Qwen3.5 Lands With a 27B Multimodal Flagship and 835K Day-One Downloads
π§ LAUNCH
Qwen3.5 Lands With a 27B Multimodal Flagship and 835K Day-One Downloads.
Qwen3.5 drops across multiple sizes β 27B, 4B, and 2B β and the open-weight community is voting with their bandwidth. The 27B multimodal flagship pulled 835K downloads and 638 likes on HuggingFace in its first day, making it one of the hottest open-weight releases this year. With image-text-to-text capabilities baked in, this is a serious contender for teams building multimodal pipelines without API dependencies. Download it and benchmark against whatever you're running now. (638 likes | 835K downloads) Read more β
GPT 5.4 Extra High tops LiveBench by a wide margin. The results are independently verified as not benchmark-maxxed β GPT 5.4 genuinely dominates LiveBench across categories. If you were still running evals from last week's launch hype, now you have independent confirmation: update your model tier list. (513 likes | 32 RTs) Read more β
Claude hits #1 in the App Store. That's not a developer tool chart or a productivity subcategory β that's the actual App Store, beating every app on the planet. Combined with the 1M daily signups milestone, Anthropic just proved it's not only an enterprise play anymore. (6,435 likes | 502 RTs) Read more β
π‘ INSIGHT
OpenAI Will Deploy Models on the Pentagon's Classified Network.
Sam Altman dropped the bombshell himself: OpenAI has reached an agreement to deploy models on the Department of War's classified network. He drew explicit red lines β no mass surveillance, no autonomous weapons β but the Rubicon is crossed. This is the moment AI companies stopped being tech vendors to the military and started becoming infrastructure. The safety principles outlined are worth reading in full, because they'll be cited in every AI-military debate for years. (9,285 likes | 1,139 RTs) Read more β
OpenAI and Google employees back Anthropic's Pentagon lawsuit. Nearly 40 employees from OpenAI and Google β including Jeff Dean β filed an amicus brief supporting Anthropic's lawsuit against the DoD's supply chain risk designation. Cross-company solidarity on AI governance is vanishingly rare, and this brief signals that the rank-and-file at frontier labs take the military deployment question seriously, even when their own employers are making different choices. Read more β
LeCun's AMI Raises $1B to Bet Against the LLM Paradigm.
Yann LeCun's AMI just raised $1.03 billion to build AI that learns from physical reality rather than text. This isn't a research grant β it's a war chest. LeCun has been arguing for years that language models are a dead end for human-level AI, and now he has the funding to prove it. Whether he's right or not, the sheer size of this raise forces every AI strategy to at least hedge against the possibility that scaling LLMs isn't enough. (226 likes | 39 RTs) Read more β
Meta acquires Moltbook, a social network for AI agents. Yes, you read that correctly β Meta bought a Reddit-like platform where AI agents post, interact, and form communities, then folded the team into Meta Superintelligence Labs. Agent-to-agent communication is now a real product surface at one of the world's largest tech companies. If you're building agents, start thinking about what happens when they talk to each other. Read more β
Anthropic crosses 1 million daily signups. That growth rate is staggering for any software product, let alone one competing against ChatGPT's brand recognition. Anthropic's consumer momentum is now matching its developer-platform push β the flywheel is spinning. (3,838 likes | 224 RTs) Read more β
π§ TOOL
Figma MCP Server Closes the Design-to-Code Loop.
The Figma MCP server now supports full roundtrip β pull design context into your coding agent, push rendered UI back to Figma as editable frames. This isn't one-way screenshot-to-code anymore; it's a bidirectional workflow where design and implementation stay in sync. If you're doing any frontend work, set up this MCP integration today β the design-to-code gap just collapsed. (85 likes | 10 RTs) Read more β
Kali Linux ships a fully local AI pentesting stack via MCP. Kali published a guide for AI-assisted penetration testing that runs entirely locally β Ollama for inference, 5ire for the interface, and MCP Kali Server to give the LLM access to security tools. No cloud APIs, no data leaving your machine. This is one of the most practical MCP use cases outside of coding. (96 likes | 15 RTs) Read more β
Claude Opus 4.6 now generates designs directly inside Figma. divRIOTS shipped a Figma plugin powered by Claude Opus 4.6 β type a prompt, get a design. The html-to-design crowd is now going the other direction, and the results are surprisingly usable for rapid prototyping. (255 likes | 9 RTs) Read more β
HuggingFace ships community evals and dataset chat agents. HuggingFace dropped community-contributed benchmark evals, conversational dataset agents, and spreadsheet-style Data Studio UX in a single release. The Hub is quietly becoming a full data development environment, not just a model registry. (84 likes | 19 RTs) Read more β
π¬ RESEARCH
Qwen3.5-4B outscores GPT-4o on classic benchmarks. Simon Willison flagged what might be the most important number of the week: a 4-billion-parameter model beating last year's frontier on established benchmarks. The efficiency curve isn't just steepening β it's breaking assumptions about what requires a massive model. If you're not testing small models for edge and local deployment, you're leaving money on the table. (567 likes | 32 RTs) Read more β
LeCun argues human-level AI requires mastering the physical world. In a WIRED profile, LeCun lays out his full thesis: language is a lossy compression of reality, and models trained only on text will never achieve genuine understanding. With $1B now backing this bet through AMI, it's worth engaging with the argument seriously β even if you think he's wrong. (422 likes | 97 RTs) Read more β
Open weights isn't open training β and it matters. A sharp essay argues that releasing model weights without training code, data, and methodology isn't truly "open" in any meaningful scientific sense. Reproducibility requires the full recipe, not just the cake. Worth reading before you call your next release "open source." (10 likes | 1 RT) Read more β
π TECHNIQUE
The unbounded agent skills problem: how to give one agent unlimited tools. Brendan Falk poses the question every production agent team hits eventually β what's the best architecture for giving a single agent access to a potentially unbounded number of skills without tanking reliability? The thread replies are a goldmine: tool registries, dynamic discovery, hierarchical routing, and hard-won failure modes. Bookmark this one. (87 likes | 4 RTs) Read more β
ChatGPT vs Claude in Excel: a 1,000-year stress test. Ethan Mollick threw 100+ tabs of macro-economic data spanning a millennium at both ChatGPT and Claude's Excel integrations. Both handled it, but with telling differences β ChatGPT stayed in-app while Claude preferred to export and analyze externally. If you have a genuinely hard spreadsheet problem, try both. (1,253 likes | 82 RTs) Read more β
ποΈ BUILD
SOTA video generation running locally on a MacBook via MLX. LTX 2.3, an open-source video generation model, is now running on Apple Silicon through a custom MLX runtime β and the runtime itself was built using GPT 5.4. A ComfyUI adapter and full open-source release are incoming. Local video gen just went from "technically possible" to "actually usable." (55 likes | 4 RTs) Read more β
π MODEL LITERACY
Parameter Efficiency and Knowledge Distillation: Qwen3.5-4B beating GPT-4o on classic benchmarks isn't a fluke β it's the result of knowledge distillation, where a smaller "student" model is trained to mimic the behavior of a larger "teacher" model, combined with aggressive data curation that prioritizes quality over quantity. The key insight: you don't need to discover intelligence from scratch in every model; you can compress what a frontier model already knows into a fraction of the parameters. This means the cost curve for "good enough" AI is collapsing faster than most product roadmaps assume β the model that cost $100M to train last year can be approximated by one that costs $1M this year. If your architecture assumes only large models can handle your workload, it's time to re-benchmark.
β‘ QUICK LINKS
- UV scripts for HuggingFace dataset conversion: One-liner format conversion between COCO, YOLO, and other object detection formats. (53 likes | 8 RTs) Link
- Dataset β GPU embeddings β interactive visualization: Single-command pipeline for quick dataset exploration via HF Inference. (33 likes | 7 RTs) Link
- Building a programming language with Claude Code: A developer walks through creating an entire language with agentic coding β solid case study. (31 likes | 34 RTs) Link
- Grammarly used real authors' names as AI personas without consent: Their Expert Review feature listed journalists as AI editors. Opt-out only. A cautionary tale. Link
π― PICK OF THE DAY
LeCun's $1B raise isn't just a funding round β it's a paradigm bet with real money behind it. Yann LeCun has been the most prominent critic of the "just scale language models" thesis for years, and now AMI has $1.03 billion to prove him right. The argument is simple: language is a lossy compression of reality, and models that only learn from text will hit a ceiling on genuine reasoning and understanding. What makes this moment different from every other "LLMs aren't enough" blog post is the price tag β this is the loudest signal yet that serious insiders believe the pure language-model scaling paradigm has limits. For the rest of us, the practical takeaway is hedging: if your entire AI strategy assumes LLMs will keep scaling to AGI, you now have a billion-dollar counterargument to account for. Whether LeCun is right or not, the race to world models is now a funded thesis, not a contrarian position. Read more β
Until next time βοΈ