Nvidia's free AI transcribes 40 languages

Hi there, this is your daily ☕️ AIpresso.

Nvidia's open AI transcribes 40 languages LINK

• Nvidia released Nemotron 3.5 Transcribe, a free, open-weight model that turns speech into text across 40 languages, translates 31 of them to or from English, labels up to eight speakers, and answers questions about recordings up to an hour long.
• The model has 34B parameters but activates only about 4B per step, letting it run on a single Nvidia chip from Jetson boards to data-center GPUs while topping Hugging Face's English speech-recognition leaderboard.
• Its license permits commercial use and fine-tuning cut errors 45.8% on Indian medical speech in about 14 hours on one GPU, though the model is built for recorded audio, not live transcription.

Gemini 4 Carbon beats GPT-6 but access is limited LINK

• Google's Gemini 4 Carbon tops GPT-6 Astra and Claude Opus 5.5 on most of 17 benchmarks, but access is restricted to a small group of cybersecurity researchers with no public release date.
• Carbon scored 59 on version 5.0 of the Artificial Analysis Intelligence Index at its high setting, one point ahead of Claude Opus 5.5, and beat Claude Opus 5.5 on Terminal-Bench 4.0, where Google cited Artificial Analysis rather than its own numbers.
• Google says Carbon will match Argon's pricing at $2 per million input tokens and $10 per million output, undercutting Claude Opus 5.5 and GPT-6 Astra, though most comparison figures come from Google's own reporting and access stays limited to approved defenders.

Pluto Security reveals Hugging Face cryptomining malware campaign LINK

• Pluto Security's Load/Mine report, published today, details a cryptomining and backdoor campaign spread through 31 malicious Hugging Face model repositories across seven accounts plus a ModelScope mirror, logging roughly 30,000 downloads.
• The attack abuses Transformers' documented remote-code loading, which runs a repository's top-level Python with the caller's privileges before weights load, dropping a UPX-packed miner disguised as a kernel worker plus reverse shells; it also cites CVE-2026-22807 affecting earlier vLLM versions.
• The flagship second-wave payload returned "safe" from every scanner and stayed live eight days after disclosure, though Pluto notes Hugging Face's own docs call its ClamAV and pickle scanning "not 100% foolproof," placing safety checks on the user.

Mistral open-sources its 1T-parameter Large 4 model LINK

• Mistral has open-sourced the weights of Mistral Large 4, the 1T-parameter model it previewed three weeks ago, posting it to Hugging Face after three extra weeks of customer-guided reinforcement learning sharpened its agentic coding and cyber defence.
• Nicknamed "Le Chonk," ML4 activates 52B parameters per task, reads text and images, and jumped on Terminal-Bench 4.0 from 28% at preview to 41%, edging GLM-5.3 while using 47% fewer tokens per task on the Coding Agent Index.
• Mistral added low and medium reasoning modes, made high mode 50% more efficient, and ships a quantised build running on four Nvidia B200 or B300 GPUs, though API pricing has not yet been published and distilled laptop-sized versions arrive on no committed date.

Copilot Fusion nearly matches Opus LINK

• GitHub unveiled Copilot Fusion at Universe on October 28, an orchestration engine that routes coding requests among models and, in Artificial Analysis testing, nearly matched Claude Code on Opus 5.5 for roughly half the cost.
• On the Coding Agent Index v2.0, Fusion Max averaged $5.59 per task versus $11.42 for Opus 5.5 Max while scoring 60 to Claude Code's 62, using about half the tokens by escalating only when a cheaper model's output falls short.
• Fusion is generally available across Copilot plans in VS Code, JetBrains and CLI, billing each called model at standard rates, though it lagged Opus 5.5 on the FrontierCode repository-change evaluation, scoring 51 to 56.

Strategies & Tactics

> How Uber Built A Gateway For AI Agents To Access Tools: Uber built a shared gateway that lets AI agents discover and safely use thousands of internal tools through one standard protocol, sparing each team from solving the same integration problems.

> Meta Andromeda: Powering the Next Generation of Ads Retrieval Through Model-System Co-Design: Meta rebuilt its ads retrieval engine to score every eligible ad at over a billion per second, matching ads across far larger pools to lift advertiser conversions.

> How to Prevent Prompt Injection: 13 Attacks, Tested: Stop prompt injection by checking each tool action against fixed rules the model can't override, since filtering the text alone never fully works.

> Why Agents Love Supabase: AI coding agents pick Supabase because its database describes its desired state in editable files, letting an agent read and change the backend without reconstructing its history.

> How to build evidence-grounded agent memory with OpenWiki and Graphs: Explains how to make AI agents answer cross-source questions accurately by compiling documents into a linked wiki with an evidence graph tracing claims to their sources.

Other news you might like

• OpenAI's Astra becomes the main target for Chinese AI labs after Anthropic tightens its defenses LINK

• Microsoft recorded $10 billion in revenue from OpenAI LINK

• GitHub saw 7.4B commits in September, adding 4M CPU cores and 160 PB of storage to keep up LINK

• Artificial Analysis Overhauls Its Intelligence Index With Harder, Private Benchmarks LINK

• Deepgram acquires Pipecat-maker Daily to bolster its voice AI infrastructure LINK

• Unreleased Claude model found attack chains in Chrome, Safari: Anthropic LINK

• Anthropic just launched tools to show companies what their AI spending is buying LINK

• iCloud Mail and Calendar plugins now work with ChatGPT LINK

• Formalizing the Enormous Theorem LINK

Trending tools

App Store Screenshots: open-source skill for Claude Code, Codex, and Cursor that builds a local editor to style and export app store screenshots in every size and language LINK

Viggle JST-2: physics-aware motion model that generates and controls characters, scenes, and 3D animation for controllable video and real playable game builds. LINK

Intrascope 2.0: unifies leading AI models, company knowledge, and agentic workflows in one workspace, with model routing, RAG, and admin controls for budgets and access. LINK

OpenLLM: routes Claude Code, Codex, and OpenCode through one local layer, unifying subscriptions, local models, and API keys with automatic fallback chains. LINK

Kin: connect GitHub, Gmail, or Stripe once and set allow, ask, or block rules controlling what AI assistants like ChatGPT and Claude can do. LINK

Basedash: an AI-native business intelligence platform that builds dashboards from natural language, connecting your data and generating visualizations without writing SQL. LINK

Trending papers & reports

Reproducible AI answers get a quality gate that stops the same question from producing different results when a system bundles it with other users' requests, lifting consistency from ~68% to ~92% while keeping ~92% of the speed. LINK

AI agent tool outputs often get silently chopped before the model sees them, with a third of results cut and key evidence deleted, quietly sabotaging the agent's answers even when memory sits nearly empty. LINK

Model swap detection gives buyers of language model APIs a reliable way to catch providers quietly rerouting some requests to a cheaper model while billing for the pricier one, even at low swap rates. LINK

Model compression shrinks large language models to run on far less memory while beating leading rival methods on language, coding, and reasoning tests, making cheaper AI deployment practical without sacrificing accuracy. LINK

Smart cache cleanup picks the best way to decide which saved chatbot computations to discard as live traffic shifts, lifting reuse rates by 5 points on average and cutting serving costs without reading users' prompts. LINK

See you tomorrow for a new dose of ☕️ AIpresso!