Spaces OS dispatch #11

Week 14 - Aug 06

Hermes is now the distro's default agent

When we started the project, we wanted to dogfood the OS integration as fast as possible. We used pi, and built our own agent harness and UI. It was useful for us to experiment freely. Now that experimentation phase is over and the code is becoming a liability.

We are thus keeping all the sandboxing primitives we built previously; the sandboxed skills and agent, but running Hermes as the default agent instead. A nice -50k LoC removal which lets us focus on what is important as an OS/distribution, and also leads to a nicer default user experience.

New models

While doing this work, we also took the opportunity to refresh the default models shipping with the OS.

What went out

• qwen2.5:0.5b: a Qwen2.5-0.5B-Instruct Q4_K_M GGUF from late 2024.
• gemma4:e2b / gemma4:e4b: Google's Gemma 4 small variants (April 2026), at 2.3B and 4.5B effective parameters (5.1B/8B real, shrunk via per-layer embeddings), 128K context, Q6_K quants. gemma4 was the default alias.

What came in

• g9v3:3b: a dense 3B from AI9Stars, released July 23, 2026. Llama architecture, Apache 2.0, 131K context; Artificial Analysis rates it well above average for open-weight models its size. Roughly fills the gemma4:e2b slot.
• qwen3.5:0.8b / qwen3.5:2b: a more recent Qwen generation (0.8B, 2B), the direct successor to the deleted qwen2.5:0.5b. The catalog uses unsloth's MTP builds — multi-token prediction gives roughly 1.4–2.2× faster generation in llama.cpp, though the gain is smaller at these sizes. The bare qwen3.5 alias (2B) takes over gemma4's role as default.
• minicpm5:1b: OpenBMB's on-device model, claimed 1B-class open-source SOTA (42.57 average across reasoning/code/math benchmarks vs 35.61 for the previous best in class). Its distinguishing feature is hybrid reasoning: one checkpoint acts as a fast assistant or a deliberate reasoner via an enable_thinking chat-template flag. Standard Llama architecture, so llama-server loads it without forks.
• nemotron3-nano:4b: NVIDIA's edge model from March 2026, a hybrid Mamba-Transformer aimed at agentic use on Jetson/RTX-class hardware. Replaces gemma4:e4b as the "big" entry, under the NVIDIA Open Model License (commercial use allowed).

Offline Wiki

You're in the jungle. You have your laptop hooked to a solar panel, but no Internet. A snake bites you.

That's where the new Offline Wiki skill comes in handy. Or, you know, to run queries, purely offline against any zim, respecting your privacy.

Next up

Next week will include more stabilization and polish of the Hermes integration.