antirez

to be precise it's this antirez/ds4 on my 128gb m5 max MacBook running DeepSeek V4 Flash 0731 Q2 + opencode + t3 code as frontend sounds like a rocket engine but it works great

Finally got around to downloading DeepSeek V4.1 Flash Q2 which is 341 GB! Compared to DeepSeek V4 Flash Q2 which was only 81G. On an 128GB machine, you need SSD streaming on antirez/ds4 but it works rather well: about 20 t/s. (V4 Flash was almost twice as fast).

Chaining image-to-video clips from the last frame: what 10 links did to a face (MiniMax-H3, measured) --- title: "Chaining image-to-video clips from the last frame: what 10 links did to a face (MiniMax-H3, measured)" published: true description: "I started each MiniMax-H3 clip from the previous clip's last frame. Across a 10-clip chain, fine detail fell to 61-65 % of the drawn keyframe, contrast rose 14 % and near-white pixels grew up to 2.8 times. Numbers per link, pictures, and the rule I use now." tags: ai, machinelearning, python, video canonical_url: --- For a dialogue scene I generated clips with MiniMax-H3, MiniMax's open-weights video model that can start from a given first frame, end on a given last frame, or both. To make one shot run longer than one clip, I started every clip from the last frame of the clip before. After ten clips the woman on screen no longer looked like the woman I had drawn: !First frames of clips 1, 3, 6, 9 and 10 in one chain Then I measured what each link did. People on forums describe this drift a lot (LTX-2.3, Wan, Seedance), but I didn't find numbers per link, so here are mine. The setup - Model: MiniMax-H3, the first/last-frame checkpoint (FL2VA), run locally on an Apple Silicon Mac through a fork of h3.c, with an INT8 copy of the transformer and lightx2v's 8-step Turbo LoRA (fl2v_turbo_8step_v1.0_768p). - Every clip: 768x1344, 124 frames (5.17 s at 24 fps), 8 steps, seed 42. - Conditioning: a first frame only, no end frame. Clip 1 starts from a keyframe I drew. Every later clip starts from the previous clip's last frame: ffmpeg -sseof -0.05 -i clip_n.mp4 -frames:v 1 -update 1 clip_n_last.png # clip_n_last.png becomes the first frame of clip n+1 Three chains: A, 10 clips of a woman in an armchair; B, 7 clips of a man by a door; C, 9 clips of a woman with a phone. Each prompt gave the person on screen one line of dialogue (the model voices it too), and every prompt said "The camera holds a static shot" and "natural skin texture". What I measured On the first frame of each clip, which is the last frame of the clip before, I computed four numbers on the whole frame and divided each by its value on the drawn keyframe: g = np.asarray(Image.open(path).convert("L"), np.float64) lap = -4*g[1:-1,1:-1] + g[:-2,1:-1] + g[2:,1:-1] + g[1:-1,:-2] + g[1:-1,2:] edge_sharpness = lap.var() # Laplacian variance F = np.abs(np.fft.rfft2(g - g.mean()))**2 r = np.hypot(np.fft.rfftfreq(g.shape[1])[None, :], np.fft.fftfreq(g.shape[0])[:, None]) finest_detail = F[r > 0.25].sum() / F[r > 0.01].sum() # energy above 0.25 cycles/px contrast = (g / 255).std() near_white = (g / 255 > 0.92).mean() # share of pixels !Four metrics per link, chains A and B, relative to the drawn keyframe Detail goes first Finest detail dropped on the first link: 77 % of the keyframe in chain A and 80 % in chain B at the start of clip 2. It bottomed out at clip 4 (61 % and 65 %) and never returned to the keyframe's level: chain A ended at 65 %, chain B at 88 %. Edge sharpness fell too at first, to 73 % and 82 % at clip 3, and then it overshot. From clip 5 on it was above the keyframe, and it reached 174 % in chain A (clip 9) and 159 % in chain B (clip 7). Detail wasn't coming back: the finest band stayed low. What rose were harder, coarser edges. On the shirts you can see it as a crunchy pattern that the keyframe didn't have. Then the faces turn oily Contrast rose with almost every link: by 14 % in chain A (clip 9) and 16 % in chain B (clip 7). Near-white pixels grew faster: from 0.20 % to 0.55 % of the frame in chain A, 2.8 times, and from 0.054 % to 0.124 % in chain B, 2.3 times. On a face this shows up as shine on the forehead, nose and cheeks. Up close: !The same face at clips 1, 4 and 9 The freckles from the drawing are gone by clip 4, the skin is glossy although every prompt asked for natural skin texture, and by clip 9 the bone structure has changed. I didn't score identity with a face-embedding model, so that last part is by eye. One oddity: the first frame of chain A's clip 10 is brighter overall (mean luma up 10 %, saturation down 15 %), which pulls its contrast and near-white numbers back. I haven't looked into why. Mistakes ride along Chain C shows a different failure. At the end of clip 2 the model drew a chat bubble with made-up English over the frame, and since clip 3 starts from that frame, the bubble was in the scene from then on: !First frames of clips 1, 3, 5, 7 and 9 in chain C, with the chat bubble It stayed through the next seven clips, with the text inside it changing from clip to clip. The model treats whatever is in the conditioning frame as part of the scene. A last-frame chain has no way to drop it. The rule I use now When I watched the long chain, I saw smeared text on the phone and faces that looked oiled. Since then I work like this: 1. At most 3 clips in a row start from a previous clip's last frame. Then a fresh drawn keyframe. 2. If a new person or object enters, the clip still starts from the last frame, but it ends on a freshly drawn end keyframe: an edit of the start frame with the new detail painted in. 3. A face close-up or a new location always starts from a freshly drawn keyframe. The cap of 3 is a judgment call, not a threshold the numbers found. In these two chains, the first three clips lose fine detail (to 66 % and 69 %), while contrast has barely moved (100 % and 106 %) and near-white pixels are up 16 % and 11 %. From clip 5 on, all four metrics move away from the keyframe together. You could argue for 2. Check your own chain The measurement needs only the PNGs you already feed the model. Compute the four numbers above for every first frame and divide by the keyframe. A falling finest-detail share with rising contrast is the pattern to watch for. Before you trust a long chain, look at the first, middle and last frame of every clip side by side. A contact sheet per chain would have caught the chat bubble at clip 3. I found it much later, in a rough cut. These are two clean chains with one model, one seed and one resolution, and the metrics cover the whole frame while people move inside it. If you run the same check on another model, I'd like to compare curves.

@DantesClown @RustyAiLab Tensorfold has guides on running across 2x: tensorfold.dev ds4 also works well with running over 2x: antirez/ds4

Open source do GitHub: #15 a #11 | Vol. 17 #shorts Os projetos #15 a #11 do Top 15 open source do GitHub, Vol. 17, 10 segundos para cada um. #15 iris.c: antirez/iris.c #14 PortKiller: productdevbook/port-killer #13 Rootprint: rootprint/rootprint #12 xfr: lance0/xfr #11 FOSSBilling: FOSSBilling/FOSSBilling Vídeo completo com os 15: youtu.be Oferecimento: Hallen Tech, hallentech.com #shorts #opensource #github

Mais 15 projetos open source do GitHub que valem o seu tempo | Vol. 17 Mais 15 projetos open source do GitHub que valem o seu tempo, em contagem regressiva e com 10 segundos para cada um: imagens geradas em C puro pelo criador do Redis, seus agentes de código no celular, um assistente de voz com Raspberry Pi em cada cômodo, sites inteiros clonados para ver offline e muito mais. 📺 Décimo sétimo volume da série Open source do GitHub. Escolhidos pela análise da lista inteira (estrelas, frequência de commits, último commit, README, documentação, capturas e site), com licença e atividade conferidas em 26/10/2026. Qual desses você vai testar primeiro? Conta nos comentários! ⏱️ CAPÍTULOS E LINKS 0:00 Abertura 0:10 #15 iris.c: gera imagens a partir de texto com FLUX.2 e Z-Image em C puro, sem dependências, na GPU (Metal, CUDA e ROCm), do antirez, criador do Redis (2,1 mil estrelas) antirez/iris.c 0:20 #14 PortKiller: app de barra de menu (macOS) e da bandeja (Windows) que mostra as portas abertas, encerra processos com um clique e gerencia port-forwards do Kubernetes (5,1 mil estrelas) productdevbook/port-killer 0:30 #13 Rootprint: logs e traces self-hosted com busca em texto completo, com os índices no seu próprio bucket S3, GCS, R2 ou MinIO, sobre o Quickwit rootprint/rootprint 0:40 #12 xfr: alternativa ao iperf3 em Rust, com TUI ao vivo, servidor para vários clientes, TCP, UDP, QUIC e MPTCP lance0/xfr 0:50 #11 FOSSBilling: cobrança e gestão de clientes para quem vende hospedagem e serviços online: faturas automáticas, lembretes e vários meios de pagamento (1,7 mil estrelas) FOSSBilling/FOSSBilling 1:00 #10 geiger: um comando só, somente leitura, que lista os agentes de IA, servidores MCP e plugins da máquina e o que cada um pode acessar Atomburstofficial/geiger 1:10 #9 getbased: painel de saúde no navegador para exames, DNA e wearables, com os dados no seu aparelho e IA opcional. Ajuda a entender, mas não substitui o médico elkimek/get-based 1:20 #8 GoLive: skill que põe no ar o app feito pelo agente: hospedagem, banco, login, domínio, e-mail e pagamentos nas suas contas, com a sua aprovação em cada passo (1,3 mil estrelas) mikehasa/golive-skill 1:30 #7 kage: clona um site para ver offline: abre cada página num Chrome de verdade, espera carregar e tira todo o JavaScript (3,5 mil estrelas) tamnd/kage 1:40 #6 Pixel2Motion: skill para Codex e Claude que transforma um logo em PNG ou JPG num SVG limpo e anima em HTML, GIF ou vídeo (2,5 mil estrelas) nolangz/pixel2motion 1:50 #5 scroll-craft: skill para o agente criar sites premium animados pela rolagem, com dez sites aprovados como referência (4,0 mil estrelas) nateherkai/scroll-craft 2:00 #4 ollaya: o Ollama dos modelos de decisão: em vez de gerar texto, devolve probabilidades calibradas em milissegundos (1,4 mil estrelas) ollaya-dev/ollaya 2:10 #3 HAPI: controla Claude Code, Codex, Cursor Agent e outros agentes pelo celular, pela web ou pelo Telegram, com relay criptografado (5,3 mil estrelas) tiann/hapi 2:20 #2 Renfield: assistente de IA self-hosted com satélites de voz em Raspberry Pi, mais de 100 ferramentas via MCP, RAG e Home Assistant, tudo offline ebongard/renfield 2:30 Oferecimento Hallen Tech 2:44 #1 Wake: app nativo em Rust que junta as sessões de Claude Code, Codex e outros agentes num só lugar, para buscar e retomar qualquer conversa (1,5 mil estrelas) iAmCorey/Wake 2:54 Encerramento 🧡 Oferecimento: Hallen Tech, liderança de tecnologia sob demanda, software, cloud e segurança para a sua empresa. Site: hallentech.com E-mail: contato@hallentech.com Instagram: instagram.com LinkedIn da Hallen Tech: linkedin.com LinkedIn do Eddwan: linkedin.com Se curtiu, deixa o like, segue o canal e ativa as notificações para a próxima lista. #opensource #github #programacao #ferramentas

Redis 之父下场做本地推理了🚀 DwarfStar，我把他仓库从头到尾扒了一遍，24k stars 不是白来的 先说这是谁：antirez = Salvatore Sanfilippo，Redis 作者本人。仓库 antirez/ds4%EF%BC%8C%E4%BB%8A%E5%B9%B4 5 月创建，GitHub API 实测 24,017 stars / 2,324 forks，4 天前还在推代码。注意语言是**纯 C**，不是 Python 套壳玩具。 这次是核验型实测：我本机没 Mac 也没 GPU，跑不了推理，所以把 README、文档、API 数据逐项核了一遍。不冒充跑过，数字都有出处👇 1️⃣ 支持的模型（README 原文）：DeepSeek V4 Flash（含实验性 vision）、V4.1 Flash、GLM 5.2/5.3、GLM 5.3 Flash、DeepSeek V4 PRO、Qwen3.8 Flash Next。策略很反常识：**故意做窄，不做通用 GGUF 运行器**——必须用项目自己产出的 GGUF 文件。作者原话 "Model support is intentionally opportunistic"，只跟最好的开源权重，有更好的就换掉旧的 2️⃣ 126 t/s 的小字我找到了：8×L40S、Q4 量化、16 会话、10 万上下文。文档原文写得明明白白——"This is total throughput, not the speed of one user." **这是 16 个会话加起来的总吞吐，单用户约 7.9 t/s**。宣传没说谎，但不读小字真会误会 3️⃣ 硬件门槛（README 原文）：Metal 是主目标，Mac 64GB 跑 Qwen Flash Next、96GB+ 跑 DeepSeek/GLM，内存不够有 SSD streaming 兜底；CUDA 以 DGX Spark 为主目标，老 Ada 卡（如 L40S）也能跑；ROCm 支持 Strix Halo。两台 128GB Mac 还能用 RDMA 做 tensor parallel 4️⃣ 最有时代感的一段：README 里有个 "AI full disclosure"——作者公开承认**这软件是 AI coding agent 强力辅助、人主导想法和测试写出来的**，不喜欢 AI 写代码的可以直接劝退。2026 年了，顶级程序员也开始给 AI 署名了 5️⃣ 模型选得准：V4 Flash 284B 总参数但只有 13B 激活（仓库 MODEL_CARD 扒的官方卡片），1M 上下文——Flash 类 MoE 小激活模型正是本地化的甜点，又强又塞得进消费级硬件 ⚠️ 劝退预警也是原文：beta quality，变化极快，"instabilities and regressions are definitely possible"。现在上手=当小白鼠，要有心理准备 💰 算笔账：软件 MIT 免费，真正的价格是硬件——一台 96GB+ 的 Mac 或几张 GPU。对比云端 API，这是"一次性硬件投入换长期零边际成本"的账，跑得多的人才划算 说实话我的判断：这是本地推理赛道目前最认真的个人项目之一——不贪大、专精几个模型、工程扎实。但门槛是真实的：特定 GGUF、大内存、特定硬件，三条至少占一条。126 t/s 等独立第三方复测，我先按"作者自报+QA 记录"看。 📌 觉得有用记得收藏，持续亲手实测，只说真话

antirez built a C64 simulator at the pin level. Lift pin 17 of the color RAM and the machine boots clean: CPU and VIC share the same wrong address, so the RAM just aliases. antirez/c64-hw-simulator

This PR from @CerebralCoding_ merged in my DwarfStart fork. I'll keep it vertical for Apple Silicon. At the end I use ds4 on M3 Ultra for now. ds4#1135

🔥 Android tv без рекламы и своя почта в docker - находки GitHub #shorts #github #opensource #viral Больше полезных находок каждый день - в Telegram: t.me Подборка интересных находок с GitHub за 18.10.2026: antirez/c64-hw-simulator, Gathin23/tv-vibehack, docker-mailserver/docker-mailserver. Сохраняй, чтобы не потерять 💾 antirez/c64-hw-simulator: antirez/c64-hw-simulator Симулятор железа Commodore 64 на C от Сальваторе Санфилиппо (antirez, автор Redis). Подойдёт тем, кто хочет разобраться, как устроены чипы и шины классического компьютера, и интересуется эмуляцией и низкоуровневым программированием. Gathin23/tv-vibehack: Gathin23/tv-vibehack Набор инструментов для очистки Android TV от предустановленной рекламы и лишних приложений через ADB, плюс минималистичный лаунчер в стиле tvOS. Пригодится владельцам ТВ-приставок и телевизоров (в том числе Xiaomi), которым надоели реклама и тормоза стокового интерфейса. docker-mailserver/docker-mailserver: docker-mailserver/docker-mailserver Готовый почтовый сервер в одном Docker-контейнере: Postfix и Dovecot (SMTP, IMAP, POP3), антиспам Rspamd, fail2ban и опциональная интеграция с LDAP. Подойдёт тем, кто хочет поднять собственную почту на своём домене без ручной сборки всех компонентов. #shorts #dev #github #useful #opensource гитхаб, гитхаб подборка, лучшее на гитхаб, новое гитхаб, гитхаб бесплатно, гитхаб ии, гитхаб мсп, гитхаб репы, гитхаб репозитории, лучшие репозитории, рейтинг гитхаб, github ai, best github, new github