Загрузка видео...

Не удалось загрузить видео

На главную

Let's go! 😍 Qwen just released Qwen3-Embedding, a new series of embedding models: 🏆 SOTA performance on MMTEB, MTEB, and MTEB-Code 📏 Three different sizes (0.6B / 4B / 8B) 🌍 Multilingual (119 languages) 💻 Can run in-browser w/ Transformers.js (+ WebGPU acceleration)

xenovacom's profile picture

Xenova

13,051 subscribers

47,143 просмотров • 1 год назад •via X (Twitter)

Новости и политикаНаука и технологииОбразование
LIVE

Anya Rossi• Live Now

Private livecam show

Комментарии: 3

Фото профиля Xenova
Xenova1 год назад

These models can be used for document retrieval, RAG, classification, clustering, sentiment analysis, code search, and much more! 🚀 Links: - Collection: - ONNX weights: - Demo:

Фото профиля Rainmaker
Rainmaker2 лет назад

Here I share an XGBoost model that delivers a 25% CAGR with minimal drawdown on Visa stock. In this free Substack post I share code and commentary for a powerful Machine Learning strategy that delivers powerful returns.

Фото профиля ander
ander1 год назад

@Alibaba_Qwen Does it output embeddings of sentences? Of the whole text? @grok @AskPerplexity

Похожие видео

NEW: Alibaba just released Qwen 3.5 Small — a family of powerful multimodal models available in a range of sizes (0.8B, 2B, 4B, and 9B parameters). Perfect for on-device applications! They can even run 100% locally in your browser on WebGPU, powered by Transformers.js! 🤯
0:46

NEW: Alibaba just released Qwen 3.5 Small — a family of powerful multimodal models available in a range of sizes (0.8B, 2B, 4B, and 9B parameters). Perfect for on-device applications! They can even run 100% locally in your browser on WebGPU, powered by Transformers.js! 🤯

Xenova

102,592 просмотров • 5 месяцев назад

NEW: Perplexity just released pplx-embed, a series of powerful embedding models. Inspired by their logo, I built a small semantic search demo: enter a query and the model retrieves the most relevant page. Everything runs locally in your browser with Transformers.js and Three.js
0:40

NEW: Perplexity just released pplx-embed, a series of powerful embedding models. Inspired by their logo, I built a small semantic search demo: enter a query and the model retrieves the most relevant page. Everything runs locally in your browser with Transformers.js and Three.js

Xenova

11,092 просмотров • 5 месяцев назад

IBM just released Granite 4.0, their latest series of small language models! These models excel at agentic workflows (tool calling), document analysis, RAG, and more. 🚀 The "Micro" (3.4B) model can even run 100% locally in your browser on WebGPU, powered by 🤗 Transformers.js!
0:34

IBM just released Granite 4.0, their latest series of small language models! These models excel at agentic workflows (tool calling), document analysis, RAG, and more. 🚀 The "Micro" (3.4B) model can even run 100% locally in your browser on WebGPU, powered by 🤗 Transformers.js!

Xenova

82,893 просмотров • 10 месяцев назад

Google just proved that bigger isn't always better. Their 308M parameter model is outperforming models 2x its size. Google just released 𝗘𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝗚𝗲𝗺𝗺𝗮, and it's proving that lightweight embedding models can punch way above their weight class. At just 308M parameters (578MB), it's the new state-of-the-art for models under 500M parameters across MTEB multilingual, English, and code benchmarks. But the really impressive part is that it ranks 8th overall on MTEB(Multilingual, v2) - that's 𝟭𝟳 𝗽𝗹𝗮𝗰𝗲𝘀 above the second-best sub-500M model, and it's delivering performance 𝗰𝗼𝗺𝗽𝗮𝗿𝗮𝗯𝗹𝗲 𝘁𝗼 𝗺𝗼𝗱𝗲𝗹𝘀 𝗻𝗲𝗮𝗿𝗹𝘆 𝗱𝗼𝘂𝗯𝗹𝗲 𝗶𝘁𝘀 𝘀𝗶𝘇𝗲. There are three key parts of their training recipe that sets it apart: 𝟭. 𝗘𝗻𝗰𝗼𝗱𝗲𝗿-𝗗𝗲𝗰𝗼𝗱𝗲𝗿 𝗜𝗻𝗶𝘁𝗶𝗮𝗹𝗶𝘇𝗮𝘁𝗶𝗼𝗻 Instead of starting from a decoder-only Gemma 3 model, they first adapted it to encoder-decoder, then used just the encoder. By basing EmbeddingGemma off an LLM that already has world and language understanding, it gives it a stronger starting point. 𝟮. 𝗧𝗵𝗿𝗲𝗲-𝗟𝗼𝘀𝘀 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 They combine three different loss functions, instead of just having one: • Contrastive loss (NCE) with in-batch negatives and hardness weighting • Spread-out regularization to ensure embeddings utilize the full space (for quantization and ANN retrieval) • Embedding matching distillation from Gemini Embedding - not just learning from relevance scores, but directly aligning the embedding space with the teacher model 𝟯. 𝗠𝗼𝗱𝗲𝗹 𝗦𝗼𝘂𝗽𝗶𝗻𝗴 Rather than just averaging checkpoints from the same training run, they use optimization techniques to find multiple specialized training mixtures. Each mixture creates an "expert" model in different domains, and averaging all their parameters creates a final model that's actually better than individual models. Extras: • Matryoshka embeddings supporting 768, 512, 256, and 128 dimensions • Quantization-aware training - maintains quality even at int4 precision • 100+ languages from Gemma 3 pretraining • Exceptional performance on low-resource languages (check their XTREME-UP results) Is it the absolute best embedding model? No - Gemini Embedding still leads overall. But that's not really the point. EmbeddingGemma proves you can achieve state-of-the-art performance in a small package that's actually deployable on-device, in low-latency applications, and in resource-constrained environments. This makes good embeddings accessible for use cases that I'm seeing more and more: offline applications, privacy-sensitive deployments, and high-throughput scenarios where inference cost actually matters. Full paper: Shoutout to the EmbeddingGemma team at Google DeepMind for this awesome open source work 💙 and to Daniel Williams for helping me with this video! 🫶
5:14

Google just proved that bigger isn't always better. Their 308M parameter model is outperforming models 2x its size. Google just released 𝗘𝗺𝗯𝗲𝗱𝗱𝗶𝗻𝗴𝗚𝗲𝗺𝗺𝗮, and it's proving that lightweight embedding models can punch way above their weight class. At just 308M parameters (578MB), it's the new state-of-the-art for models under 500M parameters across MTEB multilingual, English, and code benchmarks. But the really impressive part is that it ranks 8th overall on MTEB(Multilingual, v2) - that's 𝟭𝟳 𝗽𝗹𝗮𝗰𝗲𝘀 above the second-best sub-500M model, and it's delivering performance 𝗰𝗼𝗺𝗽𝗮𝗿𝗮𝗯𝗹𝗲 𝘁𝗼 𝗺𝗼𝗱𝗲𝗹𝘀 𝗻𝗲𝗮𝗿𝗹𝘆 𝗱𝗼𝘂𝗯𝗹𝗲 𝗶𝘁𝘀 𝘀𝗶𝘇𝗲. There are three key parts of their training recipe that sets it apart: 𝟭. 𝗘𝗻𝗰𝗼𝗱𝗲𝗿-𝗗𝗲𝗰𝗼𝗱𝗲𝗿 𝗜𝗻𝗶𝘁𝗶𝗮𝗹𝗶𝘇𝗮𝘁𝗶𝗼𝗻 Instead of starting from a decoder-only Gemma 3 model, they first adapted it to encoder-decoder, then used just the encoder. By basing EmbeddingGemma off an LLM that already has world and language understanding, it gives it a stronger starting point. 𝟮. 𝗧𝗵𝗿𝗲𝗲-𝗟𝗼𝘀𝘀 𝗧𝗿𝗮𝗶𝗻𝗶𝗻𝗴 They combine three different loss functions, instead of just having one: • Contrastive loss (NCE) with in-batch negatives and hardness weighting • Spread-out regularization to ensure embeddings utilize the full space (for quantization and ANN retrieval) • Embedding matching distillation from Gemini Embedding - not just learning from relevance scores, but directly aligning the embedding space with the teacher model 𝟯. 𝗠𝗼𝗱𝗲𝗹 𝗦𝗼𝘂𝗽𝗶𝗻𝗴 Rather than just averaging checkpoints from the same training run, they use optimization techniques to find multiple specialized training mixtures. Each mixture creates an "expert" model in different domains, and averaging all their parameters creates a final model that's actually better than individual models. Extras: • Matryoshka embeddings supporting 768, 512, 256, and 128 dimensions • Quantization-aware training - maintains quality even at int4 precision • 100+ languages from Gemma 3 pretraining • Exceptional performance on low-resource languages (check their XTREME-UP results) Is it the absolute best embedding model? No - Gemini Embedding still leads overall. But that's not really the point. EmbeddingGemma proves you can achieve state-of-the-art performance in a small package that's actually deployable on-device, in low-latency applications, and in resource-constrained environments. This makes good embeddings accessible for use cases that I'm seeing more and more: offline applications, privacy-sensitive deployments, and high-throughput scenarios where inference cost actually matters. Full paper: Shoutout to the EmbeddingGemma team at Google DeepMind for this awesome open source work 💙 and to Daniel Williams for helping me with this video! 🫶

Victoria Slocum

21,610 просмотров • 8 месяцев назад

NEW: Mistral AI releases Mistral 3, a family of multimodal models, including three start-of-the-art dense models (3B, 8B, and 14B) and Mistral Large 3 (675B, 41B active). All Apache 2.0! 🤗 Surprisingly, the 3B is small enough to run 100% locally in your browser on WebGPU! 🤯
0:53

NEW: Mistral AI releases Mistral 3, a family of multimodal models, including three start-of-the-art dense models (3B, 8B, and 14B) and Mistral Large 3 (675B, 41B active). All Apache 2.0! 🤗 Surprisingly, the 3B is small enough to run 100% locally in your browser on WebGPU! 🤯

Xenova

225,178 просмотров • 8 месяцев назад

RF-DETR, the state-of-the-art model series for real-time object detection, can now run 100% locally in your browser on WebGPU with 🤗 Transformers.js v4! The models are Apache-2.0 licensed, making them a perfect fit for both personal and commercial applications. Try the demo 👇
0:42

RF-DETR, the state-of-the-art model series for real-time object detection, can now run 100% locally in your browser on WebGPU with 🤗 Transformers.js v4! The models are Apache-2.0 licensed, making them a perfect fit for both personal and commercial applications. Try the demo 👇

Xenova

77,147 просмотров • 5 месяцев назад

Text-to-speech is moving way too fast. Just a few days ago, I tweeted about PersonaPlex-7B, NVIDIA's new open source TTS ( And today, Qwen just open-sourced Qwen3-TTS 🤯 It’s a revolutionary text-to-speech model built for control. Not just about generating speech, but about shaping how it sounds directly from language. You can guide the pace, the tone, and the expressiveness straight from text, without touching audio graphs or hand-tuning parameters. That’s the real shift! What makes Qwen3-TTS stand out is how practical it already is: → voice cloning from just a few seconds of audio → voice creation without any reference sample → support for 10 languages out of the box → end-to-end latency down to ~97ms → works in both streaming and non-streaming setups The models come in two sizes (0.6B and 1.7B), so you can trade off quality and hardware cost depending on your setup. You can work with curated voices, designed voices, or cloned ones, and it integrates cleanly with vLLM for production use. It also ships as a simple Python package you can pip install. If you’re building real-time voice systems, this removes a lot of friction! 100% free and open source. I put the repo in the 🧵↓
0:41

Text-to-speech is moving way too fast. Just a few days ago, I tweeted about PersonaPlex-7B, NVIDIA's new open source TTS ( And today, Qwen just open-sourced Qwen3-TTS 🤯 It’s a revolutionary text-to-speech model built for control. Not just about generating speech, but about shaping how it sounds directly from language. You can guide the pace, the tone, and the expressiveness straight from text, without touching audio graphs or hand-tuning parameters. That’s the real shift! What makes Qwen3-TTS stand out is how practical it already is: → voice cloning from just a few seconds of audio → voice creation without any reference sample → support for 10 languages out of the box → end-to-end latency down to ~97ms → works in both streaming and non-streaming setups The models come in two sizes (0.6B and 1.7B), so you can trade off quality and hardware cost depending on your setup. You can work with curated voices, designed voices, or cloned ones, and it integrates cleanly with vLLM for production use. It also ships as a simple Python package you can pip install. If you’re building real-time voice systems, this removes a lot of friction! 100% free and open source. I put the repo in the 🧵↓

Charly Wargnier

59,144 просмотров • 6 месяцев назад

Locked and loaded 💥 Openwork is now Accomplish v0.3.4 released today (download from link in bio): - Following the announcement from Kimi.ai a couple of hours ago, we are happy to introduce support for Kimi K2.5 with a new Moonshot AI integration! This is a global SOTA model for CUA BrowseComp (74.9%) - Log in with your OpenAI account (using OAuth) instead of providing an API key! This is both more seurce and it allows you to leverage existing OpenAI plan/subscription. - Support for Microsoft Azure AI Foundry thanks to our friends from Microsoft (cheers Ido Frizler!) - Support for LM Studio to run models locally (cheers to Yossi Elkrief LM Studiodevs!) . Try qwen/qwen3-vl-8b! - Support for MiniMax (official), the awesome open source models we love (thanks marcho) - You can now voice chat with openwork, courtesy of ! Thank you Max (Zihan) ZENG! Watch Openwork is now Accomplish one-shotting SOC 2 Compliance gap analysis (scanning your GitHub, Jira, Notion, websites and more):
1:44

Locked and loaded 💥 Openwork is now Accomplish v0.3.4 released today (download from link in bio): - Following the announcement from Kimi.ai a couple of hours ago, we are happy to introduce support for Kimi K2.5 with a new Moonshot AI integration! This is a global SOTA model for CUA BrowseComp (74.9%) - Log in with your OpenAI account (using OAuth) instead of providing an API key! This is both more seurce and it allows you to leverage existing OpenAI plan/subscription. - Support for Microsoft Azure AI Foundry thanks to our friends from Microsoft (cheers Ido Frizler!) - Support for LM Studio to run models locally (cheers to Yossi Elkrief LM Studiodevs!) . Try qwen/qwen3-vl-8b! - Support for MiniMax (official), the awesome open source models we love (thanks marcho) - You can now voice chat with openwork, courtesy of ! Thank you Max (Zihan) ZENG! Watch Openwork is now Accomplish one-shotting SOC 2 Compliance gap analysis (scanning your GitHub, Jira, Notion, websites and more):

Or Hiltch

33,932 просмотров • 6 месяцев назад

Introducing NexaSDK for Android (Beta) — run the latest AI models locally, 9× more energy-efficient and 2× faster, on Android devices, powered by the Qualcomm Hexagon NPU. This is the first SDK to support NPU, GPU and CPU, unlocking the full power of every Android device — for example, LFM2-1.2B achieves 85 t/s on NPU vs 37 t/s on CPU. With just 3 lines of code, you can run the latest state-of-the-art models across every task, for example: Multimodal (Vision, Audio, Text): OmniNeural-4B Embedding: EmbeddingGemma from Google ASR: Parakeet-v3 from NVIDIA OCR: PaddleOCR from Baidu Inc. Rerank: Jina-reranker from Jina AI LLM: LFM2-1.2B from And we continue to deliver Day-0 model support across our framework. Build on-device AI into your Android app today and enjoy no cloud API cost, full privacy, and offline availability. Check out our Quickstart guide and example app below to get started in minutes.
0:54

Introducing NexaSDK for Android (Beta) — run the latest AI models locally, 9× more energy-efficient and 2× faster, on Android devices, powered by the Qualcomm Hexagon NPU. This is the first SDK to support NPU, GPU and CPU, unlocking the full power of every Android device — for example, LFM2-1.2B achieves 85 t/s on NPU vs 37 t/s on CPU. With just 3 lines of code, you can run the latest state-of-the-art models across every task, for example: Multimodal (Vision, Audio, Text): OmniNeural-4B Embedding: EmbeddingGemma from Google ASR: Parakeet-v3 from NVIDIA OCR: PaddleOCR from Baidu Inc. Rerank: Jina-reranker from Jina AI LLM: LFM2-1.2B from And we continue to deliver Day-0 model support across our framework. Build on-device AI into your Android app today and enjoy no cloud API cost, full privacy, and offline availability. Check out our Quickstart guide and example app below to get started in minutes.

NEXA AI

26,091 просмотров • 8 месяцев назад

new-day-new-hack so last time i dropped nocturne, a verifiable AI trading agent on hyperliquid and many people said "just run it for me". this time, we go one level higher. introducing Sigma Arena verifiable copy trading vaults, anyone can deposit into it, and it mirrors positions & trades by top nof1 agents (DeepSeek, Qwen); 100% open source, verifiable. i'm starting with $10K deposit limit on each vault, will lift if as the vault gets more traction and good performance. execution verified by eigencloud SigmaArena [dot] com (deposit at your own risk, NFA) code in replies + info in the video
2:03

new-day-new-hack so last time i dropped nocturne, a verifiable AI trading agent on hyperliquid and many people said "just run it for me". this time, we go one level higher. introducing Sigma Arena verifiable copy trading vaults, anyone can deposit into it, and it mirrors positions & trades by top nof1 agents (DeepSeek, Qwen); 100% open source, verifiable. i'm starting with $10K deposit limit on each vault, will lift if as the vault gets more traction and good performance. execution verified by eigencloud SigmaArena [dot] com (deposit at your own risk, NFA) code in replies + info in the video

Gajesh

85,529 просмотров • 9 месяцев назад

NVIDIA just released a new open source transcription model, Nemotron Speech ASR, designed from the ground up for low-latency use cases like voice agents. Here's a voice agent built with this new model. 24ms transcription finalization and total voice-to-voice inference time under 500ms. This agent actually uses *three* NVIDIA open source models: - Nemotron Speech ASR - Nemotron 3 Nano 30GB in a 4-bit quant (released in December) - A preview checkpoint of the upcoming Magpie text-to-speech model These models are all truly open source: weights, training data, training code, and inference code. This is a big deal! Jensen said in the CES keynote yesterday that he expects open source models to catch up to proprietary models this year in a number of categories. NVIDIA is putting their weight behind making this happen. (As Alan Kay said, the best way to predict the future is to invent it.) The code for this agent is open source too, of course. You can deploy it to production with Modal and Pipecat AI cloud, or run locally on an NVIDIA DGX Spark or RTX 5090.
1:24

NVIDIA just released a new open source transcription model, Nemotron Speech ASR, designed from the ground up for low-latency use cases like voice agents. Here's a voice agent built with this new model. 24ms transcription finalization and total voice-to-voice inference time under 500ms. This agent actually uses *three* NVIDIA open source models: - Nemotron Speech ASR - Nemotron 3 Nano 30GB in a 4-bit quant (released in December) - A preview checkpoint of the upcoming Magpie text-to-speech model These models are all truly open source: weights, training data, training code, and inference code. This is a big deal! Jensen said in the CES keynote yesterday that he expects open source models to catch up to proprietary models this year in a number of categories. NVIDIA is putting their weight behind making this happen. (As Alan Kay said, the best way to predict the future is to invent it.) The code for this agent is open source too, of course. You can deploy it to production with Modal and Pipecat AI cloud, or run locally on an NVIDIA DGX Spark or RTX 5090.

kwindla

274,562 просмотров • 6 месяцев назад

Two new #threejs projects on the way 😍 Working on my own gem renderer with a custom refraction material in TSL for #WebGPU, combining Screen Space Reflections + SSGTAO for superior results, as this allows diamonds to generate reflections on the metal mesh. The performance is incredible, even on gems with over 200 diamonds. This time, without any additional libraries. Just React Three Fiber and a lot of magic with custom shaders in TSL. GemGPU™ will also come with a very minimalist gem and metal customization panel that hides and lets the jewelry be the star. I also created a beautiful new stats panel that completely hides in the scene when not needed. It provides relevant optimization information, a performance gauge, summary and expanded modes, and day/night toggle. In the video, you can also see SSGI in action, with a pearl that reflects, in real time, all the objects in the scene, just as the objects in the scene reflect the pearl. It's magical. There are still a few things to fix, but I'll share the live link soon. Stay tuned.
3:43

Two new #threejs projects on the way 😍 Working on my own gem renderer with a custom refraction material in TSL for #WebGPU, combining Screen Space Reflections + SSGTAO for superior results, as this allows diamonds to generate reflections on the metal mesh. The performance is incredible, even on gems with over 200 diamonds. This time, without any additional libraries. Just React Three Fiber and a lot of magic with custom shaders in TSL. GemGPU™ will also come with a very minimalist gem and metal customization panel that hides and lets the jewelry be the star. I also created a beautiful new stats panel that completely hides in the scene when not needed. It provides relevant optimization information, a performance gauge, summary and expanded modes, and day/night toggle. In the video, you can also see SSGI in action, with a pearl that reflects, in real time, all the objects in the scene, just as the objects in the scene reflect the pearl. It's magical. There are still a few things to fix, but I'll share the live link soon. Stay tuned.

Anderson Mancini

12,113 просмотров • 5 месяцев назад

Alibaba just released a coding model that hits 82 percent on SWE-Bench Verified. That is the highest score ever published for an open-source model. The weights are free. The license is Apache 2.0. You can run it today. The model is Qwen 4 Coder 32B. Here is what 82 percent on SWE-Bench Verified actually means. SWE-Bench Verified tests whether an AI can autonomously resolve real bugs pulled from real production GitHub repositories. Not synthetic exercises. Real open-source projects that real teams depend on. A model gets a bug report, reads the code, writes a fix, and either passes the test suite or it does not. At 82 percent, Qwen 4 Coder 32B resolves 82 out of every 100 real production bugs it is given. Without a human guiding it. On code it has never seen before. For comparison: Qwen 4 Coder 32B: 82 percent SWE-Bench Verified. Open source. Apache 2.0. Claude Fable 5: 80.3 percent SWE-Bench Pro. $10 input / $50 output per million tokens. Currently suspended. GPT-5.6 Sol: Competitive on Terminal-Bench. $5 input / $30 output per million tokens. An open-weight model that you can download and run for free just beat both of them on the benchmark designed to measure real software engineering capability. Here is the architecture. Qwen 4 Coder 32B is a 32 billion parameter dense model. Not a Mixture-of-Experts. Every parameter is active on every request. This matters for inference: a dense 32B model runs on 22 gigabytes of VRAM, which fits on a single high-end consumer GPU or a MacBook Pro with 64GB of unified memory. The smaller variant, Qwen 4 Coder 4B, runs at approximately 135 tokens per second on an M5 Max and fits inside 8 gigabytes of RAM. For a model with usable coding capability, that is a new bar for what fits in a single laptop. The training methodology continued Alibaba's approach of reinforcement learning on verifiable coding tasks. The model gets rewarded when its code passes tests. It gets penalized when it fails. Over millions of training steps, the model learns to write code that actually runs rather than code that looks plausible. License: Apache 2.0. Full commercial use. No attribution requirement. No revenue threshold. No monthly active user ceiling. Weights: Hugging Face, available today. Runs on: vLLM, Ollama, SGLang, and any standard GGUF-compatible inference engine. Qwen 4 32B also runs at approximately 135 tokens per second on an M5 Max chip, setting a new bar for what a sub-8GB model can do on Apple Silicon. The open-source coding model just beat the best closed-source model in the world on the benchmark designed to test whether AI can actually do software engineering. The weights are free. The subscription is optional. Source: Autom8Labs AI Insight July 2026, State of Open Source LLMs June 2026, Kunal Ganglani blog June 2026.
0:25

Alibaba just released a coding model that hits 82 percent on SWE-Bench Verified. That is the highest score ever published for an open-source model. The weights are free. The license is Apache 2.0. You can run it today. The model is Qwen 4 Coder 32B. Here is what 82 percent on SWE-Bench Verified actually means. SWE-Bench Verified tests whether an AI can autonomously resolve real bugs pulled from real production GitHub repositories. Not synthetic exercises. Real open-source projects that real teams depend on. A model gets a bug report, reads the code, writes a fix, and either passes the test suite or it does not. At 82 percent, Qwen 4 Coder 32B resolves 82 out of every 100 real production bugs it is given. Without a human guiding it. On code it has never seen before. For comparison: Qwen 4 Coder 32B: 82 percent SWE-Bench Verified. Open source. Apache 2.0. Claude Fable 5: 80.3 percent SWE-Bench Pro. $10 input / $50 output per million tokens. Currently suspended. GPT-5.6 Sol: Competitive on Terminal-Bench. $5 input / $30 output per million tokens. An open-weight model that you can download and run for free just beat both of them on the benchmark designed to measure real software engineering capability. Here is the architecture. Qwen 4 Coder 32B is a 32 billion parameter dense model. Not a Mixture-of-Experts. Every parameter is active on every request. This matters for inference: a dense 32B model runs on 22 gigabytes of VRAM, which fits on a single high-end consumer GPU or a MacBook Pro with 64GB of unified memory. The smaller variant, Qwen 4 Coder 4B, runs at approximately 135 tokens per second on an M5 Max and fits inside 8 gigabytes of RAM. For a model with usable coding capability, that is a new bar for what fits in a single laptop. The training methodology continued Alibaba's approach of reinforcement learning on verifiable coding tasks. The model gets rewarded when its code passes tests. It gets penalized when it fails. Over millions of training steps, the model learns to write code that actually runs rather than code that looks plausible. License: Apache 2.0. Full commercial use. No attribution requirement. No revenue threshold. No monthly active user ceiling. Weights: Hugging Face, available today. Runs on: vLLM, Ollama, SGLang, and any standard GGUF-compatible inference engine. Qwen 4 32B also runs at approximately 135 tokens per second on an M5 Max chip, setting a new bar for what a sub-8GB model can do on Apple Silicon. The open-source coding model just beat the best closed-source model in the world on the benchmark designed to test whether AI can actually do software engineering. The weights are free. The subscription is optional. Source: Autom8Labs AI Insight July 2026, State of Open Source LLMs June 2026, Kunal Ganglani blog June 2026.

Harman

41,278 просмотров • 25 дней назад

I'm up late with the rest of you building AI agents with the new AI browser from Genspark. We can see where this is all going: a new kind of operating system -- one that is very different than the Microsoft centric way that I've been working for 20 years. There are several things that these new agentic browsers bring to you: 1. They let you change how you browse. With an old browser like Google Chrome, you go to your email, Facebook, or X. 2. With these new browsers, you tell it where to go and what to do for you. 3. It can even build software for you. At the end of this video, I have it building me a little YouTube uploading utility, which is very helpful. 4. They have a ton of "applications" built in. Think of it as a new kind of office suite. Docs. Spreadsheets. Slide decks. And much more. All built with AI, not bolted on the side like with Microsoft's Office. 5. They have AI models built "underneath" so you can work privately and cheaply. There’s a lot of new choices you have to make with browsers like this. I’ve been playing with a bunch of them. Some have better user interfaces than others. Some have different versions, slide components, or applications. The reason I like Genspark is because they ship so fast. I’ve been watching this company since its very beginnings, and every week they ship new things. Just yesterday, they shipped a new photo editing feature for my iPhone. I upload a photo and then I can just talk to it and edit it with my voice. It's really cool. I try to reward companies that ship at such a fast rate and that are shipping innovation that improves our lives. It's not that I'm going to stop using Google Chrome. My whole life has been there for, I don't know, almost 20 years now. This is a different way of working and it gives me a space to run my AI tasks that's different than Google Chrome. I run them side by side. One doing old stuff, one doing new stuff. I can keep using Google Chrome for my old stuff, like my email and my calendar. And I use GenSpark or one of the new AI browsers to do new AI-centric things. All sorts of new things that these new agentic browsers open up! Have you tried it, or one of the other new ones yet? How has it changed your work? It takes a little time to get used to AI-centric ways of doing things. Pretend your browser is a team of interns. Give them a task, in this case I said "help me upload my videos to YouTube." You might be shocked at what Genspark does to improve your life. I am everytime I use it. Give it a try and let me know what you think! Oh, and I used another little tool to "write" this post. Typeless -- I push a button and talk and it writes. With fewer typos than I usually type in, to boot. It works great with Genspark's new browser too. Download it here:
3:41

I'm up late with the rest of you building AI agents with the new AI browser from Genspark. We can see where this is all going: a new kind of operating system -- one that is very different than the Microsoft centric way that I've been working for 20 years. There are several things that these new agentic browsers bring to you: 1. They let you change how you browse. With an old browser like Google Chrome, you go to your email, Facebook, or X. 2. With these new browsers, you tell it where to go and what to do for you. 3. It can even build software for you. At the end of this video, I have it building me a little YouTube uploading utility, which is very helpful. 4. They have a ton of "applications" built in. Think of it as a new kind of office suite. Docs. Spreadsheets. Slide decks. And much more. All built with AI, not bolted on the side like with Microsoft's Office. 5. They have AI models built "underneath" so you can work privately and cheaply. There’s a lot of new choices you have to make with browsers like this. I’ve been playing with a bunch of them. Some have better user interfaces than others. Some have different versions, slide components, or applications. The reason I like Genspark is because they ship so fast. I’ve been watching this company since its very beginnings, and every week they ship new things. Just yesterday, they shipped a new photo editing feature for my iPhone. I upload a photo and then I can just talk to it and edit it with my voice. It's really cool. I try to reward companies that ship at such a fast rate and that are shipping innovation that improves our lives. It's not that I'm going to stop using Google Chrome. My whole life has been there for, I don't know, almost 20 years now. This is a different way of working and it gives me a space to run my AI tasks that's different than Google Chrome. I run them side by side. One doing old stuff, one doing new stuff. I can keep using Google Chrome for my old stuff, like my email and my calendar. And I use GenSpark or one of the new AI browsers to do new AI-centric things. All sorts of new things that these new agentic browsers open up! Have you tried it, or one of the other new ones yet? How has it changed your work? It takes a little time to get used to AI-centric ways of doing things. Pretend your browser is a team of interns. Give them a task, in this case I said "help me upload my videos to YouTube." You might be shocked at what Genspark does to improve your life. I am everytime I use it. Give it a try and let me know what you think! Oh, and I used another little tool to "write" this post. Typeless -- I push a button and talk and it writes. With fewer typos than I usually type in, to boot. It works great with Genspark's new browser too. Download it here:

Robert Scoble

70,991 просмотров • 10 месяцев назад

‘Gargantuan’ Turn In US Small Business Data 📊🚀 The pulse of the U.S. economy just delivered a jolt of optimism. NFIB Small Business data released on Tuesday surged to its highest level in more than three years, marking a dramatic turnaround that has analysts and investors taking notice. In this clip from "The Macro Show," Hedgeye CEO Keith McCullough breaks down the implications of this significant move and its potential to fuel an extended market bull run. "This is a very good leading indicator when you come out of the bottom," McCullough explains. "This is a gargantuan turn in small business. I can only go back so many times in my career and find something like this." * * * Get a free copy of Keith's new 52-page eBook "Master The Market: A Hedge Fund Manager's Guide to Process and Profit":
4:17

‘Gargantuan’ Turn In US Small Business Data 📊🚀 The pulse of the U.S. economy just delivered a jolt of optimism. NFIB Small Business data released on Tuesday surged to its highest level in more than three years, marking a dramatic turnaround that has analysts and investors taking notice. In this clip from "The Macro Show," Hedgeye CEO Keith McCullough breaks down the implications of this significant move and its potential to fuel an extended market bull run. "This is a very good leading indicator when you come out of the bottom," McCullough explains. "This is a gargantuan turn in small business. I can only go back so many times in my career and find something like this." * * * Get a free copy of Keith's new 52-page eBook "Master The Market: A Hedge Fund Manager's Guide to Process and Profit":

Hedgeye

102,594 просмотров • 1 год назад

More moves to AI-first living. Got the Genspark browser that just came out. One thing I notice is that I like having separate browsers working on different tasks. I'm not a big user of MCPs, since I'm not really a developer. Are you? They have a store of a bunch of MCPs available to use. Most of my AI work is research related, or answering emails and scheduling things, I'm focusing more of my effort there. Genspark has been shipping a ton of stuff the past few weeks, from ability to make slides to a new AI secretary that can schedule things, or help you answer your emails. First I looked at, "does it make sense to move over to this new way of working?" The problem with that is you gotta learn a new browser. Well, it looks like Google Chrome, except it has a prompt window. Click the icons underneath it and you will see you can add X, Notion, Google stuff like calendar and gmail, or a variety of others. But the real power is just talking to it. I asked it what kinds of tasks it can do, and it answered: "Just tell me what you want to accomplish! For example: "Research the latest developments in spatial computing and create a presentation" "Analyze my Gmail for any important emails from this week" "Find information about AI robotics companies and create a spreadsheet" "Generate a video about emerging tech trends" "Help me plan a trip to CES 2026" And so further into AI-first living I go. Have you tried it yet? What do you suggest I do next to use more AI to run my life and build my business? I have another browser tab building a video for me. What a beautiful time to be alive. Download it at:
1:09

More moves to AI-first living. Got the Genspark browser that just came out. One thing I notice is that I like having separate browsers working on different tasks. I'm not a big user of MCPs, since I'm not really a developer. Are you? They have a store of a bunch of MCPs available to use. Most of my AI work is research related, or answering emails and scheduling things, I'm focusing more of my effort there. Genspark has been shipping a ton of stuff the past few weeks, from ability to make slides to a new AI secretary that can schedule things, or help you answer your emails. First I looked at, "does it make sense to move over to this new way of working?" The problem with that is you gotta learn a new browser. Well, it looks like Google Chrome, except it has a prompt window. Click the icons underneath it and you will see you can add X, Notion, Google stuff like calendar and gmail, or a variety of others. But the real power is just talking to it. I asked it what kinds of tasks it can do, and it answered: "Just tell me what you want to accomplish! For example: "Research the latest developments in spatial computing and create a presentation" "Analyze my Gmail for any important emails from this week" "Find information about AI robotics companies and create a spreadsheet" "Generate a video about emerging tech trends" "Help me plan a trip to CES 2026" And so further into AI-first living I go. Have you tried it yet? What do you suggest I do next to use more AI to run my life and build my business? I have another browser tab building a video for me. What a beautiful time to be alive. Download it at:

Robert Scoble

84,349 просмотров • 1 год назад

In our latest Box AI Enterprise Eval, we tested Paul Jankura’s Claude 4 Sonnet and Opus models, now integrated into Box AI, across enterprise Q&A tasks, technical workflows, and advanced coding scenarios—revealing major advancements in developer productivity and content intelligence. AI-assisted coding and development just reached a new milestone! Here's what we discovered: Claude 4 significantly improves understanding, generating, and debugging code across multiple programming languages. Developers can: ↳ Accelerate code generation ↳ Improve debugging ↳ Enhance technical documentation ↳ Build smarter AI agents 👉 Automating Financial Analysis with Code Generation: We evaluated Claude 4 by using the Box AI API to analyze ten complex 10-K financial reports. Claude 4 dynamically generated Python code to fetch file IDs from a Box folder, automating data extraction. Within two minutes, it accurately extracted key company data such as revenues, metrics, and highlights—demonstrating its potential to streamline demanding analytical tasks. 👉 Understanding Enterprise Content: Our evaluation confirms Claude 4 maintains strong performance on enterprise Q&A tasks, effectively extracting precise details from single documents and reliably synthesizing information across multiple sources. This ensures seamless integration of structured and unstructured data alongside powerful coding capabilities. 🔓 Developer-Centric Use Cases Unlocked: Organizations can leverage Claude 4 within Box AI to: ↳ Create custom engineering agents referencing technical documents stored in Box, pulling real-time data from Jira, or finding solutions on Stack Overflow. ↳ Build intelligent technical support bots capable of analyzing user-provided code snippets against internal manuals. ↳ Automate secure code reviews by evaluating repository code (stored in Box) against security policies. ↳ Efficiently migrate legacy systems by translating old codebases into modern languages or platforms. Ready to empower your developers and accelerate innovation? To explore Claude 4 Sonnet and Opus through Box AI Studio and APIs, contact us at ailabs@box.com and request early access today! Learn more:
4:25

In our latest Box AI Enterprise Eval, we tested Paul Jankura’s Claude 4 Sonnet and Opus models, now integrated into Box AI, across enterprise Q&A tasks, technical workflows, and advanced coding scenarios—revealing major advancements in developer productivity and content intelligence. AI-assisted coding and development just reached a new milestone! Here's what we discovered: Claude 4 significantly improves understanding, generating, and debugging code across multiple programming languages. Developers can: ↳ Accelerate code generation ↳ Improve debugging ↳ Enhance technical documentation ↳ Build smarter AI agents 👉 Automating Financial Analysis with Code Generation: We evaluated Claude 4 by using the Box AI API to analyze ten complex 10-K financial reports. Claude 4 dynamically generated Python code to fetch file IDs from a Box folder, automating data extraction. Within two minutes, it accurately extracted key company data such as revenues, metrics, and highlights—demonstrating its potential to streamline demanding analytical tasks. 👉 Understanding Enterprise Content: Our evaluation confirms Claude 4 maintains strong performance on enterprise Q&A tasks, effectively extracting precise details from single documents and reliably synthesizing information across multiple sources. This ensures seamless integration of structured and unstructured data alongside powerful coding capabilities. 🔓 Developer-Centric Use Cases Unlocked: Organizations can leverage Claude 4 within Box AI to: ↳ Create custom engineering agents referencing technical documents stored in Box, pulling real-time data from Jira, or finding solutions on Stack Overflow. ↳ Build intelligent technical support bots capable of analyzing user-provided code snippets against internal manuals. ↳ Automate secure code reviews by evaluating repository code (stored in Box) against security policies. ↳ Efficiently migrate legacy systems by translating old codebases into modern languages or platforms. Ready to empower your developers and accelerate innovation? To explore Claude 4 Sonnet and Opus through Box AI Studio and APIs, contact us at [email protected] and request early access today! Learn more:

Box

285,676 просмотров • 1 год назад

APIs are a bottleneck, not a solution! AI agents can't access 95% of the web because most websites simply don't have APIs. Supplier portals, appointment systems, regional job boards, etc., none of them have developer APIs. The valuable data lives behind logins, multi-step forms, and interfaces built for humans, not machines. That's why, for 25 years, users have been stuck with: - Search engines that index 5% of the web (nothing behind authentication) - Manual data entry or fragile scrapers that break with every CSS update That's the problem TinyFish's Mino solves. It is a web automation API that can simultaneously navigate 100s of websites and turn them into structured data. You send it URLs and a goal in plain English. It returns JSON. The approach is different from typical AI agents: Most browser agents use vision models, which screenshot the page, reason about it, decide what to click, screenshot again, and repeat. Every action needs a model call. It's slow and expensive. Mino uses AI to learn the website structure once, then executes it through deterministic code. First run figures out the site. Every run after that is code-level precision in milliseconds. It provides three core capabilities: → Navigate - Handles logins, forms, and multi-step workflows → Extract - Pulls structured JSON from any site layout → Execute - Runs 100+ sites in parallel with stealth mode The performance is production-grade: - 85-95% success rate on complex workflows - 10-30 seconds per task - Pennies per run Moreover, it works on authenticated sites, bypasses anti-bot protections, and returns clean JSON every time. Lastly, you can also integrate its MCP server with clients like Claude Desktop. I have recorded a walkthrough in the video below. Try now, link in the next tweet.
1:10

APIs are a bottleneck, not a solution! AI agents can't access 95% of the web because most websites simply don't have APIs. Supplier portals, appointment systems, regional job boards, etc., none of them have developer APIs. The valuable data lives behind logins, multi-step forms, and interfaces built for humans, not machines. That's why, for 25 years, users have been stuck with: - Search engines that index 5% of the web (nothing behind authentication) - Manual data entry or fragile scrapers that break with every CSS update That's the problem TinyFish's Mino solves. It is a web automation API that can simultaneously navigate 100s of websites and turn them into structured data. You send it URLs and a goal in plain English. It returns JSON. The approach is different from typical AI agents: Most browser agents use vision models, which screenshot the page, reason about it, decide what to click, screenshot again, and repeat. Every action needs a model call. It's slow and expensive. Mino uses AI to learn the website structure once, then executes it through deterministic code. First run figures out the site. Every run after that is code-level precision in milliseconds. It provides three core capabilities: → Navigate - Handles logins, forms, and multi-step workflows → Extract - Pulls structured JSON from any site layout → Execute - Runs 100+ sites in parallel with stealth mode The performance is production-grade: - 85-95% success rate on complex workflows - 10-30 seconds per task - Pennies per run Moreover, it works on authenticated sites, bypasses anti-bot protections, and returns clean JSON every time. Lastly, you can also integrate its MCP server with clients like Claude Desktop. I have recorded a walkthrough in the video below. Try now, link in the next tweet.

Akshay 🚀

62,659 просмотров • 7 месяцев назад