Three MLX quants of Carnice-V2-27B (by @kaiostephens) up on @huggingface
On M4 Pro vs Q5_K_M / llama.cpp baseline:
• mixed_3_6 (12 GB): ~50% faster
• 4bit (14 GB): ~42% faster
• 6bit (20 GB): ~31% faster
huggingface.co/Tranquil-Flow…
huggingface.co/Tranquil-Flow…
huggingface.co/Tranquil-Flow…
Anyone using Carnice on Mac for #Hermes should be able to get a boost using this, although ofc results vary based on hardware :) @Prince_Canuma @awnihannun #MLX
Recommend setting chat_template_kwargs {enable_thinking: false} server-side to keep the model concise on agent loops.
Well, we are using agents so we don't just do one entry for the @NousResearch hackathon... we do two. Introducing hermes-neurovision! It's a visualiser that hooks into your Hermes agent session, takes that input and converts it into visuals! See your agents in action :D
Announcing Hermes-Aegis built for the @NousResearch hackathon! Adds multiple layers of security measures to Hermes-Agent so you can use its sleek agent harness while feeling secure that your device is protected. 🛡️
Hermes v0.2.0 does address some of the attacks portrayed in the demo :D However, Aegis still provides deeper coverage, and I've deployed a patch to integrate the new security features from Hermes such as redactpy and Tirith.
github.com/NousResearch/herm…
Then, enter @OpenClaw. It becomes the most starred repo on Github within weeks, overtaking Linux. People are WILD for autonomous agents. Soon we see many others with @NanoClaw_AI, ZeroClaw, Moltis, @PicoClaw, Nanobot, IronClaw, NullClaw, @microclaw , OpenFang + more.
✅ Audit trail - every BLOCKED and ANOMALY decision is logged with host, path and reason
If this all interests you, check out the github for instructions to install. Testing out the bugs by using Hermes for another build 👀. Once installed just run `hermes-ageis run` to start!
