Qwen 3.8 Max Drops 2.4 Trillion Parameters Aimed Squarely at Coding

🧠 LAUNCH

Qwen 3.8 Max Drops 2.4 Trillion Parameters Aimed Squarely at Coding

Qwen 3.8 Max is here with 2.4 trillion parameters and a 27B sibling optimized for coding and collaborative workflows. The Max model goes head-to-head with GPT-5 and Claude on code benchmarks — and the 27B variant hits a sweet spot for teams that want strong coding performance without frontier-scale inference costs. Two models, one message: Alibaba's AI lab isn't just competing, it's forcing the frontier to look over its shoulder. Read more →

Liquid AI Ships a 2.6B Agent Model Built for Your Phone. LFM2.5-2.6B is a small language model designed from the ground up for on-device agent deployment — already at 47K downloads on HuggingFace. The architecture trades raw benchmark scores for efficient tool-calling and multi-step reasoning in constrained environments, which is exactly what edge agents need. If you're building anything that runs without a cloud GPU, this is your new baseline. (144 likes | 47.4K downloads) Read more →

Mistral's Shieldstral: A 3B Open-Weight Safety Filter You Can Run Locally. Shieldstral is a purpose-built 3B moderation model handling both text and images — small enough to slot into any pipeline as a local guardrail without adding cloud latency or cost. At 3B parameters, you can run it alongside your main model on the same hardware, giving every deployment a built-in content filter. (284 likes | 69 RTs) Read more →

LG AI Research Enters the Frontier Ring with a 750B MoE Model. K-EXAONE 2.0 from LG AI Research is a 750B Mixture-of-Experts model with only 37B active parameters — meaning it delivers frontier-class reasoning while keeping inference costs closer to a mid-range dense model. It's a serious statement that frontier-scale development is globalizing beyond the usual US and Chinese players. (116 likes | 325 downloads) Read more →

🔧 TOOL

Anthropic Finally Publishes the Claude Cost Control Playbook

Anthropic drops its first comprehensive guide to monitoring and managing Claude costs — covering usage dashboards, rate limits, spend alerts, and org-level controls. If you're running Claude in production and your finance team keeps asking "how much are we spending on AI?", this is the doc you send them. It's less about new features and more about making existing controls discoverable — which is arguably more valuable. Read more →

Interconnects Launches a Public Dashboard Tracking Which Open Models People Actually Use. Interconnects ships an adoption dashboard with real download numbers, deployment counts, and community forks — finally giving the open-model ecosystem data on what's actually getting used versus what's just getting benchmarked. Bookmark it before your next model selection meeting. Read more →

💡 INSIGHT

Latent Space Reverse-Engineers How ChatGPT Work Actually Works

Latent Space publishes the most detailed external teardown of ChatGPT Work's agent architecture to date — breaking down Memory, Proactivity, Scheduling, Browser Use, Plugins, Skills, and Tools as distinct orchestration layers. The key insight: OpenAI isn't just shipping a better chatbot, they're building a persistent coworker with its own agency loop. If you're building agents, this is required reading — not for the model, but for the orchestration patterns. Read more →

Google's July 2026 AI Recap: Everything You Missed in One Place. Google consolidates every AI announcement from July — model updates, product launches, and research milestones — into a single reference doc. Not groundbreaking on its own, but genuinely useful as a catch-up if you blinked during a busy month. Read more →

Interpol: AI Now Powers Over Half of Cybercrime Across Africa. Interpol reports that AI-driven voice cloning, deepfake scams, and automated phishing now account for more than 50% of cybercrime across Africa — at a scale that's outpacing law enforcement capacity. A sobering reminder that AI's most impactful "use cases" aren't always the ones celebrated at product launches. (108 likes | 52 RTs) Read more →

📝 TECHNIQUE

353,000 People Took Google's Free AI Agents Course — Here's What They Built

Kaggle's AI Agents Intensive attracted 353,000 learners — and Google is now sharing course structure, completion rates, and what participants actually built. The scale alone is the story: agent development is crossing from "niche developer skill" to "thing your PM took a course on." The course materials are free and cover tool use, multi-step planning, and real deployment patterns — worth reviewing even if you've been building agents for months. Read more →

🔬 RESEARCH

When Benchmarks Stop Working: A Systematic Study of Saturation. This paper documents how major AI benchmarks lose discriminative power as models saturate them — and proposes metrics for detecting when a benchmark has stopped being useful. If you've ever wondered whether that 2% SOTA improvement is signal or noise, this gives you a framework to answer. (72 likes | 78 RTs) Read more →

A Ternary 20B MoE Model Hits 120 Tokens/Second on an iPhone. Maple-Preview is a ternary-weight 20B Mixture-of-Experts model running at 120 tok/s on consumer Apple hardware — a concrete proof point that on-device inference at useful model sizes is arriving faster than anyone's deployment roadmaps assumed. The trick is ternary quantization combined with MoE routing, which slashes both memory and compute requirements. (31 likes | 10 RTs) Read more →

OpenAI Highlights Ten Math and CS Breakthroughs with ML Implications. OpenAI curates ten recent advances in mathematics and theoretical computer science that connect back to machine learning — spanning optimization, complexity theory, and algorithm design. Strong HN engagement suggests the list is genuinely useful for researchers scanning adjacent fields. (389 likes | 673 RTs) Read more →

🏗️ BUILD

MiniMax H3 Now Runs Natively on Your Mac via MLX. Simon Willison highlights the PipeNetwork port that brings MiniMax H3 to Apple Silicon through MLX — meaning you can run the open-weight video model locally without cloud GPU costs. If you've been waiting for a zero-cost way to experiment with H3, this is it. Read more →

The MiniMax H3 Ecosystem Explodes: ComfyUI, GGUF, and INT8 Variants. The community is making MiniMax H3 production-ready across hardware tiers — Comfy-Org's ComfyUI integration leads with 440 likes, GGUF quantizations are already at 40K downloads, and INT8 variants open up deployment on consumer GPUs. Pick the variant that matches your hardware and start building. (440 likes | 40.0K downloads) Read more →

🎓 MODEL LITERACY

Active Parameters vs. Total Parameters in MoE Models: Today's launches include a 2.4T Qwen Max and a 750B K-EXAONE with only 37B active — so what do those numbers actually mean? In a Mixture-of-Experts (MoE) model, total parameters is the full weight count across all expert sub-networks, but only a fraction — the "active" parameters — fire for any given input. A learned router selects which experts handle each token, so a 750B MoE model might only activate 37B parameters per forward pass, using roughly the same compute and memory as a dense 37B model. This is why raw parameter counts are increasingly misleading: a 750B MoE can run on hardware that would choke on a dense 70B, while potentially matching or exceeding its quality. When comparing models in 2026, ask "how many active parameters?" before asking "how many total?"

⚡ QUICK LINKS

• OpenAI Cyber Evaluations: Publishes results from independent third-party security evaluations of its models. (31 likes | 2 RTs) Link
• Simon Willison on Steve Yegge: Surfaces a Yegge quote on AI development that aged perfectly. Link
• Palantir posts $1B profit: CEO Karp calls AI industry "Marxist," positions Palantir as the enterprise alternative. Link

🎯 PICK OF THE DAY

Latent Space's teardown of ChatGPT Work is the agent architecture doc every builder should read. The piece reverse-engineers six distinct orchestration layers — Memory, Proactivity, Scheduling, Browser Use, Plugins, and Skills — and reveals that OpenAI's real moat isn't model quality. It's the orchestration layer that turns a chatbot into a persistent coworker. Memory gives it continuity across sessions. Proactivity lets it act without being prompted. Scheduling makes it reliable. The lesson for anyone building agents: the hard part isn't making the LLM smarter, it's building the scaffolding that makes it useful day after day. Every team shipping agent products should study this playbook before reinventing the wheel — the patterns are general, even if the implementation is OpenAI-specific. Read more →

Until next time ✌️