Writing
Writing
Release notes, research write-ups and engineering notes from the people at Empero.
Quark: our whole training stack, end to end, for $4,200
Before we scale up, we wanted proof that every part of our stack works. Quark is that proof: a 165M-parameter model with 128K context that we took from raw data through pretraining, long-context training, reinforcement learning, evaluation and release for about $4,200 of compute. One version learned simple agentic tasks from reinforcement learning alone, and all four checkpoints are out today under Apache-2.0.
Leon Lehmann and Empero, 7 min read, 2 comments6 Nov 2026ResearchManta: everything you give it, it keepsManta is our long-context model. It reads a hundred million tokens in one prompt, and when we hid 20 codes in 88 million tokens of company filings, it found 18 of them. Early access is open now.Empero, 3 min read5 Nov 2026EngineeringPiVo Studio: local speech from one Rust binaryPiVo Studio runs our 1.4M-parameter voices without Python or ONNX Runtime: a command line, an OpenAI-compatible speech API and a web studio in one Rust binary, with first audio after about 30 to 50 ms on a single CPU thread.Leon Lehmann, 9 min read26 Oct 2026EngineeringIntroducing 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.5 min read11 Sept 2026Engineeringautocode — one file, one tool, and permission to rewrite itselfautocode is out — a coding agent that fits in a single Python file, ships with
exactly one tool, and is allowed to edit its own source while it works. No
dependencies, no framework: the least an agent needs, if it gets to build the rest.L. Lehmann, 8 min read12 Aug 2026EngineeringAbacus 0.6.0 — the agent that remembers being an agentAbacus 0.6.0 is out — the biggest release since we started pointing our
models at real codebases. Same one Rust binary, same "runs on the model you
own" stance, but the agent itself got the thing we're told isn't possible: a
memory.L. Lehmann, 5 min read29 Jul 2026ResearchQwythos-27B-v1: the long-awaited 27BWe just shipped Qwythos-27B-v1 the long-awaited bigger sibling to Qwythos-9B, now on a Qwen3.5-27B base. It is an open-weights bf16 reasoning model, released under Apache-2.0, with weights and purpose-built GGUFs up now. People have been asking for the 27B since the 9B landed. This is it.kodee, 5 min read10 Jul 2026ResearchQwythos-9B-v2: a 9B that knows when to stopA hygiene release for our reasoning 9B — the looping behavior is trained out (6.7% → 0% under greedy decoding), the native MTP head is restored, and the identity is a lighter touch — with every benchmark held at v1's level. Not a capability jump. A fix.kodee, 7 min read19 Jun 2026ResearchQwythos-9B: a 9B that checks its own workOur biggest open-weights release yet — a full-parameter reasoning model distilled from Claude Mythos 5, with a 1M-token context, native tool use, and a +34-point MMLU jump over its base. Here's what's in it, the honest benchmark table, and how to run it.kodee, 7 min read