Which local models can actually handle tool calling? I built a framework to find out. 15 scenarios. 12 tools. Mocked responses. Temperature 0. No cherry-picking. Tested every Qwen3.5 size from 0.8B to 397B, and since some of you asked after the distillation tests: yes, I included Jackrong's Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled too. Only two models went all green: the 27B dense and the distilled 27B. The 397B? Failed two tests. The 122B? Failed one. The 35B? Failed two. The timed-out results — mostly on the smaller models, are cases where the model got stuck in a loop, repeating the same tool call until it hit the 30-second limit. The test that exposed the most models: "Search for Iceland's population, then calculate 2% of it." Simple, but 35B, 122B, and 397B all used a rounded number from memory instead of the actual search result. They didn't trust their own tool output. Small models hallucinate data. Big models ignore data. The 27B just threaded it through.

You know that "But, wait..." moment in every LLM thinking trace? I made it visible. I asked 8 models the same tricky probability question and rendered their reasoning as trees. Every time a model rejects its own idea and pivots, every "But...", every "Wait, actually...", a new branch grows. Same question. Completely different minds.

🐱 LongCat-2.0 is now fully open-source — MIT licensed, no restrictions. Since our launch a few days ago, the response from the community has been incredible. Thank you for all the feedback, discussions, and interest. Today, we’re releasing the model weights and inference code to everyone. ◆ 1.6T MoE · ~48B active · 1M token context ◆ Agent-native: Integrates directly with Claude Code, OpenClaw, and Hermes Agent ◆ Deployment: Support both GPU and NPU platforms— verified on large-scale domestic clusters 📑 Tech Blog: longcat.ai/blog/longcat-2.0/ 🤗 HuggingFace: huggingface.co/meituan-longc… 💻 GitHub: github.com/meituan-longcat/L… 🪄 ModelScope: modelscope.ai/collections/me… 👇 Inference Code GPU: github.com/sgl-project/sglan… NPU: github.com/meituan-longcat/S…

Which LLM has the guts to answer first? I built a game to find out. I made 4 LLMs play a guessing game: an image hidden behind 100 tiles, revealed one by one. The only question: who dares to answer first (and correctly)? The results say a lot about model "personality": > 🎈Balloon: Qwen3.6 35B A3B guessed after just 2 tiles. TWO. > 🌕Lighthouse (with a moon as a trap): 35B guessed too early again "moon". Meanwhile Gemma4 26B A4B waited for 15 tiles before committing. > 🦉Camouflaged owl: everyone got it, 3–7 tiles. 35B again the first one to answer. Takeaways: > Qwen3.6 35B A3B = the gambler. Fastest when right, but the only one that failed. > Qwen3.6 27B & Gemma4 31B = balanced. > Gemma4 26B A4B = the accountant. Correct on everything, but needed 2–5x more evidence. Speed vs certainty is a real tradeoff in these models, and you can literally watch it happen.

Anyone else enjoy reading an LLM's thought process?

Claude Fable 5 vs Opus 4.8: three rounds of canvas animation. Single HTML file, no libraries. These models can disassemble a Mac Studio and install a CPU correctly, but neither can fold a paper airplane. 🔧 Mac Studio 360° teardown Both delivered clean exploded views with labeled internals. But the motion tells the story: Fable's rotation is butter-smooth, easing into a slowdown as the parts separate. Opus's movement feels awkward by comparison. 🧠 CPU installation Both know the ritual: lever, gold triangle alignment, lock, paste, cooler. One catch: Fable's thermal paste just materializes out of nowhere. No syringe. ✈️ Paper airplane origami The round nobody won. Neither folded the paper correctly. Fable got closer but still not perfect.

By the way, this is how GLM 5.2 handled the CRT Typing test. It costs $0.0534, about 1/4 the cost of Sonnet and 1/12 the cost of Opus.

"Its (Sonnet 5) performance is close to Opus 4.8, at lower prices." So I ran 4 canvas test through both. > Opus 4.8, 4/4 actually animating. > Sonnet 5, 2/4 came back as static images. And "lower price"? On the paper shredder task, Sonnet 5 spent $0.36 for a static image. Opus 4.8 spent $0.18 and it actually animated. The 4 tests: > Win 98 drag-to-BSOD > Self-typing keyboard CRT > Letter burning > Paper shredder

just dropped huggingface.co/nvidia/Qwen3.…

We’ve received notice that the Department of Commerce has lifted export controls on Claude Fable 5 and Mythos 5. We'll begin restoring access tomorrow, and will share an update soon. We’re grateful to our users for their patience, and to everyone who worked with us on redeploying the models.

Meituan's LongCat-2.0 reportedly lands near GPT-5.5 on SWE-bench. So I threw 5 HTML canvas animation prompts at both. 🥷 Paper sliced fruit-ninja style. 💧 An ink drop diffusing in water. 🔥 A letter burning. 🗑️ Paper crumpling into a ball. ✂️ A strip-cut shredder. Here's how they did 👇

.@grok

Top-tier AI learning resources are free. The roadmap I'd follow: 1. 3Blue1Brown: watch first, build the intuition before any code. > Essence of linear algebra: youtube.com/watch?v=fNk_zzaM… > The essence of calculus: youtube.com/watch?v=WUvTyaaN… > Neural networks: youtube.com/watch?v=aircAruv… 2. Google ML Crash Course: the vocabulary and core concepts, fast. developers.google.com/machin… 3. Karpathy's Zero to Hero (@karpathy): build a GPT from scratch, line by line. youtube.com/watch?v=VMj-3S1t… 4. fast.ai: Practical Deep Learning for Coders course.fast.ai 5. d2l.ai: Dive into Deep Learning, the free textbook you'll keep open forever. 6. Hugging Face LLM Course, ship a real transformer. huggingface.co/huggingface-c… 7. Stanford CS231n / CS224n: go deep, vision and NLP. > CS231n: Deep Learning for Computer Vision youtube.com/playlist?list=PL… > CS224n: Natural Language Processing with Deep Learning youtube.com/playlist?list=PL… Pick one and start tonight.

3 ways to destroy a piece of paper. Qwen 3.5 35B A3B vs. Ornith 1.0 35B, side-by-side canvas test. (Why 3.5 not 3.6? Ornith is post-trained on Qwen 3.5 and Gemma 4, so this shows what the post-training adds.) Same 3 challenges: 🔪 Slice: three blade swipes, fruit-game style 📄 Shredder: desktop strip-cut 🗑️ Crumple: balled up and tossed Winner: not close. Ornith, decisively. The post-training quality is REAL.

How it started How it's going

Introducing a limited preview of GPT-5.6 Sol, our next generation frontier model, as well as GPT-5.6 Terra, a balanced model for efficient, everyday work, and GPT-5.6 Luna, a fast and affordable model for high-volume work. openai.com/index/previewing-…

This test "Pixelate" is now live on BenchLocal. Canvas size is configurable, 8×8 all the way up to 64×64. Install: Settings > Bench Packs > Pixelate, then run it on your own images models. Tested the 😂 emoji with other three frontier models: > Claude Opus 4.8 done in 8 turns > GPT 5.5 done in 10 turns > Gemini 3.5 Flash still nudging 2 pixels at a time at turn 95, so I stopped the test. Go pixelate something!

Memory price hikes have finally started hitting Apple. Today, they raised some base prices: Macbook Neo: $599 -> $699 Macbook Air: $1099 -> $1299 Macbook Pro: $1699 -> $1999 Mac Studio: $1999 -> $2499 iPad Air $599 - $749 iPad Pro: $999 -> $1199

Ugh...