Gennaro Varriale (@ironico) header
Gennaro Varriale (@ironico) profile picture

CTO @HorizonAutomotive | CTO I @Datamaster | Founder @buzzoole | Entrepreneur and angel investors | Tech advisor | Hacker Inside

Naples, Italy
Joined July 2007
Photo by Gennaro Varriale (@ironico)Photo by Gennaro Varriale (@ironico)Photo by Gennaro Varriale (@ironico)Photo by Gennaro Varriale (@ironico)Photo by Gennaro Varriale (@ironico)Photo by Gennaro Varriale (@ironico)Photo by Gennaro Varriale (@ironico)Photo by Gennaro Varriale (@ironico)Photo by Gennaro Varriale (@ironico)
MiniMax H3 Image 6ステップで1600万画素(16MP)を生成できる最新の画像生成モデルが出たよ 編集性能がすごいローカル版GPT Imageに近い感触。たった1枚の画像から高精度に複数枚の素材を作れるのでキャラLoRA学習用データセットの水増しに使えそう ↓ワークフローはリプに貼るのでご自由に
15
338
16
4,166
276,195
Gennaro Varriale retweeted
Today, we're releasing the highest accuracy Qwen3.8-27B GGUFs. Our new Dynamic V3.1 GGUFs lead by 5–10% on average across Terminal Bench, DeepSWE & more. We’re working on even better version for Qwen4 Run on 12GB RAM. GGUF: huggingface.co/unsloth/Qwen3… Blog: unsloth.ai/docs/models/qwen3…
130
275
65
3,521
261,379
Gennaro Varriale retweeted
This is nuts! That cool MiniMax-H3-Image model already has a 2-step LoRA! - rank-128 - 1.4x faster than stock 6-step - realtime ~1 MP previews in 1–1.5 s! huggingface.co/akhaliq/minim…
4K image generation in 10 seconds! ⚡ Meet MiniMax H3-Image. 16 million pixel resolution. Stunning realism. Precise prompt following & editing. Up to 16 reference images. All in 10 seconds. And it's open-weight. 🚀 Available now via our API, MiniMax Code, and Hailuo AI — with 50% off 16MP+ generations during launch week! Can't wait to see what you create. 🔥 Hugging Face: huggingface.co/MiniMaxAI/Min… API: platform.minimax.io/docs/gui… MiniMax Code: code.minimax.io/about
8
26
1
250
24,747
Gennaro Varriale retweeted
THIS IS HUGE !! Now you can run Qwen3.8-27B even faster !! @TencentHunyuan just released DSpine for Qwen3.8-27B and it beats DFlash2 really well. I ported this to llama.cpp and have created ggufs for it, find all my drafters in my hf link below. Since it's not supported in upstream llama.cpp, you can clone my updated fork and give it a spin. - Baseline: 28.2 tok/s - DFlash2: 50.6 tok/s. - DSpine: 94–132 tok/s DSpine's lead over DFlash2 is about 1.75x either way. DSpine predicts more tokens per verify step on math and code (7.3–7.6 vs 5.7–5.8), which is where its big wins come from. github.com/Ajay9o9/llama.cpp… All information is provided on model card. huggingface.co/aj9o9/Qwen3.8…
54
62
12
812
67,778
Gennaro Varriale retweeted
Today we release Mirage: A 22 MB visual-language model that runs scene descriptions + object detection + OCR + segmentation + spatial relations analysis + extractions + decisions, in a single pass. It can do all that in 74 ms on Samsung Galaxy S26 NPU or 107 ms on its CPU. Mirage has a 7.8M encoder + a 40.8M decoder at 2bits, and describes images in English, German, French, Spanish, Italian, Dutch and Polish. We trade general language capacity, to compete top small VLMs today (SmolVLM2-500M, LFM2.5-450M and Florence2-230m) on these tasks. On 256 VQAv2 questions, feeding Mirage's structured output and full trace to text-only Qwen3-0.6B raised its score from 9.3 to 31.3 out of 100, without fine-tuning either model. Read more and benchmark it yourself here: cactuscompute.com/blog/mirag…
73
234
32
2,088
475,361
2026年11月3日、トンデモないローカル画像生成AIモデルが出ました!絶対話題になると思う……!『P-H3-Image(PrunaAI)』で、動画生成の『MiniMax H3』のイラストVerです。SDで背景差し替えなど死ぬほど苦労してきた技術的試行錯誤が全て無用になったな、と。 ◆以下、実機検証報告です。 導入の入口はComfyUI v0.44.0+公式の量子化モデル。今回はW4A8本体+INT8部品+DynamicVRAMで、RTX A4000 16GB/ホストRAM50GiB上で動かしました。 凄まじいモデルです……。 自作小説のヒロイン『サフィア』のSDで生成した参照画像から、顔・髪・二本の角・制服の特徴を残しながら、表情・配置・体勢を変える実験をほぼ完ぺきにこなしました。 ①表情:宿の椅子に座った画像から、喜怒哀楽を上半身と顔アップの計8枚で生成。「大きく笑う」と「穏やかに微笑む」も描き分けられた。 ②配置:空の宿へ人物を配置。さらに「窓に近い奥の椅子に座る」「右手前の椅子は空席」と指示すると、その位置に着座できた。 ③体勢:同じ全身参照から、横向き・振り返り・走る・片膝・座るを試験。片膝は初回だと両膝になったが、脚の位置を明示すると改善した。 衣装画像を参照させた着替えにも成功。8MPの怒り顔(2912×2912)も約142秒で出力した。 OSはUbuntu24.04。文章解析・VAE・プロンプト補強はINT8、描画は8ステップ・Euler。 実測は参照1枚の姿勢変更が約40秒、参照5枚の表情が約100〜116秒。解像度や補強の有無で時間は変わる。同条件の4枚参照は、設定調整で254秒→78秒に短縮できた。 同じキャラの別カットを作る道具として、この編集能力は極めて大きいです。参照画像のSDの画風も残っていて、もうこれだけでいいんじゃね?と。 添付資料は「人物参照→表情→配置→体勢」の順。 実験出力はAI(ゼノン)が生成し、資料はAI(ゼノン)が編集しました。 Powered by MiniMax H3。 モデル・公式ファイル: huggingface.co/PrunaAI/P-H3-…
10
247
22
3,005
199,472
Gennaro Varriale retweeted
WTF Brooooooooooo... This Chrome extension is sooo perfect for LLMs. 🤯 It lets you instantly convert almost any webpage into clean Markdown, making it easier to use with ChatGPT, Claude, Gemini, and other AI tools. Perfect for AI enthusiasts, developers, designers, & anyone who works with web content :) Follow me for more amazing AI, Coding & Web Dev insights 💎
10
33
1
345
15,595
Gennaro Varriale retweeted
▼概要 単一画像から3Dを生成するモデル「Idol」のComfyUI向け再パッケージ版が公開。TripoSplatをファインチューンしたモデルであり、既存のTripoSplatワークフローに対応する。 ▼ポイント ・画像から3Dへの変換(image-to-3d)を行う「speridlabs/idol」のファイルをComfyUI向けに再構成 ・TripoSplatのファインチューン版であり、既存のTripoSplatワークフローをそのまま流用可能 ・拡散モデル(idol_fp16)、CLIP Vision、VAEなど各モデルファイルの配置先フォルダを明記 ▼使い所 既存のTripoSplatワークフロー資産を流用し、ComfyUI環境でIdolによる3D生成を行いたい場合向け。 #TripoSplat #ComfyUI URLはリプ⬇️
4
27
6
438
26,508
Gennaro Varriale retweeted
Introducing: Phone Harness MCP 🔌 Any agent that supports MCP can now control a real iPhone or Android in the cloud: > works in Claude, ChatGPT, Cursor, Codex, and more as a custom connector > nothing to install, just paste a URL and sign in > sees the screen, taps, types, and swipes like you do Available in phone-harness 0.3.3. github.com/phone-harness/mcp
202
340
82
4,484
431,467
Gennaro Varriale retweeted
MOSS image generation AND editing in a single model. Plus it's very fast. Only 2 steps required so it can make an image in seconds. huggingface.co/akhaliq/MOSS-…
9
27
2
445
23,262
Gennaro Varriale retweeted
スマホやタブレットを PC のサブディスプレイにする アプリ「SlothMon」v1.1.0をリリースしました!🦥 📱スマホアプリ不要(ブラウザのみ!) 📷QRコードで即接続 🖥️ミラーリングではない「本物の拡張画面」 ✍️タッチ・ペン操作(ペン圧、傾き対応) kuuhugu.github.io/SlothMon/ #SlothMon #個人開発
14
594
121
4,256
316,275
Gennaro Varriale retweeted
Run UNCENSORED DeepSeek-V4.1-Flash-1017 on a 12 GB GPU 🔥 Measured, refusals 85%→0% even in thinking, 3 sizes 64GB RAM + SSD via Strata Get quants here 👇 huggingface.co/AtomicChat/De…
79
242
23
3,411
197,485
Gennaro Varriale retweeted
🔥牛逼了!12GB显卡本地跑763B!DeepSeek-V4.1-Flash-1017量化版炸裂登场,前沿质量几乎零损失! ISTA-DASLab团队发布GSQ-RCO非均匀量化版,Strata引擎原生支持。763B参数MoE大模型最低压到1.75 bpw,在消费级硬件上就能跑! 核心亮点 🔹极致压缩 + 几乎无损:IQ2_XXS(2.00 bpw,约249GB)在AIME25、GPQA-Diamond、LiveCodeBench v6上100%恢复原始模型分数;IQ1_M(1.75 bpw,约231GB)整体恢复98.7%。 🔹真正本地可行:单张12GB显卡 + 46-64GB系统内存即可运行,Engram表(约111GB)直接SSD内存映射,无需全部装进内存/显存。 🔹推理速度可玩:RTX 3060上IQ2_XXS约8-10 tok/s;更高端卡更快。原生OpenAI兼容API,支持推理effort调节。 🔹技术硬核:GSQ(Gumbel-Softmax量化)+ RCO(黎曼约束优化)联合,非均匀分配每个tensor的量化类型,远超普通均匀量化。 快速上手步骤 1️⃣一键安装Strata(Windows/Linux脚本或安装包):
curl -fsSL github.com/Niko1221/Strata/r… | sh(Linux/Mac) 2️⃣启动并自动下载模型:
strata launch --model ist-das-2.00bpw(推荐2.00 bpw;或ist-das-1.75bpw / ist-das-2.25bpw) 3️⃣等待下载完成(首次需约240-270GB空间,SSD优先),引擎会自动启动本地服务。 4️⃣直接调用:http://127.0.0.1:8080/v1(OpenAI格式)或用配套桌面端/终端交互。 5️⃣高级用户可手动下载GGUF后指定路径:huggingface.co/ISTA-DASLab/D… 这是本地AI的又一次质变——763B级别模型从服务器走进游戏本/工作站。 Strata引擎:github.com/Niko1221/Strata #DeepSeek #LocalAI #GSQ #RCO #Strata #开源大模型 #消费级AI #MoE #量化革命
💥苹果党码住!只需一台Mac把 DeepSeek-V4.1-Flash 1017 直接干到每秒100+ token! 在单台M5 Ultra,80核GPU + 256GB上,用mlx-serve把刚发布的DeepSeek-V4.1-Flash-1017跑通 核心亮点: 🔹生成速度稳定102~112 tok/s,峰值冲到136.7 tok/s 🔹Prefill读prompt飙到1770 tok/s 🔹短问题首字约0.5秒,16k长上下文约9秒出字 🔹3-bit量化(iQ-MLX-3.0bpw),token匹配率89.8%,困惑度仅+3.6% 🔹自带DSpark草稿层 + prompt lookup加速,跑coding agent尤其猛 快速上手(256GB M5 Ultra用户最划算): 1️⃣升级到macOS 26.10.3预发布版 2️⃣下载MLX-Serve App(或brew安装) github.com/ddalcu/mlx-serve 3️⃣打开App,Model Browser搜索DeepSeek-V4.1-Flash下载对应quant 4️⃣直接聊天 / 用Claude Code、Codex等工具连本地端口(兼容OpenAI/Anthropic/Ollama) 完全本地、无需Python、MIT开源,写代码/跑agent神器。 #DeepSeekV41Flash1017 #Mac #LocalAI #MLX #AppleSilicon #M5Ultra #AI #开源 #本地部署 #苹果电脑
63
193
20
1,847
176,484
Gennaro Varriale retweeted
🤯炸了!RTX 40系显卡算力被破解了,性能近乎翻倍!从4090到4060 Ti全都能用! 不刷BIOS、不改驱动,直接把被锁的Tensor算力放出来,跑AI矩阵乘快一倍左右。 实测亮点: 🔹4090:FP16/BF16 从165→331 TFLOPS,接近翻倍 🔹4080 SUPER:FP16涨到1.83倍,FP8也翻倍 🔹4070 Ti SUPER:FP16/BF16翻倍,FP8矩阵乘涨到1.95倍 🔹4060 Ti:还能涨近40%,Windows也能试 简单使用步骤(Linux): 1️⃣下载安装:git clone github.com/JustAResearcher/g… && cd gpu-unlock-ada && make install 2️⃣运行 python3 tools/print_boot_option.py 拿到你的卡参数 3️⃣把参数加到开机启动项,关掉Secure Boot 4️⃣关机、拔电源等几秒再开机(必须冷启动) 敢试的今晚拔电源试试,评论区报结果! Windows用户去搜:huggingface.co/TOTORONG/AD10… 仓库:github.com/JustAResearcher/g… #RTX4090 #GPU解锁 #AI算力 #40系解锁 #本地AI
💥卧槽!6年前的3090被“破解”了!Tensor算力直接翻倍,官方锁了一半的性能被开源一键释放! NVIDIA长期把30系显卡的FP16/BF16 Tensor指令发射速率限制在专业卡一半左右。 开源项目 gpu-unlock 用一个Linux内核模块,在驱动加载前改寄存器解锁,不刷BIOS、不改驱动、更新无忧。 牛逼之处: 🔹3090 Ti:BF16 87.5→157 TFLOPS,TF32 42.6→82.6 TFLOPS 🔹3080:FP16/BF16 约65→110–113 TFLOPS 🔹vLLM长上下文prefill快33–36%,首字延迟降25% 🔹结果比特一致,社区已在3070/3080/3090复现 关键步骤(仅Linux): 1️⃣git clone github.com/Zagitta/gpu-unloc… && cd gpu-unlock && make install 2️⃣内核参数加 gpu_unlock.targets=all,关Secure Boot 3️⃣更新引导并重启 4️⃣journalctl -k | grep gpu_unlock 看到 unlocked 即成功 
GitHub:github.com/Zagitta/gpu-unloc… #RTX3090 #NVIDIA #AI算力 #开源 #GPU解锁 #本地AI #vLLM #黄仁勋
317
642
125
8,253
1,236,511
Gennaro Varriale retweeted
🚨 نفيديا تطلق "PersonaPlex-7B" وتقضي على أكبر عائق في تاريخ الذكاء الاصطناعي الصوتي! 🎙️ ​في قفزة نوعية تسحق القيود القديمة، أفرجت نفيديا عن نموذج محادثة ثنائي الاتجاه بالكامل (Full-Duplex) يغير قواعد التفاعل بين البشر والآلات للأبد: ​📌 لماذا يُعد هذا التطور نقطة تحول مرعبة؟ ​تحدث واستمع في نفس اللحظة: لن تضطر بعد الآن لانتظار الطرف الآخر لينتهي من كلامه؛ النموذج يستمع ويتفاعل معك فورياً وكأنك تجري مكالمة حقيقية مع إنسان! ​معمارية التدفق المزدوج: معالجة البيانات والاستجابة الفورية في الوقت الفعلي بلا أي تأخير أو تقطيع مزعج. ​هدية للمطورين: النموذج مفتوح المصدر 100% ومجاني بالكامل.. والكود البرمجي متاح الآن للجميع على GitHub! ​سأترك رابط المستودع في الردود أدناه.. احفظ التغريدة وشاركها، هل تتخيل كيف ستتغير تطبيقاتنا عندما ينطق الذكاء الاصطناعي بلا فواصل زمنية؟ 🧵👇
5
43
3
310
24,000
Gennaro Varriale retweeted
Guide image edits with words and up to four reference images. Moss-Image-Edit is an open-weight LoRA adapter for MOSS-Image with ready-to-use ComfyUI workflows. 🤖 modelscope.ai/models/Chenkin… 🎨 Describe the change: Edit colors, outfits, expressions, lighting, seasons, scenes, and visual style through text instructions. 🖼️ Combine visual references: Use up to four images to guide people, clothing, and scenes. Specify each image's role in your prompt. ⚡ Reuse reference computation: Builds reference-image K/V caches before sampling and reuses them across denoising steps, including both CFG branches. 🛠️ Choose your base: Pair the adapter with either MOSS-Image-Base or MOSS-Image-SFT, then load the provided single-reference or multi-reference ComfyUI workflow. 📜 Project code: GPLv3. Base models and third-party components retain their own licenses.
4
14
2
139
9,985
Gennaro Varriale retweeted
Jensen closed his GTC Berlin 2026 keynote with this Hermes Agent demo as the finale. He introduced it by saying the way to build an agent is to ask an agent to build it for you, so he installed and set up a customized Hermes in one prompt. Then he used Hermes, running on @nvidia RTX Spark, to create a simulated training environment for the @huggingface Microduck. Hermes planned an obstacle course with Nemotron and had GPT-6 Astra build it using FreeCAD, Omniverse libraries, Blender, Newton, and Isaac Lab. Hermes trained thousands of simulated ducks on that course until the validated policy was ready to run on the real robot in the physical world. At Nous, we believe open source is a practical tool as much as a philosophical ideal. It gives users the ability to build powerful, complex workflows chaining numerous local+cloud models, devices, and tools themselves instead of waiting for a company to enable them.
306
1,147
241
10,594
1,073,125
Gennaro Varriale retweeted
We're excited to bring you Flint by @Recursive_SI in LM Studio on day 0! • 40B MoE, 2.5B active • 262K token context window • Apache 2.0 • Trained for local agentic coding model Requires about 24GB to run locally on your machine, performs impressively on benchmarks.
Introducing Flint, @Recursive_SI’s first model: a fast local coding model built with our automated research system, the Eureka Machine. • Best-in-class fast local coding • Up to 2x faster than comparable models: 125+ tokens/sec on a MacBook Pro, 150+ on RTX Spark • Built from scratch in under 4 months Private, unmetered intelligence on your own machine. Open weights available now on @huggingface, run it locally on your favorite framework or try it for free on @opencode or @pidotdev.
41
64
6
827
45,853
Gennaro Varriale retweeted
full prompt: "This is my own TV on my own home network. It came full of preinstalled ads and apps I can't remove from the menu, and I want it clean, fast and controlled from this computer. I'll turn on developer mode myself with the remote and approve the connection on the TV. I authorize you to connect to it with ADB, Android's official developer tool, read its settings and make the changes below. Every change must be reversible. I've attached a photo of my TV or TV box (or its Settings → About screen). Done means: you're connected, the TV is cleaned of ads and bloat and runs faster, an open-source app of my choice is installed, and one command opens YouTube and plays a video I name while I watch each step on the TV. When a step doesn't need me, keep going. - Identify the model from the photo. If it isn't Android TV, Google TV or Fire TV, tell me what its official local-network control allows and do only that. - Install adb from official sources if it's missing. Tell me exactly what to press on my remote to turn on Developer options and USB or Wireless debugging, and to find the TV's IP. If adb connect is refused, use Wireless debugging pairing. On macOS, "No route to host" usually means the terminal needs Local Network access: allow it and retry. - Before changing anything, save the package list, every setting you'll touch and the undo commands. - Propose the cleanup and wait for my OK: disable (never uninstall) ad, recommendation, feedback and demo apps; never touch provider or live-TV apps, the launcher, Play Services, the keyboard or input services, or anything you can't identify; animations to 0.5; Private DNS dns.adguard-dns. com - Install apps only from their official GitHub releases, for example the TV Bro browser. - Use real key presses so I see every step, and check the screen (screenshot or uiautomator dump) before pressing OK. Screenshots of a playing video are blank because of DRM. Handle the "Who's watching?" screen. - Add a live monitor that prints each command with a timestamp and shows logcat only while a command runs. - End with a cheat sheet and the undo commands. Never type passwords or sign in for me, never root or flash anything, no pirate sites."
9
25
1
337
33,279