ThursdAI - Jul 16 - Inkling 975B open weights, Kimi K3 at 2.8T, a 27B model on a phone & Codex hits 9M
From W&B CoreWeave - Just a few min before midnight, but still on Thu! Here's everything big important that happened this week that you need to know about!
Hey yall, Alex here,
Huge thanks to Wolfram for running point on the live show this week. Didn’t have tons of time to edit this one, so please skip the first 10 minutes, it’s a loop of our new “wait for the live show to start” vid, that I build with HyperFrames and can’t wait to tell you about, next week!
Today it seems that OpenSource is biting back, with Kimi K3 getting released just a short while after Thinking Machines (Thinky) has released Inkling, their near 1T model.
I’m attaching the TL;DR and timestamps for the full show (my AI agents, yes even Fable and Sol are not a match yet at editing down hehe) and I’ll spare you the long Fable recap (please do let me know in the comments if you were expecting it)
0:00 – Intro, Alex on vacation, TLDR overview
11:35 – TLDR: Thinking Machines, open source, OpenAI news
12:34 – Banter: impressions of Sol/Codex, over-verification behavior
37:22 – TLDR restart & detailed breakdown
48:40 – Open Source AI section begins (Bonsai/Prism ML, Kimi K3)
58:42 – Inkling (Thinking Machines) deep dive & 3D model visualization
1:10:33 – Kimi K3 discussion & demo comparisons
1:27:02 – Frontier Labs: AGI governance framework discussion (Demis Hassabis essay)
1:47:04 – Grok Build CLI data leak & OpenAI file deletion incident
2:02:15 – This Week's Buzz: Wolfbench results on GPT 5.6 Sol/Terra/Luna
2:09:52 – Closing remarks & sign-off
The one-minute version: Mira Murati's Thinking Machines released Inkling, a 975B parameter open-weights MoE under Apache 2.0, the top US open-weights model right now. Moonshot's Kimi K3 went from rumor to released API during the show, confirmed at 2.8 trillion parameters with open weights promised within days, and it's already topping early arena boards. PrismML's Bonsai 27B squeezes a full 27B model into 3.9 gigabytes so it runs on a phone. Codex and ChatGPT Work blew past 9 million users, OpenAI confirmed and explained the Sol file-deletion bug (back up your machines, folks), and xAI's Grok Build CLI got caught uploading entire private repos before open-sourcing the whole thing in response. Plus Wolfram's fresh Wolfbench numbers on the GPT-5.6 family in This Week's Buzz 🐝, where Sol on max thinking came out both cheaper and better than GPT-5.5's best.
TL;DR and show notes
Hosts and Guests
Wolfram Ravenwolf, guest host this week (@WolframRvnwlf), while Alex Volkov (@altryne) is on vacation
Co-hosts: @yampeleg, @nisten, @ldjconfirmed, @petergostev
Open Source LLMs
Thinking Machines releases Inkling - 975B total / 41B active MoE, trained from scratch on 45T multimodal tokens, Apache 2.0, top US open-weights model at 41 on the Artificial Analysis Index, encoder-free text/image/audio, Inkling-Small (276B/12B) previewed (X, Blog, HF)
PrismML Bonsai 27B - 1-bit (3.9GB, ~90% retention) and ternary (5.9GB, ~95% retention) versions of Qwen 3.6 27B, multimodal, 262K context, Apache 2.0; Nisten demoed it live on a phone and a 6GB 1660 Ti (X, Blog, HF)
MOSS-VL-Realtime - open source 11B VLM for real-time streaming video with proactive speaking and proactive silence, SOTA on all three open proactivity benchmarks, ~22.7GB, base model included (X, HF, GitHub, Arxiv)
Kimi K3 API drops mid-show - confirmed 2.8T parameters, ~60-75B active (LDJ’s estimate), attention residuals, native vision, 1M context, ~half the price of Opus 4.8 / GPT-5.6 Sol, open weights promised within days; post-show, Arena reports K3 debuting #1 on Frontend Code Arena above Fable 5 (early results, caveats apply) (X, Arena)
Big CO LLMs + APIs
Codex + ChatGPT Work unified app hits 9M active users, up from ~6M days earlier and 1M in February; 5-hour windows replaced with banked, expiring resets (X)
OpenAI confirms GPT-5.6 Sol file-deletion bug: $HOME override in full-access mode without sandbox or auto-review can nuke real home directories; mitigations and post-mortem promised (X, Techzine)
OpenAI ships first hardware, the $230 kbd-1.0-codex-micro Codex controller with a reasoning-effort dial, built with Work Louder; sold out (X, Work Louder)
GPT-Red - OpenAI’s internal automated red-teamer finds prompt injections at 84% vs 13% for humans, makes Sol 6x more injection-resilient, discovers the fake chain-of-thought attack class (X, Blog)
ChatGPT returns to WhatsApp in the EEA after an EU antitrust order forces Meta to reopen to third-party AI bots; Kakao and Viber rollouts too (X)
Google patches the Gemma 4 family - Flash Attention 4 (25-70% prefill speedup), tool calling fixes, reduced laziness, configurable vision resolution; criticized for shipping new weights with no version bump (X, HF)
xAI’s Grok Build CLI caught silently uploading full private repos (history, deleted files, secrets) to Google Cloud Storage despite opt-outs; xAI deletes data, disables retention, and open-sources the CLI under Apache 2.0 (X, xAI response, GitHub)
Demis Hassabis publishes an AGI governance essay proposing a FINRA-style Frontier AI Standards Body; endorsed by Altman, Nadella, Pichai, and Suleyman; the panel debates it hard on the show (X, Essay)
This Week’s Buzz
Wolfbench adds GPT-5.6 Sol, Terra, and Luna on Terminal Bench 2.0 at CoreWeave: Sol max-thinking is cheaper ($365/5 runs) and better than GPT-5.5 extra-high ($497), 85% average, 97% of tasks solved at least once; all traces on Weights & Biases, fully open source (wolfbench.ai)
Show and tell