


ANEMLL (pronounced like "animal") Advanced Neural Engine Machine Learning Library, Open Source Project


M6 ANE and GPU performance and power numbers for @ivanfioravanti @liquidai model on CoreAI TLDR: The ANE uses 0.55× the energy of Core AI GPU per decision, and 0.51× that of MLX. It is also the fastest backend on this chip.
d1-2B by @liquidai :Same model, three engines on a MacBook Pro M5 Max - MLX: 66 ms, 153 W - Core AI GPU: 67 ms, 120 W - Core AI Neural Engine: 197 ms, 37 W The Neural Engine is 3x slower (on M5 Max, I bet this is different on M4) and draws a quarter of the power. Same answers on all three. I had to use a video instead of live camera feed because ANE usage goes up as soon as you start the camera. Core AI export from the original weights, plus live GPU / Neural Engine / power charts: github.com/ivanfioravanti/ml… Download and have fun! Share your stats on other Apple Silicon machines! I'm now creating an iPhone version 😎


👀
Introducing Qwen3.8-27B-pi-adapted 🤗 We adapted Qwen's model specifically for the pi coding agent using reinforcement learning and a LoRA adapter. The results, measured as pass@1: 🚀 SWE-bench Pro (Hard 51): 29.4% → 62.7% 💻 Terminal-Bench 2.1: 39.5% → 55.8% ⚡ 55–62% fewer tool calls per task This is the power of open models: you can adapt them to YOUR tools and YOUR use cases. Let’s build! 💛 Open weights: huggingface.co/harness-hub/Q… Methodology: huggingface.co/blog/harness-…

It’s here. Let the race for the best inference engine begin 🚀🚀🚀
Token efficiency and cost: Qwen4-Max outperforms Qwen3.8-Max with fewer tokens. On Work360, token usage drops by approximately 24% (Xhigh) and 35% (High); on Terminal-Bench 4.0, by about 48% and 49%. In our Work360 evaluation, the average cost per task drops from $4.43 to $1.67.

📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery! Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window. Highlights: - Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation. - Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables. - Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery. - $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching — all lower than Qwen3.8-Max, with implicit caching down by 70%. The API is live on QwenCloud. You can also try Qwen4-Max on Qwen Studio, QwenWork and Qoder. Start building today! 🚀 - Blog: qwen.ai/blog?id=qwen4-max - API: qwencloud.com/models/qwen4-m… - Token Plan: qwencloud.com/pricing/token-… More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀

It begins
📢 Meet Qwen4-Max, your new AI teammate for work, creativity, and discovery! Qwen4-Max is built on a new MoE architecture with native vision and a 1M-token context window. Highlights: - Outperforms Qwen3.8-Max with up to ~50% fewer tokens, while average cost per task drops over 60% in our internal evaluation. - Advances real-world productivity through greater proficiency with professional tools, stronger long-horizon execution, and higher-quality deliverables. - Handles real tasks, including 3D modeling, game development, manufacturing, finance, law, healthcare, and frontier scientific discovery. - $1.5/M tokens for input, $4.5/M tokens for output, and $0.075/M tokens for implicit caching — all lower than Qwen3.8-Max, with implicit caching down by 70%. The API is live on QwenCloud. You can also try Qwen4-Max on Qwen Studio, QwenWork and Qoder. Start building today! 🚀 - Blog: qwen.ai/blog?id=qwen4-max - API: qwencloud.com/models/qwen4-m… - Token Plan: qwencloud.com/pricing/token-… More to come! The open weights of smaller Qwen4 models will be released. Stay tuned. 👀

To add : one problematic area for macOS with streaming weights is RAM compression and cache eviction. MacOS will aggressively compress cached weights slowing down agentic workflows (eg during short tool calls). So both OS compression state watchdog and careful memory management are required.


Back to SoCal for the winter, SpaceX @USSpaceForce launch lit up night skies tonight 😍

🚀🚀🚀
Apple's Core AI models repo is now accepting third-party contributions 🥳 If you have developed some cool ports, this is the best place to share your recipes with the community. Looking at you @JackdeS11 @mattt @kevinqz 😂 github.com/apple/coreai-mode…

👀
We're releasing new Unsloth Dynamic v3.1 GGUFs for Qwen3.8-Flash-Next! They're better by 10-15% accuracy on DeepSWE, Terminal Bench, Aider Polyglot and more vs GSQ-RCO. We also provide a new engram switching guide for BF16, Q8_0, Q5_1, IQ4_NL! This builds upon our DeepSeek-V4.1-Flash UD3.1 work, which improved over 10-20% accuracy vs others. All benchmarks are 256K for agentic ones, and 128K for the rest. We used temperature 1.0, top_p 0.95, top_k 20, min_p 0, no presence or repetition penalty. preserve_thinking and xhigh reasoning, and ran over 26B of tokens for benchmarking. The RMSE vs BF16 for each of the engrams: BF16 102.4 GB 0 Q8_0 54.4 GB 0.0053 Q5_1 38.4 GB 0.0376 IQ4_NL 28.8 GB 0.0757 More details at unsloth.ai/docs/models/qwen3…. We also moved all benchmarking, grading, infra issues and more to unsloth.ai/docs/basics/dynam…

Open source, iOS app included: github.com/Anemll/system1-an… PS video is created from Duo screen capture using ChatGPT 6.1 Terra and voice command

System 1 ANE now supports iPhone Duo! My first attempt at adapting a simple on-device AI app to the Duo layout: Jeff (0.8B) playing Tetris on the Neural Engine, nothing in the cloud. Opt in and your games become training data for better System 1 models. Only the moves are shared, no account or device ID. Same TestFlight, just update: testflight.apple.com/join/4M…


Good to see KLD defense across multiple datasets 👍
We’re releasing new DeepSeek-V4.1-Flash-1017 GGUFs, ranking highest across benchmarks. Dynamic V3.1 leads across nearly all benchmarks: GPQA Diamond, AIME25, LiveCodeBench v6, MMLU-Pro, Aider & more. Run on 64GB+ RAM. GGUF: huggingface.co/unsloth/DeepS… Blog: unsloth.ai/docs/models/deeps…