Born2.Three models. Real purpose.

Born2, the second generation of the Born model programme, is in its final phase at Leemer Labs. We are post-training open-weight Qwen models on NVIDIA B200 GPUs, using a 50-billion-token dataset we built for the job.

The programme now has three models, each with one clear job: a 9B Irish model for DáilDex and for organisations that want to run it themselves, a 35B-A3B code-review specialist for Critique.sh, and a very small model that runs locally on a phone. API access is coming soon. It will be served through Critique’s inference API and hosted in Europe, and every account gets €20 in free credit during the first week.

Three models, three jobs

All three are under development. DáilDex and Critique.sh are where the first two are headed; neither product runs a Born2 model today.

What Born2 is built on

Born2 does not start from nothing, and it does not pretend to. The foundations come from Qwen; the training runs on NVIDIA hardware. Our contribution is the layer on top: the data, the post-training and the evaluation that turn a capable general model into one that does a specific job well.

What Qwen3.5 brings to an Irish model

The Irish model starts from Qwen3.5-9B, released by Qwen in March 2026 as part of the Qwen3.5 series. Pretraining is where a model picks up its languages, its general knowledge and its habits. Qwen’s announcement describes what changed in Qwen3.5:

1. More data

   A larger pretraining run than Qwen3.

   Qwen trained Qwen3.5 on a significantly larger scale of visual-text tokens than Qwen3, with enriched multilingual, STEM and reasoning data under stricter filtering.

2. 201 languages

   Irish is on the list.

   Coverage grew from 119 to 201 languages and dialects, with priority on lower-resource languages. Irish, Welsh and Scottish Gaelic are all named.

3. 250k vocabulary

   A tokenizer that wastes less.

   The vocabulary grew from about 150k to 250k entries. Qwen reports 10–60% better encoding and decoding efficiency across most languages.

4. Native multimodal

   Text and vision from the start.

   Early text-vision fusion during pretraining rather than a vision adapter bolted on afterwards.

5. Hybrid attention

   Built for long context.

   Gated DeltaNet linear attention interleaved with gated full attention, plus multi-token prediction. Qwen3.5-9B has 32 layers and a 262,144-token native context.

Qwen does not publish a single pretraining token total for Qwen3.5. It says the series was trained on a significantly larger scale of visual-text tokens than Qwen3, whose technical report gives about 36 trillion tokens across 119 languages and dialects, and that its training pipeline scales stably to tens of trillions of tokens.

Irish appearing on that list is a starting point, not a finished capability. A language somewhere in a pretraining mix is not the same as a model that reads Dáil records accurately, handles Irish legal and civic terminology, and follows people as they move between Gaeilge and English mid-sentence. Closing that gap is what Born2’s post-training is for.

A 50B-token dataset, built for the next phase

Qwen’s pretraining gives the models breadth. Our 50B-token dataset is about depth: post-training, task-specific adaptation and evaluation in the settings where these models are meant to work. Each track is evaluated separately for its own job.

The 9B Irish model, in DáilDex or on your own servers

DáilDex is the public home for the Irish model. Picture someone trying to understand a parliamentary answer while switching between Gaeilge and English. Fluent sentences are only the beginning. Names, institutions, terminology and the line between a source and an inference matter just as much. The aim is a model that explains the record and keeps it within reach. It should never quietly replace it.

The same model is built for enterprises. At 9B parameters it is small enough for an organisation to host on infrastructure it controls, on premises or in a private cloud. Contracts, case files, citizen correspondence and internal documents stay inside the organisation; prompts and outputs never pass through a third party. That matters for public bodies, legal and regulated teams, and businesses that work in Irish, where sending documents to an outside API is not an option.

Organisations that would rather not run hardware will be able to use the same model through the hosted API in Europe. The model is the same either way.

The review specialist that earns the interruption

The larger model builds on Qwen3.6-35B-A3B for Critique.sh. Our target is frontier-level code review from a model that activates only about 3B parameters per token: understand a change, challenge its assumptions and explain findings an engineer can verify.

Think of a retry that charges twice. A permission check that trusts the wrong identifier. A test that passes while the failure path stays untested. These are the defects we want the specialist to find. Restraint is part of the job: a list of speculative warnings costs more time than it saves.

Qwen3.6-35B-A3B is a mixture-of-experts model: 256 experts, with 8 routed and 1 shared expert active per token, so about 3 billion of its 35 billion parameters do the work on any step. The active count describes computation; it does not mean the full model fits in the memory of a three-billion-parameter one. Qwen’s model card reports 73.4 on SWE-bench Verified and 51.5 on Terminal-Bench 2.0 for the upstream checkpoint, up from 70.0 and 40.5 for Qwen3.5-35B-A3B. Born2’s own review results will be published with their evaluation methodology.

A Born2 model for your phone

The third model is the smallest. It is built to run entirely on a phone: no network round trip, help that still works without a signal, and nothing sent to a server. For a lot of everyday tasks, that is the most private and the fastest place a model can run.

We will share its size, foundation and supported devices with the release.

Three ways to use Born2

Born2 models will be available through the API, served by Critique’s inference API and hosted in Europe. For the first week, every account gets €20 in free credit: enough to try the models on real work before committing to anything. Pricing and the opening date will be announced with the launch.

What we are aiming for

A better explanation of an Irish civic document. A review that catches a failure hidden inside a reasonable-looking diff. A helpful model that never leaves your pocket. Those are outcomes worth the training.

Our open source commitment

Born2 is meant to be useful beyond Leemer Group: something other researchers and builders can inspect, evaluate and build on. Our plan is to release model weights with supporting documentation, including intended uses, evaluation methods and known limitations, subject to the applicable licences and rights. The exact artefacts, licence terms and access arrangements will come with the release; nothing is being released in this announcement.

We are post-training existing foundations, not pretraining a new model from scratch. Qwen’s upstream models are published under Apache 2.0. That licence does not by itself set the terms of a future Born2 release.

The plan from here

1. 01

   Evaluate in the real setting.

   Work with early users on Irish-language and civic tasks, on code-review cases and on-device use. Examine source fidelity, useful findings and false alarms.

   Next
2. 02

   Open the API.

   Serve Born2 through Critique’s inference API, hosted in Europe, with €20 of free credit for every account in the first week.

   Soon
3. 03

   Document what holds up.

   Prepare model cards, intended uses, evaluation methods and known limitations. Report results we can explain.

   Before release
4. 04

   Release, then keep learning.

   Publish weights and release details under the applicable licences. Bring feedback into the next iteration and the product integrations.

   Planned

The post-training runs on NVIDIA B200 GPUs, part of NVIDIA’s Blackwell generation. The hardware gives us room to run, examine and iterate. The data, the training decisions and the resulting behaviour still decide whether a model earns its release.

We are not announcing benchmark scores or a release date today. Those belong with results we can explain.

Become an early user

Tell us which model you want to try, what you are building and the difficult cases you want it to handle. Irish speakers, researchers, civic organisations, enterprises planning to self-host and engineering teams can all bring useful perspectives.

Follow the Journal or RSS feed for the API launch and release updates.

Sources

• Qwen3.5 announcement: pretraining, languages, vocabulary and architecture.
• Qwen3.5-9B: official model card and Apache 2.0 licence.
• Qwen3.6-35B-A3B: official model card, MoE design and benchmark results.
• Qwen3 technical report: the 36-trillion-token pretraining baseline.
• NVIDIA B200 and the Blackwell platform.

The programme status, hardware, hosting plans, API credit offer and release plans in this post are reported by Leemer Labs. The sources describe upstream technology; they do not independently verify Born2’s performance. The film illustrates the programme’s aims rather than showing measured model outputs.