📆 ThursdAI - Dec 18 - Gemini 3 Flash, Grok Voice, ChatGPT Appstore, Image 1.5 & GPT 5.2 Codex, Meta Sam Audio & more AI news

Episode Summary

ThursdAI’s pre-holiday episode is a speedrun through one of the most stacked AI news weeks of 2025. The crew breaks down Google’s Gemini 3 Flash leap in price-to-performance, OpenAI’s rapid-fire releases (GPT Image 1.5, ChatGPT App Store, and breaking GPT 5.2 Codex), and NVIDIA’s major open-source Nemotron 3 Nano drop. They also cover the rapidly commoditizing voice stack with xAI Grok Voice, Chatterbox Turbo, and Meta’s SAM Audio. With Kwindla joining the panel, the conversation stays practical on what actually matters for builders shipping agents right now.

In This Episode

• 🔓 Open Source LLMs
• 🏢 Big CO LLMs + APIs
• ⚡ This Week’s Buzz
• 🔊 Voice & Audio
• 🛠️ FunctionGemma & Edge Agents
• 📰 Year-End Coverage Preview

Hosts & Guests

Alex Volkov

Host · W&B / CoreWeave

@altryne

Kwindla Hultman Kramer

Co-Founder & CEO · Daily.co

@kwindla

Wolfram Ravenwolf

Weekly co-host, AI model evaluator

@WolframRvnwlf

Yam Peleg

AI builder & founder

@Yampeleg

Nisten Tahiraj

AI operator & builder

@nisten

LDJ

Weekly co-host of ThursdAI

@ldjconfirmed

Ryan Carson

AI educator & founder

@ryancarson

By The Numbers

🔥 Breaking During The Show

🔓 Open Source LLMs

The panel highlights NVIDIA Nemotron 3 Nano as the most consequential open release of the week, not only for performance but for releasing full training data and recipes. They also cover Allen AI’s BOLMO and OLMO multimodal progress plus Mistral OCR 3’s aggressive pricing and document performance gains.

• NVIDIA Nemotron 3 Nano: 30B params, 3B active, hybrid Mamba-MoE
• NVIDIA released weights, reports, recipes, and 25T-token data details
• BOLMO: byte-level parity breakthrough from Allen AI
• OLMO multimodal video models (4B/7B/8B)
• Mistral OCR 3 claims 74% win-rate over OCR v2

🏢 Big CO LLMs + APIs

Google and OpenAI trade major launches in the same week. Gemini 3 Flash stands out for frontier capability at flash-tier price, while OpenAI pushes GPT Image 1.5 and then drops GPT 5.2 Codex during the show as breaking news.

• Gemini 3 Flash: 78% SWE-bench Verified at flash pricing
• Google tool-calling scale: up to 100 simultaneous function calls
• OpenAI GPT Image 1.5: 4x faster, 20% cheaper
• GPT 5.2 Codex: 400K context with context compaction
• ChatGPT App Store submissions opened via MCP app model

⚡ This Week’s Buzz

A community moment: Alex announces Wolfram joining Weights & Biases/CoreWeave as an AI Evangelist and ‘AIvaluator.’ The segment frames 2026 as a more benchmark-driven era for the show and the broader AI community.

• Wolfram Ravenwolf announced as joining W&B/CoreWeave
• Focus on deeper public evals and model benchmarking
• Weave highlighted for practical AI evaluations

🔊 Voice & Audio

Voice AI competition tightens with lower prices and more capable real-time stacks. xAI ships Grok Voice Agent API with Tesla integration and strong audio benchmark positioning, while open-source Chatterbox Turbo and Meta SAM Audio push accessible audio generation and separation.

• Grok Voice Agent API: $0.05/min pricing and Tesla integration
• Big Bench Audio leadership claim for Grok voice stack
• Resemble Chatterbox Turbo: MIT-licensed 350M open TTS
• Meta SAM Audio: source separation with multimodal prompting

🛠️ FunctionGemma & Edge Agents

Google’s tiny FunctionGemma release gets a dedicated discussion for what it signals about on-device agents. The model is small enough for constrained hardware and points toward privacy-first local function-calling assistants.

• FunctionGemma at 270M parameters
• ~500MB RAM footprint for edge usage
• Strong improvement after fine-tuning for mobile actions
• On-device tool use for private assistant workflows

📰 Year-End Coverage Preview

The episode closes by setting up the full 2025 recap planned for the next show. This installment is framed as week-of-news triage, with the year-in-review coming as a separate deep retrospective.

• Dec 18 episode intentionally focused on weekly drops
• Full 2025 month-by-month recap queued for next week
• Team emphasizes pace and acceleration of releases