The Born2 family from Leemer Labs, served from Europe. Start with Born2-9B, our code-review specialist, today. The Irish model and the 35B-A3B review specialist are next.
Already have a key? Sign in to the platform
€5 in free credit on every approved account. Top up any time at ray@critique.sh. Inference is managed by Critique.sh, in Europe.
Each Born2 model has one clear job. They are post-trained from open Qwen foundations on NVIDIA B200 GPUs, using a 50-billion-token dataset Leemer Labs built for the work.
born2-9bCode review, coding agents, terminal and tool use
A dense nine-billion-parameter specialist taught by three frontier models, each checking the others. Built to find the real defect and stay quiet when the code is fine.
born2-9b-gaeilgeIrish language, civic information, bilingual work
Post-trained from Qwen3.5-9B for Gaeilge and Irish civic context. It follows people between Irish and English mid-sentence and keeps the source record within reach.
born2-35b-a3b-reviewFrontier-level code review at ~3B active
The larger review specialist for Critique.sh. A mixture-of-experts model where about 3 billion of its 35 billion parameters work on each token.
Preliminary results across repository-level programming, terminal work and tool use. Born2-9B leads every dense open model up to 30B in this set on SWE-bench Verified.
Preliminary results reported by Leemer Labs. Methodology, run counts and configuration will accompany the technical report.
External scores are developer-reported and use different harnesses; this is not a controlled ranking. Liquid AI separately reports 66.2% for Qwen3.6-35B-A3B under its OpenCode configuration.
Prices are set in euros. Repeated repository context hits the cache at a cent per million tokens, and opting in to training takes 75% off every request.
Born2-9B shows the top of its 120–150 tok/s European serving target, not an independent measurement. Haiku 5.5 is measured at medium and high effort. OpenAI has not published, and Artificial Analysis does not yet list, an output speed for GPT-6 Luna.
Born2 is served by Critique’s inference API, which also routes and hosts 15+ leading open models: DeepSeek, Kimi, GLM, Qwen, MiniMax, gpt-oss, Gemma, Llama, Nemotron and more, at market prices that are often the best available. The open models are offered by Critique directly, on critique.sh.
Allow training on your prompts and every request costs a quarter of the price. It is off by default and opt-in only, and you can turn it off at any time.
Tell us what you are building on the API platform. Approved accounts open with €5 of free credit.
Create a key in the dashboard and point any OpenAI SDK at the base URL.
Give Born2-9B a pull request, a failing test or a terminal task, and read the findings.
curl https://leemergroup.com/api/v1/chat/completions \
-H "Authorization: Bearer $BORN2_API_KEY" \
-H "Content-Type: application/json" \
-d '{
"model": "born2-9b",
"messages": [
{ "role": "system", "content": "Review this diff. Cite file and line." },
{ "role": "user", "content": "<your diff>" }
]
}'Born2-35B-A3B-Review builds on Qwen3.6-35B-A3B. These are the upstream checkpoint’s published scores, the floor our review post-training starts from. Born2’s own results will come with their methodology.
Scores are from Qwen’s model card for the upstream checkpoint, not Born2 results.
Sign up for the API platform and try Born2-9B on real work, on us. You will be first in line for Born2-9B-Gaeilge and Born2-35B-A3B-Review.
Every approved account gets €5 in credit to try Born2-9B on real work. At preview prices that is more than 60 million input tokens on Born2-9B. To top up, email ray@critique.sh.
Europe. Inference is managed by Critique.sh, part of Leemer Group, in Europe.
Both are in post-training at Leemer Labs. They will join the API when their evaluations are ready to publish, and platform accounts will hear first.
Open weights are planned for the Born2 models, with a model card and evaluation methods. The API preview comes first.
Born2-9B results, prices and throughput targets are reported by Leemer Labs and are preliminary. External scores are developer-reported and use different harnesses. Prices are set in euros as of October 2026; pounds and dollars are shown at reference rates. Prices may change before general availability. Read the Born2-9B announcement and the first Born2 post for the full story.