Born2-9B.Small model. Serious review.

Our last Born2 announcement introduced three models: a 9B Irish model, a 35B-A3B code-review specialist and a model small enough for your phone. Today we’re introducing a fourth that wasn’t in that announcement: Born2-9B, a nine-billion-parameter dense model built for code review, coding agents, tool execution and efficient reasoning.

It is available now as an API research preview, served from Europe through Critique.sh at $0.08 per million input tokens, $0.01 per million cached input tokens and $0.20 per million output tokens.

Our objective is straightforward: make capable review intelligence practical to deploy, cheap to run and useful in real engineering work. Born2-9B is our first answer to that question in the Born2 family, and the preview is where we start measuring it on real pull requests.

More intelligence per parameter

Model size is not a reliable measure of engineering capability. A reviewer has to understand existing software, navigate an unfamiliar repository, run commands, read errors and recognise when a proposed change introduces a problem. It also has to know when the code is fine.

We think a small model can learn a lot of that, if it learns from the right examples. So rather than competing on broad knowledge benchmarks, Born2-9B concentrates on practical engineering intelligence: resolving software issues, operating a terminal, using tools reliably and, above all, reviewing code.

Taught by three, checked by three

Born2-9B learned code review from a panel of three far larger models: Xiaomi’s MiMo-V2.6, Z.ai’s GLM-5.3 and Alibaba’s Qwen3.8-2.4T, a mixture-of-experts model with 2.4 trillion total parameters and about 95 billion active per token.

Each one played two parts. As a teacher, it reviewed real changes and wrote candidate findings. As a reviewer, it cross-examined the findings of the other two. A finding only became training signal when it survived that challenge with evidence behind it: a file, a code path and a reason.

The panel matters as much for what it throws away as for what it keeps. A single teacher passes on its blind spots and its bad habits, including the habit of flagging things that are not wrong. Three teachers that disagree, and have to justify themselves to each other, leave behind a cleaner record of what a good review looks like. Born2-9B is trained on that record, and on its restraint.

Preliminary benchmark performance

We are evaluating Born2-9B across established software-engineering and agentic benchmarks: repository-level programming, terminal operations, automated workflows and tool execution.

SWE-bench Verified asks a model to resolve real issues in existing repositories; SWE-bench Pro raises the difficulty with larger code paths and coordinated changes. Terminal-Bench tests work inside a terminal, while AutomationBench and Toolathlon measure multi-step execution and tool use. Together they tell us whether Born2-9B can do engineering work rather than just write convincing snippets.

Against open models of its size

We compared Born2-9B with open models from Qwen, IBM Granite and Liquid AI, from the same 9B class up to larger mixture-of-experts architectures.

At 62.1%, Born2-9B is ahead of every dense model up to 30B in this set, and two points ahead of Qwen3.5-9B at the same size. Qwen3.6-35B-A3B, the foundation of our larger review specialist, is still clearly ahead. That is the gap the 35B-A3B track exists to close at the top end.

The margin widens on harder work. On SWE-bench Pro and Terminal-Bench 2.1, Born2-9B is more than ten points ahead of Qwen3.5-9B and ahead of IBM’s 30B model:

Terminal execution matters most to us here. A review agent has to run the tests, read the failure, poke at the environment and recover when something unexpected happens. That loop is the centre of the intended Critique.sh integration.

Against the efficient API models

Developers choosing a cheap, fast model for high-volume work are mostly comparing proprietary APIs, so we compared Born2-9B against three of them: OpenAI’s GPT-6 Luna, Anthropic’s Claude Haiku 5.5 and DeepSeek V4 Flash.

We want to be plain about what this table shows. DeepSeek V4 Flash scores well above Born2-9B on every benchmark we share, sometimes by more than 40 points. It is also a 284-billion-parameter mixture-of-experts model, more than 30 times Born2-9B’s total size. Anthropic’s results show Haiku 5.5 well ahead of Luna on terminal and computer-use work. We have not yet run Born2-9B on Anthropic’s evaluation set, and we won’t guess.

Born2-9B is not trying to beat these models everywhere. It is a specialist, and a small one: 9B dense parameters are cheap to serve and, once the weights are released, practical to run on infrastructure you control. A specialist only needs to be dependable on the work you actually give it. Code review is that work.

Speed

Our European serving configuration targets 120–150 generated tokens per second. That is in the same range as Claude Haiku 5.5 as measured by Artificial Analysis and above DeepSeek V4 Flash. It is a target, not yet an independent measurement, and we will publish measured throughput from the preview.

API pricing, and a cache that makes it cheaper

Born2-9B’s input is 20% cheaper than GPT-6 Luna and Claude Haiku 5.5, and its output is 60% cheaper. Against DeepSeek V4 Flash at off-peak rates, input is about 47% cheaper and output about 67% cheaper. Against GPT-6.1 Sol, it is 25 times cheaper for input and 50 times cheaper for output.

Then there is the cache. Review traffic is repetitive: the same repository context, style guide and system prompt arrive again and again. Cache hits on Born2-9B cost $0.01 per million input tokens, one cent and an eighth of the standard input price. That matches Luna and Haiku 5.5. DeepSeek’s cache hits are cheaper still, at $0.003 off-peak. Its output, however, costs three times as much, and in review work output is where the money goes.

Try it with your own workload:

At the default setting of 100 million input tokens with 70% cache hits and 20 million output tokens, Born2-9B costs $7.10. GPT-6 Luna and Claude Haiku 5.5 cost $13.70 each, and DeepSeek V4 Flash costs $16.71 off-peak. Without any caching, the same workload costs $12 on Born2-9B and $20 on Luna or Haiku.

For a model that runs on every pull request, summarises every failing test and prepares context for every escalation, those differences add up quickly.

Born2-9B and GPT-6.1 Sol: two models, one review

Not every step in an engineering agent needs a frontier model. Reading files, searching a repository, classifying a change and assembling evidence are high-volume and comparatively simple. Choosing between two fixes or signing off a risky migration is not.

Born2-9B does the first pass; GPT-6.1 Sol spends its reasoning only where it is needed. The number that matters is not price per token but cost per correctly completed review, and that means measuring task completion, correctness, end-to-end latency and inference cost together. We will publish those measurements as the preview matures.

Reviews that come with evidence

A useful reviewer finds a concrete problem, explains why it matters and points to the evidence. It also says clearly when something needs more investigation and when the code is correct.

 async function capture(order) {+  for (let attempt = 0; attempt < 3; attempt++) {+    const charge = await psp.charge(order.total);+    if (charge.ok) return charge;+  } }

This is the behaviour Born2-9B is trained toward. Our next stage of development concentrates on four areas:

Born2-9B is not meant to produce more code or more warnings. It is meant to support better engineering decisions.

Specifications

Research preview, then open weights

Born2-9B is part of the Born2 research programme at Leemer Labs. The API preview supports early evaluation while specialist post-training continues.

Born2-9B is going open source. The API preview comes first; the open release follows, so other developers and organisations can inspect, evaluate and build on the work.

Try it now

Born2-9B is developed by Leemer Labs, part of Leemer Group. Inference infrastructure is managed by Critique.sh. Read the first Born2 announcement for the Irish model, the 35B-A3B specialist and the phone model.

Sources

• Claude Haiku 5.5 pricing, including cache reads and the 100K long-context tier.
• GPT-6 Luna pricing, including cached input and the 272K long-context tier.
• DeepSeek API pricing: Flash cache-hit, cache-miss and output rates, peak hours and model names.
• DeepSeek V4 Flash 0731: Terminal-Bench 2.1, Toolathlon Verified, parameters and output speed.
• DeepSeek V4 Flash: SWE-bench Verified.
• Artificial Analysis: Claude Haiku 5.5 and DeepSeek V4.1 Flash output speed.
• Qwen3.8-Max: 2.4T total and 95B active parameters.
• Qwen3.6-35B-A3B: official model card and benchmark results.

Born2-9B results, prices, throughput targets and training details are reported by Leemer Labs and are preliminary. External scores are developer-reported, use different harnesses and are not a controlled ranking. Prices are standard list prices as of October 2026 and may change. The review example is illustrative, not a recorded model output.