context fixed at 8K. dsv41 ships for Windows/Linux only; Mac has a 0.2.5 dmg and Metal engine but no DeepSeek. Hardware from the 0.2.4 notes still applies: about a 12GB GPU plus about 46–64GB RAM, full download about 239 / 257 / 274GB (IQ1_M / IQ2_XXS / IQ2_XS). Busy experts stay in RAM; the rest stream from SSD. Speed figures are project-reported—for reference only.

DeepSeek V4.1 FlashRead story