Rijoda Labs

[Rijoda Labs @RijodaLabs]

Rijo vs Hermes, warm, same laptop, 3 sessions:

10 KB read: 85–241×
search: 6.4–7.2×
terminal: 7.4–9.3×

720/720 calls correct. Tool layer only — no model, no tokens.

https://t.co/KFGP8DAlm8

[Rijoda Labs @RijodaLabs]

Rijo Core vs Hermes, warm tool calls, same laptop, 3 sessions.

The fair result:
→ 10 KB read: 85–241×
→ search: 6.4–7.2×
→ terminal: 7.4–9.3×
→ skills_list: 8.1–9.0× (different libraries, indicative only)

Method: 2 warmups + 30 measured calls per case and arm, interleaved, unique inputs, every call correctness-checked. Hermes blocks repeated identical calls, so each measured call is a real call.

Tool layer only: no model, no prompt, no tokens. Search had one I/O-noisy session; the median direction held. The report has p5/p95 and the caveats. OpenCode is the next arm, not in this run.

https://t.co/B3T8Y1cxHM

What should the next arm be: OpenCode, or an end-to-end task suite?

[Rijoda Labs @RijodaLabs]

Rijo Core is open source.

The engine behind Rijo Code: 34 native Rust tools, one guard on every execution path, sessions, a workspace brain, MCP in both directions.

Apache-2.0. The desktop workbench stays private.

https://t.co/8uSkTr9J5f

CLI first, or MCP first?

[Rijoda Labs @RijodaLabs]

We rebuilt a 2B model's quantisation for one laptop: 4.7% smaller, 4% better PPL, 9-16% faster decode.

The trick was the calibration corpus, not fewer bits.

First release from Rijoda Labs.
https://t.co/X7WSj6bhn0

What quant do you run?

[Rijoda Labs @RijodaLabs]

Our Quant has already 55 downloads on @huggingface since releasing yesterday 🥳https://t.co/X7WSj6bhn0

[Rijoda Labs @RijodaLabs]

Check out our new Quant of LFM3-Omni-2B https://t.co/X7WSj6aJxs

[Rijoda Labs @RijodaLabs]

We rebuilt a 2B model's quantisation for one laptop: 4.7% smaller, 4% better PPL, 9-16% faster decode.

The trick was the calibration corpus, not fewer bits.

First release from Rijoda Labs.
https://t.co/X7WSj6bhn0

What quant do you run?