Gemma4にプロンプトを書いて貰っていて、 「日本語で入力」→「プロンプトだけ返してもらう」 設定にしていたけど、OpenWebUIの設定を変えたら、 プロンプトの前に解説が入るようになった🤔 最初、「プロンプトコピーしにくい」って思ってたけど、解説ある方が可愛くなってきた😊

Whisperで音声認識させて LLMから、Tool（python から date コマンドを実行）を呼び出す。 ModelのSystemPromptに入れるのではなく、OpenWebUIからベースのモデル(gemma4:12b)に対し、 Tools やActionsに登録したスクリプトを関連づけて、新しいモデルとしてOpenWebUIから使う。

So capable, especially when coupled with openwebui.

OpenWebUI のToolsを使って VoiceModeで「What time is it now in Japan?」って質問すると、音声で、「It is currently 12:22 PM on Sunday, June 14th in Japan.」と返答してきた。 これはOpen-terminalを使わなくても動作するので、OpenWebUIが動いているUbuntu 上でdate コマンドを実行してる感じ？

yes it can , it depends on what you run and how you manage the ram. use vllm setup the sequence parallelism, tgen install openwebui and setup vpn for your users.

I retry mistral small 4 on a dgx with openwebui and it is very good even if the cutoff knowledge is 2024. I hope a new mistral medium 3.5 Moe model release for 128 gb unified will be awesome for agentic coding and multimodal on the dgx spark

Absolutely destroyed the openwebui setup due to some obscure kea dhcp bug. Kept digging into the host and container. Nope. Took the opportunity to refresh the containers and decouple ollama and whisper. Cleaner build. So dumb. @pfsense When a Kea pool or scope changes, the existing lease database can end up holding content that no longer lines up with the pool, and Kea then refuses to allocate from that subnet even though the pool looks empty and open. redmine.pfsense.org/issues/1…

PewDiePie made a free self-hosted AI workspace. We installed Odysseus on TrueNAS and tested it vs OpenWebUI, ChatGPT & Claude. Local models, real costs & security — here's the truth. #homelab #selfhosted youtube.com/watch?v=InKHg6WZ…

He sacado nuevo vídeo en mi canal!! Como Vitaminar tu IA Local con OpenWebUI ¡100% Privado! 🚀 youtu.be/NyzqUovUvNk?si=dLlY… a través de @YouTube

It bundles with a @OpenWebUI instance wired up to your agent that can run autonomous self-improvement experiments at night self-hosted using @Cloudflare tunnels relies heavily on @nvidia inference apis for low-cost inference check out the docs.supachad.com

Carte AMD RX9060 (16 Go) VM 64 Go montable à 80 Go ollama openwebui et ça marche sur pas mal de modèles y compris un peu gros (au delà de 80B ça commence à tousser mais du 128B marche si on n'est vraiment pas pressé).

Pour du chatbot (openwebui ou librechat). Tu peux limiter la VRAM et déborder en RAM (mais avec beaucoup de RAM). Ça sera lent mais pas dramatique

I designed a management tool for a 4× DGX Spark Cluster that automatically downloads a model when given a Hugging Face model ID, distributes it across all 4 machines, and serves it through the head node. The model is first downloaded to the head node, then synchronized to the other DGX Spark nodes over the 200G fabric, and launched as distributed inference powered by vLLM/Ray. Thanks to NVFP4 support, I was able to run massive MoE models such as Qwen3.5-397B-A17B-NVFP4 across 4 nodes. The tool also displays OpenWebUI connectivity, cluster health checks, node-level unified RAM usage, and aggregate tok/sec benchmark metrics on a single dashboard. This means model selection, deployment, restart, stop, and performance testing no longer require SSH’ing into each machine one by one. 🎉 I’ll be releasing the tool this week. 🎉❤️ Huge thanks to @NVIDIAAI for building these incredible devices, and to @ASUSTR for their support. 🚀

seems cool! but also expensive :(

One interface. Every model. That's the idea behind Open WebUI OpenRouter. Instead of juggling multiple vendors, API keys, and UIs, organizations can now get a single platform with 600 models ready to go out of the box. Interface and inference, together.

Open WebUIのカスタマイズが面白い。GrokでやってたタスクオートメーションがOpenWebUIでもちゃんと出来た。カレンダーやノートなど他にも機能が多いので触りながら覚えつつ、チューニング安定したら、複数人でも共有して使いたい。