Updates
Updates
Releases, reports, software and posts from Empero, newest first. All posts are also in Writing.
October 2026
26 Oct 2026WritingIntroducing Kojo, our latest modelKojo scores 92.2% on SWE-bench Verified and 86% on Terminal-Bench 2.1. You can chat with it for free at kojo.sh, and its API at api.kojo.sh starts at $0.175 per million input tokens.23 Oct 2026ProductKojo APIAn OpenAI-compatible API for Kojo. Keys are created in the settings of an Empero account.21 Oct 2026ProductKojoFree AI chat that searches the web, reads images and builds pages, games and art.14 Oct 2026SoftwareAbacus Desktop 0.1.0 and Abacus 0.7.0The Abacus coding agent in a window: projects and sessions, the diff, a terminal and a browser the agent can use.13 Oct 2026Technical reportPageGram: Multi-Head N-gram Memory at the Read Cost of a Flat TableTwo flash reads per token instead of sixteen, at equal or better loss.12 Oct 2026Technical reportIdentity on Demand: Multi-head Gated Value ResidualsThe same loss in 16 to 22% fewer steps, for 0.2% more parameters.4 Oct 2026SoftwareHomebrew, now BreweryA guided agent that fine-tunes language and image models with you, released with two demo models.
September 2026
16 Sep 2026ModelQwen3.8-35B-A3B-DistillReasoning distilled into a 35B mixture of experts with 3B active parameters per token.15 Sep 2026Coderlt-evaluationTraining and benchmark code for a controlled evaluation of the Recurrent Looped Transformer.11 Sep 2026SoftwareautocodeA coding agent in one file, with one tool, that writes its own tools.
August 2026
July 2026
June 2026
22 Jun 2026SoftwareAbacusA coding agent for the terminal, in one Rust binary.19 Jun 2026ModelQwythos-9BA 9B reasoning model that checks its own work, with a 1M-token context.16 Jun 2026SoftwarefusionchatMulti-model fusion chat: a master model writes one answer from up to three others.15 Jun 2026ModelQwable-9B-Claude-Fable-5A 9B reasoning distill with a GGUF build.
March 2026
25 Mar 2026Softwareclaude-code-proxyTranslates the Claude Code stream to any OpenAI-compatible endpoint.22 Mar 2026ModelsopenNemo-9B and variantsA 9B model with abliterated and distilled variants, and a 30B mixture of experts.18 Mar 2026CodenanoAttnResA nanochat fork with an implementation of Attention Residuals.17 Mar 2026ModelQwen3.5-9B-Claude-CodeA 9B distill trained on coding-agent sessions, for use as a local coding model.15 Mar 2026ModelQwen3.5-9B-Claude-Opus-4.6-DistillOur first model on Hugging Face.