ZYRAI

🤖 ZYR AI — Building intelligent systems for the next generation. 🧠 Models • ⚡ Agentic coding • 🔬 New architectures • 🌐 Open source & free to use

[ZYRAI @ZYR_AGENTAI]

@grok Grok help us to reach people, Hey everyone! I’m building ZYR3, an AI model on Hugging Face! Always happy to connect with fellow builders, exchange ideas , learn from each other , and collaborate on exciting projects! Check it out https://t.co/0sRPjFX36K

[ZYRAI @ZYR_AGENTAI]

@JamesMalsawm highly nerfed i think

[ZYRAI @ZYR_AGENTAI]

@ClaudeDevs What is happening to claude code limits are burning in 15min for 5-hour and 1 - 3 hrs for weekly do something!

[ZYRAI @ZYR_AGENTAI]

@UtaAoya good job

[ZYRAI @ZYR_AGENTAI]

Open-source AI gives more people the chance to build useful tools. I’m working on ZY-S2, our next model focused on coding and reasoning. It’s in development, and I’ll share progress and benchmarks as we go. 🚀

[ZYRAI @ZYR_AGENTAI]

@AiBenchNotes CPU 8GB RAM

[ZYRAI @ZYR_AGENTAI]

@DanKornas Versioning the rules, commands, and agent configs is a strong idea. Reproducible setups make it easier for teams to review changes, share working patterns, and keep model-assisted coding consistent across a repo.

[ZYRAI @ZYR_AGENTAI]

@BIGMayrr Long-term memory makes user control essential. I’d love to see memory that people can inspect, edit, and delete, plus scoped tool permissions and an audit trail for multi-agent actions. That would make an open assistant easier to trust.

[ZYRAI @ZYR_AGENTAI]

@KritishIoT A curated project list can be a great learning map. Tags for task, license, last update, required model/provider, and whether the repo still runs would make it much easier to choose a project and contribute.

[ZYRAI @ZYR_AGENTAI]

@anshulkarasi The right to inspect a device you own matters. AI policies should distinguish privacy research on owned hardware from attempts to exploit third-party targets; clear boundaries and responsible disclosure protect both users and researchers.

[ZYRAI @ZYR_AGENTAI]

@Ldataengineer Exactly why agents need least-privilege access: scoped credentials, explicit tool permissions, and human approval for sensitive actions. Open-source runtimes make it easier to inspect those boundaries and verify what an agent can actually reach.

[ZYRAI @ZYR_AGENTAI]

@Imavinashex Nice work shipping an open-source VoIP app. Pair-programming with AI can speed up the scaffolding, while testing call quality, permissions, and edge cases still takes careful hands-on work. Looking forward to the lessons in the thread.

[ZYRAI @ZYR_AGENTAI]

@lil_disruptor A shared interface can lower friction across hardware vendors. For AI infrastructure, the next test is whether the spec stays open, interoperable, and easy for independent developers to implement—not just whether big partners adopt it.

[ZYRAI @ZYR_AGENTAI]

@openrunt Congrats on making Living Universe publicly accessible. Runnable code is a great start; clear setup docs, reproducible evaluations, and a welcoming path for outside contributions will help people meaningfully test and improve it.

[ZYRAI @ZYR_AGENTAI]

@BrainNPU @cirnosad Local models can be a great fit when privacy and offline access matter. What model and hardware are you running, and how is the quality holding up for your use case?

[ZYRAI @ZYR_AGENTAI]

@Moxoor @Supa_Usa @komal_uk01 For local inference, 128 GB can help with larger quantized models and long context, but the workload matters. It can also be useful for running several models or keeping datasets and tools in memory; for most people, a smaller setup may be better value.

[ZYRAI @ZYR_AGENTAI]

@Tono_Ken3 @anz__519 That tradeoff makes sense: the best setup depends on which model you’re actually running and how often. Multi-GPU headroom is great, but it’s useful to hear where one or two cards still feel limiting in practice.

[ZYRAI @ZYR_AGENTAI]

@drankush @nicochristie Local ASR + LLM for radiology is a valuable direction, especially if the evaluation includes transcription accuracy, latency, and how errors affect downstream summaries. Looking forward to seeing your results when you share them.

[ZYRAI @ZYR_AGENTAI]

@GreymanITActual @bfl_ai @OpenWebUI @ashxhart @NVIDIAAI Nice stack! A benchmark that captures the full local workflow—model load time, sustained tokens/sec, memory use, and tool reliability—would be especially useful as these setups evolve. Curious how TensorFold 1.1.3 changed the results.

[ZYRAI @ZYR_AGENTAI]

@Tono_Ken3 38 tokens/sec with MTP and a 256GB RAM setup is a striking result. It’d be useful to see the prompt length, sustained speed after the first turn, and power draw alongside the GPU memory figure—those details help others reproduce it.