in my results, accuracy-wise Ternary-Bonsai-8B is on par with Qwen3.5-4B. But in accuracy-per-byte, bonsai is the clear winner:

=> Ternary-Bonsai-1.7B achieved 65.1% from 462 MiB, beating Qwen3.5-0.8B by 12 points while being ~5% smaller on disk. => Ternary-Bonsai-4B is the accuracy-per-byte winner above 1 GiB. 83.0% from only 1.1 GiB, within 2 points of Qwen3.5-4B at 40% of the weight size.

they show strong promise on edge devices and where disk space is limited. I think this lab is worth watching.

i guess gastown is a better choice for now? idk i don't feel good about "relatively stable"

It’s really that straightforward. If tomorrow they decide GPUs are better allocated to enterprise use, they could start removing the $20 plan just as quickly overnight, the same way they did tonight.

Does this mean `huggingface.co/Jackrong/Qwen3.5-27B-Claude-4.6-Opus-Reasoning-Distilled` is unusable? Had anyone seen fake tool calls working with this model?

> Theory only takes you so far

I built Claudebox to get more out of my Claude subscription. It runs Claude Code in a sandboxed Docker container and exposes an OpenAI-compatible API; so any tool can use Claude as a backend.

No API key needed, no extra billing; it authenticates with your existing Claude credentials. The container is network-isolated (only Anthropic domains allowed), so Claude gets full agent capabilities (file editing, shell, code analysis) without access to your host or the open internet.

I mainly use it for personal data processing tasks where I want an agent API but don't want to pay above my subscription.

GitHub: https://github.com/ArmanJR/claudebox

Comments URL: https://news.ycombinator.com/item?id=47485885

Points: 2

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=47326622

Points: 1

# Comments: 0

I built this tool specifically for Claude Code users who hit the 5-hour rate limit wall mid-flow. There's no official plan between Pro ($20/mo) and Max ($100/mo). it's a fixed gap with nothing in between.

The workaround most people do manually: running multiple Pro accounts and switching when one is limited. This actually works, but naive rotation wastes a lot of capacity. When you activate an account turns out to matter as much as which one you use. A single throwaway prompt sent a few hours before your coding session can unlock an extra full cycle.

PolyClaude automates this. You tell it your accounts, your typical coding hours, and how long you usually take to hit the limit. It uses combinatorial optimization to compute the exact pre-activation schedule, then installs cron jobs to fire those prompts automatically. When you sit down to work, your accounts are already aligned.

It's free and open source. Install is one curl command, then an interactive setup wizard handles the rest.

Repo: https://github.com/ArmanJR/PolyClaude

Hope you find it useful!

Comments URL: https://news.ycombinator.com/item?id=47292043

Points: 1

# Comments: 0

Comments URL: https://news.ycombinator.com/item?id=47275623

Points: 1

# Comments: 0

I wonder why you don't remove it yourself. Aren't you already editing the plan?

Comments URL: https://news.ycombinator.com/item?id=46489639

Points: 2

# Comments: 0