Milk @TeleJet

I just reserved my Cloudflare Wallet tag: corex.cloudflare.pay. Reserve yours now at cloudflare.pay

Big news: MiniMax-H3 by @MiniMax_AI is now the #1 open model in Video Arena: across both Text-to-Video and Image-to-Video. This is +280pts over the next best open model, hunyuan-video-1.5, and a huge improvement from Hailuo-2.3 at #27 (1199 pts) and Hailuo-02-pro at #28 (1197 pts). In Image-to-Video specifically, it scored 1476 pts, just 2 pts away from Dreamina Seedance-2.0 (1478), making it tied for#1 overall! Learn more about how MiniMax-H3 performed in Text-to-Video Arena in thread.

MiniMax (official) @MiniMax_AI

2 days ago

MiniMax-H3 Is Now Publicly Available huggingface.co/MiniMaxAI/Mini…

212 578 5K 1.5M 2K

生成高清视频的算力成本要被这玩意打下来了。 这就是 FoundationVision 刚开源的 FlashVideo，已经被 AAAI 2026 顶会收录。它搞了个极度聪明的“两阶段超分”架构：先用极低的算力跑出 270P 的分镜草稿，觉得构图和语义没问题，再一键秒级冲到 1080P 高清。目前项目在 GitHub 采用 Apache-2.0 协议完全开源，权重和推理代码全放出来了。 低成本预警：270P 先看效果，不满意随时打回，不用盲等显卡烧半天。 高清细粒度：第二阶段 Flow Matching 技术，极少步数补全 1080P 细节。 兼容 CogVideoX：基于 3D VAE 打造，本地显卡部署体验极度丝滑。 拒绝算力黑洞：把 DiT 模型庞大的计算量拆解，显存压力直接减半。 🔗 项目地址：github.com/FoundationVisi…

Mnimax H3 💋👠 💄N$FW The Uncensored variant, built for content creators, educators, indie studios, and everyone who wants a video model that stops saying no. 👇 huggingface.co/Blackfrost-Res…

MiniMax-H3のアップスケーラーノード。要するに最初から高解像度で生成すると時間がかかるから、低解像度で生成してからこれ使ってアップスケールした方が速いという事。 github.com/Tr1dae/ComfyUI…

DeepSeek-V4-Flash-0731 is Ollama's fastest growing model ever in token usage. We are scaling capacity in US & Europe. On Ollama, this model runs with high performance (100tps+) and zero data retention. Your data stays yours. ollama run deepseek-v4-flash:0731-cloud

Running Gemma 4 26B locally with zero GPUs? Very cool. Running it on a 13-year-old Xeon CPU? Wild! It just shows us how far you can push older gear. It’s a great reminder that you don't need modern hardware to try the latest models.

🚀实时语音 Agent 正在从“演示玩具”走向真正可用，而 LiveKit Agents 可能是目前最完整的开源框架之一。 它不只是把 STT、LLM 和 TTS 串起来，还直接提供： • WebRTC 实时音视频 • 电话呼入与呼出 • 语义轮次检测，减少抢话 • 原生 MCP 工具调用 • 多 Agent Handoff • 内置测试与 LLM Judge • 支持自托管完整技术栈 开发者可以自由组合不同厂商的语音和大模型，快速构建 AI 客服、电话机器人、会议助手及实时语音应用。 GitHub： github.com/livekit/agents

huggingface.co/realrebelai/Mi…

终于在本地（RTX 4080）把 MiniMax H3 跑通了。 我感觉效果已经不错了。十秒视频耗时10分钟生成。 不过有个问题还没解决： 明明在提示词里强调了中文说话，最后生成出来的还是不是中文。 接下来准备继续试试本地工作流，看看能不能做一个简短的小动画片。 本视频的提示词我整理在评论区了。

烁皓 @eternityspring

2 days ago

如果MiniMax H3 真能在一张 3060 上跑出能用的效果，那它就是视频生成领域实打实的价格屠夫了。 我准备拿出我吃灰许久的4080来，替大家先试试效果。 等我消息！

1 2 13 77K 11

マンガのコマ割りから画像を生成・合成するComfyUIノード「ComfyUI-PanelComposer」公開。 45種以上のレイアウトに対応。コマごとのサイズで画像を生成し、非矩形を含む変形コマも含めて1ページへ自動合成。 #ComfyUIPanelComposer #ComfyUI URLはリプ⬇️

测试了一晚上：提速 24.3%，最高提速 44% SageAttention 2.2.0 跑 MiniMax H3，在 RTX 4090 48G 上的实测（开源项目地址放最后） 清空显存后的可复现结果，是单段从 425.6 秒降到 322.0 秒，每段省 103.6 秒； 连续跑 13 段，大约省 22 分钟。44% 是首轮测到的极值。 安装： pip install sageattention==2.2.0 --no-build-isolation ComfyUI 在 UNETLoader 后接 KJNodes 的 Patch Sage Attention KJ，选 auto； model 输出同时接 BasicScheduler 和 BasicGuider，4090 会自动走 cuda++。 具体开源地址👇 github.com/thu-ml/SageAtt…

huangserva @servasyy_ai

2 days ago

本地搭建了MiniMax H3（海螺3） 15秒的视频，跑了20多分钟 H3 Ref2VA 原生音画链： H3 Ref2VA INT8 ConvRot 主模型：20.97GB Qwen3-VL-32B NVFP4-AWQ 编码器：15.69GB H3 Video VAE FP16：5.21GB H3 Audio VAE FP32：0.61GB 权重合计：42.47GB ComfyUI 0.30.0 + 原生 MiniMaxH3ReferenceToVideo 节点

13 7 84 46K 69

At first I thought Bloomberg forgot to add DeepSeek's pricing to the chart.

分享一个给 codex app 用的小工具，不入侵你任何配置： 今天看到不少人在吐槽 deepseek 没有多模态，我默默的都给点了赞，然后顺手做了个 deepseek-v4-flash 图片识别 skill。 梁圣说，多模态不影响 AGI，我认可。 可是在这个年代，一个模型没有多模态，不能识图，还真挺别扭的。不能粘贴 bug 图，也别想像素级模仿了。 @iamai_omni 甚至发现 deepseek-v4-flash 会调用 fable 5 的多模态去识别图片，这属实是太奢侈了。 但也启发了我这个工具的思路，本来这个工具是想封装下，让 deepseek-v4-flash 去调用 haiku 的，转念一想，也不知道每个人都订阅了 claude。 haiku 再便宜特么也是是钱啊，关键 haiku 识图能力真不行。 我找了一圈，最后发现 antigravity 有免费额度，这不巧了么。 可能大家不知道，antigravity 拉胯是真拉胯，可是它底层的 gemini 3.6 flash，哪怕是 gemini 3.5 flash 的识图能力，都是没话说的，换句话说就是，当今无人能敌，包括 fable 5。 这个工具用起来也简单，先让你的 codex 啥的给你安装好 antigravity cli，随便登录个免费的 google 账号，不用 google ai pro，免费额度干这事情就用不完的。 然后给 codex 或 claude code 说安装这个 skills，就完了。 如果你想研究源码，源码地址在下面，支持 cli 和 skills 两种调用方式。 github.com/liustack/modle…

✧ 𝕀𝔸𝕄𝔸𝕀 ✧ @iamai_omni

5 days ago

有人测试 v4 flash 发现虽然没有视觉能力，但是会“自主”调用电脑上的 fable 5 去解析图片，没有人指示它们这样做。

64 2 9 10K 2

DeepSeek V4 原生支持 Response 格式，CC Switch 马上也会进行相关支持

Kimi K3 is now available on Ollama’s cloud. To use it with Claude Code, run: ollama launch claude --model kimi-k3:cloud Currently Kimi K3 requires a Pro or Max subscription, and consumes extra usage credits. We’re quickly working on adding capacity to expand access.

Kimi.ai @Kimi_Moonshot

a week ago

Releasing the model weights and technical report of Kimi K3. Kimi K3 is our most capable model: a 2.8T MoE model with native visual understanding and a 1M-token context window. New model architecture: 2.5x the intelligence per unit of compute, not just more params. Alongside

2K 7K 46K 14.3M 14K

@xinzhi 除了9号，其他的也都用过

@xinzhi 都见过

AntiCheatExpert (TencentVM) Devirtほぼ終わり👍 LLVMだとできない/めんどくさいことを簡単にできるので自分でIR/コンパイラ持ってると強いね

Day 08 of building the space game I saw in a dream. 🌙🚀 IT FLIES! The spacecraft I showcased on Day 05 is airworthy, certified and flies under the new realistic physics system I was working day and night for the past few days. ☀️Thanks to astrophysics Professor Fable 5 orchestrating his 4 best PhD Opus 5 colleagues we got: Under the hood: – full 6-DOF rigid-body flight, zero arcade shortcuts – real delta-wing aero: stalls, post-stall drag, the works – fly-by-wire assist so a mouse flies her easily with assist off for the hardcore, oleo landing gear, wheel brakes, speedbrake — you can actually land, an onboard black box records every flight. When physics don't obey, I got the receipts 📼 "Certified" isn't a vibe: 65 automated flight tests green + scripted QC sorties (takeoff, ridge crossing, high-speed decel) flown by the machine before I ever touch the stick. These were also done and orchestrated by the agents under the new skillset that gives them specific controls and camera angles for QC. The other part of this physics system is real weather, wind, lift/drag, shears, turbulence. Yes, I managed to crash several times by underestimating these forces - if high enough, stall recovery was easy, but under 1km, difficult. Skill issue i guess Each part of the spacecraft is programmatic, so you can see the ailerons moving, landing gear retracting, cockpit sliding, even the engine lattices waving. The particle system hasn't been worked on yet, so ignore the random snowflakes going out the engine hahah Another team of agents led by another Fable is currently working on the world generator, so the unfinished island won't be so unfinished soon. Oh, as a bonus, GPT 5.6 Sol built a whole hangar module where we can choose ships and have the tug car haul them to the runway. I'll capture that next time! Next up: a glass cockpit worth flying blind in. Demo video soon. Full avionics!

Startracker 🔺 @startracker

2 weeks ago

Day 05 of building the space game I saw in a dream. 🌙🚀 3D spacecraft made COMPLETELY FROM SCRATCH by Fable 5 + Grok 4.5, not a single downloaded asset. The entire ship is generated by Python code in Blender: geometry, UVs, textures, rigging, even the marketing renders. The

25 28 404 76K 257