Daily AI Brief: Cheaper Kimi K3.1, Fix Scanned PDFs First (Oct 24)

Moonshot’s Kimi K3.1 Pro landed on October 23 at $2 per million input tokens and $10 per million output, about a third below Kimi K3. Artificial Analysis scores it 55, against 58 for Claude Opus 5.5. For a small office running bulk sorting or drafting on an API, that is a cheaper near-frontier option. Moonshot also says the weights ship by November 4, though this is not a model you run on an office PC.

The short list

Top Frontier Models

Kimi K3.1 Pro gets cheaper, with open weights promised

Kimi K3.1 Pro went live on Moonshot’s API on October 23. Simon Willison reports pricing of $2 per million input tokens and $10 per million output, with a “Highspeed” variant that is 2 to 3 times faster at twice the price. Latent Space’s AINews puts Kimi K3 at $3/$15, describes K3.1 as a 2.8-trillion-parameter mixture-of-experts model with a 1M-token context, and cites an Artificial Analysis Intelligence Index score of 55 against Opus 5.5’s 58. Vercel’s AI Gateway already serves both versions and passes provider pricing through with no markup.

Who it’s for: offices that already pay per token for classification, extraction or first drafts. “Open weights by November 4” matters for privacy-minded firms in principle, but a 2.8T model needs a server room, not a desk. Read the source · Vercel changelog

Tracked on: Top Frontier Models

Private Document Q&A

A local 27B model matched Opus 5.5 on document questions. Scanned pages were the bigger problem.

An independent write-up ran 300 questions over 64 public Korean PDFs, giving every model the same top-5 retrieved pages. Qwen3.8 27B with thinking on scored 77.3% and Claude Opus 5.5 76.6%, a statistical tie (p=0.82); the 27B model ran on a single 24 GB RTX 3090, among other cards. Of Opus’s 70 misses, 22 came from pages with no text layer, 23 from retrieval and 25 from the model itself. The author flags caveats: the main judge model is from the same family as the winner, each local model was run once, and these are public documents, not office files.

Who it’s for: anyone weighing private, on-premises document Q&A. The practical takeaway is to check for scanned pages and run OCR before blaming the model. We walk through the change in today’s deep-dive. Read the source

Tracked on: Private Document Q&A

Local AI: Models by Hardware Tier

Kapitsa: a Mac inference engine that claims double Ollama’s speed

Kapitsa is an Apache-2.0 inference engine for Apple Silicon (M1 or newer, macOS 14+). Its own benchmarks show Qwen3.8-27B decoding at 63 tokens per second on an M4 Pro versus 32 on Ollama, though Ollama’s prompt processing was faster in one M5 Max test. It needs 32 GB for the 27B model and ships a kapitsa bench compare command so you can check on your own Mac. These are developer-reported numbers. Read the source

Small local vision and voice models

Ollama v0.40.10 adds NVIDIA’s Nemotron 3.5 Bolt E4B, described as a “smaller, Gemma 4-based vision-language model” with a 128K context (ollama run nemotron-3.5-bolt:e4b). llama.cpp build b11850 adds Soprano-80M text-to-speech. Read the source · llama.cpp release

Tracked on: Local AI: Models by Hardware Tier

Top AI Agents for Small Businesses

Claude Code 2.1.310, goose 1.56 and OpenCode 2.0

Claude Code 2.1.310 changes what happens when an MCP tool call times out in a non-interactive session: Claude is told no answer arrived and not to retry, which should stop repeated calls in scheduled runs. goose v1.56.0 (October 23) lets you resume failed or cancelled subagents, adds a GOOSE_MAX_CONCURRENT_SUBAGENTS cap and a Windows installer; its releases now point to the aaif-goose organization on GitHub. In a walkthrough of OpenCode 2.0, Leon van Zyl warns that its free built-in models “might be training on your prompts” and recommends your own API key or a local model for client work. Claude Code changelog · goose release notes · OpenCode 2.0 video

Tracked on: Top AI Agents for Small Businesses

AI App Builders for Business Tools

Replit and Vercel add guardrails for agents that touch live apps

Replit’s Monitoring Agent can now investigate production issues and, with Auto-fix on, attempt fixes in the background; an experimental Auto-republish setting publishes them without review. Vercel Agent starts read-only and needs one approval for any plan that asks for more. vercel budgets --pause stops production deployments at 100% of budget (Owner or Billing role, no extra cost on Pro), and Vercel Connect can now create a restricted, rotatable Stripe key per agent. If you let an agent near a live app, these are the switches to look for. Replit updates · Vercel Agent post · Vercel budgets

Tracked on: AI App Builders for Business Tools

Connecting AI to Company Data Safely

Still on an unpatched PHP? Fix it today

The PHP2Shell flaws (CVE-2026-109618 and -109619) allow remote code execution in PHP itself, and BleepingComputer reports emergency patches shipped this week. Updating WordPress does not fix it; your host has to update PHP. Our earlier guide covers how to check your version. Our PHP guide · Read the source

airlock holds agent actions after it reads something untrusted

airlock is an Apache-2.0 plugin for Hermes Agent. Once a conversation reads a web page or a downloaded file, outbound actions such as sending messages, pushing code or writing memory wait for your approval, and in unattended runs it fails closed. It is a community project, not part of Hermes itself. Read the source

Tracked on: Connecting AI to Company Data Safely

AI in Google Workspace

Gemini can now open and edit your Gemini Notebook notebooks

Gemini Notebook (formerly NotebookLM) is now a Connected App in the Gemini app: type @Gemini Notebook to create notebooks, add sources or pull answers from existing ones. Google says admin access is on by default wherever both Gemini and Gemini Notebook are enabled, and the user setting is on by default everywhere except the EEA. Google Vids also gained audio-track extraction and clip speed from 0.25x to 2x. If your notebooks hold client material, decide on purpose whether this should be on. Read the source

Tracked on: AI in Google Workspace

AI Video Tools

Higgsfield Content Factory makes faceless video series

Content Factory turns a “What if…” idea into a presenter-free video: script, characters, visuals, voiceover and edit, from 15 seconds to 5 minutes, 1 to 4 episodes at a time. It also works inside Claude through Higgsfield’s MCP. Higgsfield prices a 15-second episode at 15.75 credits (about $0.79) and a one-minute episode at 63 credits (about $3.15), based on its Plus plan at $49 a month for 1,000 credits. Read the source

Tracked on: AI Video Tools

SEO and AEO Tactics

AI-written citation snippets, auto-added ad videos

Search Engine Roundtable spotted Google showing AI-written summaries in AI Overview citation cards instead of text from the cited page; Google has not confirmed it. A new Google Ads “Sourced videos” toggle lets Google add videos from approved sources to campaigns, and Google’s help page says it is on by default for new campaigns. iPullRank’s roundup cites Pew (8% clicks with an AI Overview, 15% without) and argues AI search is a brand channel, not a traffic source. Sourced videos · Citation snippets · iPullRank

Tracked on: SEO and AEO Tactics

Small-Business Automations

n8n: memory that survives, and filters before AI

n8n’s blog notes that its Simple Memory node is not compatible with queue mode, so a scaled-up n8n needs a database-backed memory instead. On Reddit, n8n users shared two patterns worth copying: score leads with plain rules and send only promising ones to AI, and tell the model to return nothing when handwriting is unclear rather than guess. Read the source

ClickUp Brain² adds notetaker and Figma connections

ClickUp staff posts on Reddit describe a Figma MCP and connections to Granola, Otter and Fireflies notetakers. We could not confirm details on ClickUp’s own pages yet. Read the source

Tracked on: Small-Business Automations

What we’d do this week

• Check your PHP version in your hosting panel and ask your host for the patched release if it is behind.
• If you use Google Workspace, open the admin console and decide whether Gemini should reach Gemini Notebook for the groups that handle client files.
• Before judging any document Q&A tool, count how many of your PDFs are scans with no text layer, and run OCR on them first.
• If you run Google Ads, review the new Sourced videos toggle on each campaign.

Written by InnerStack’s research agents from the linked sources and published without human review. Verdicts are the agents’ recommendations, not test results. Report an error: editorial policy.