Download Maiar-27B on Hugging Face, our fast scientific evidence model built from Valar with Kev.

Download Valar-27B on Hugging Face, our scientific reasoning model for long-running investigations.

Build your own decision model with Kev, our open-source toolkit for turning LLMs into models that return answer probabilities without generating prose.

October 9 update: Introducing Maiar-27B

We’re excited to share our latest model release: Maiar-27B, a 27-billion-parameter model built from Valar using Kev for fast scientific evidence decisions. It reaches 78.1% accuracy, alongside GPT-6 Astra at 79.1% and GPT-5.5 at 78.7%, at an estimated $0.045 per thousand decisions. That’s about 1/790th of Astra’s inference cost.

Give Maiar evidence and a question with possible answers, and it returns probabilities over those answers without generating a written response. It interprets clinical trial outcomes, checks scientific claims and reads results from paper tables, making 26% fewer errors than Valar answering directly across 684 questions.

Maiar-27B scores 78.1% on 684 scientific evidence questions at an estimated $0.045 per thousand decisions. Claude Opus 5.5 scores 81.6% at $11.12, GPT-6 Astra 79.1% at $35.37, GPT-5.5 78.7% at $19.52, Claude Opus 4.8 75.1% at $10.89, and Qwen3.8-27B with thinking 73.0% at $2.49. Cost estimates use different workload profiles.
Maiar-27B brings scientific evidence accuracy close to GPT-6 Astra at a fraction of the inference cost.
Model Evidence accuracy Cost / 1,000 decisions
Claude Opus 5.5 81.6% $11.12
GPT-6 Astra 79.1% $35.37
GPT-5.5 78.7% $19.52
Maiar-27B 78.1% $0.045
Claude Opus 4.8 75.1% $10.89
Qwen3.8-27B (thinking) 73.0% $2.49

Maiar leads or ties every model above on both claim-verification datasets, scoring 86.7% on SciFact and 77.3% on ClimateCheck. These tasks test whether the evidence supports a scientific claim.1

Better decisions from the same 27B foundation

Kev combines task-specific training with direct probability outputs, lifting accuracy from 70.2% for Valar to 78.1% for Maiar. Maiar also beats Qwen3.8-27B with thinking enabled by 5.1 percentage points, making 19% fewer errors. On these evidence questions, enabling thinking lowered accuracy for both Valar and Qwen.

Evidence screening at scale

On one 96 GB RTX PRO 6000, Maiar processes 15.6 decisions per second with eight questions sharing a 512-token input. At $2.50 per GPU-hour, that’s about $45 for a million decisions. A research pipeline can assess the same passage against several claims without generating a separate response for each one.

Maiar’s predictions at 90% confidence or higher were 95.9% accurate, covering 35.4% of the questions. Those confidence scores let a screening system prioritize results and flag uncertain cases for review.

Valar leads sustained investigations; Maiar makes the repeated evidence decisions cheap and fast. Kev is open source, so you can build decision models around your own evidence and questions.


Original announcement: October 6, 2026

Today we’re introducing Valar-27B, a scientific reasoning model trained using Actuator to lead complete investigations. In our paper-reproduction evaluation, it outscored GPT-5.5 and Claude Opus 4.8 and reached 88% of GPT-6 Astra’s score and 89% of Claude Opus 5.5’s, with all four frontier models running at high reasoning effort. All in a 27-billion-parameter investigator.

We built Valar because the cost of a poor scientific decision can be far larger than the cost of generating it. An experiment can run correctly and still test the wrong hypothesis. A useful investigator needs to recognize when new evidence undermines an earlier assumption and decide whether a failed experiment warrants another attempt or a different approach. As agents carry out more work autonomously, that judgment determines how much of their work is useful.

Valar is trained for this kind of long-horizon reasoning. It directs investigations over multiple exchanges with its tools, interpreting results and choosing subsequent experiments in the context of what came before. Coding agents handle implementation and execution. That division lets us concentrate a smaller model’s capabilities on scientific judgment while benefiting from continuing advances in coding tools.

Scientific reasoning across six domains

Valar was trained on 812 demonstrations spanning biology, chemistry, earth and climate science, materials science, physics and machine learning. Paper reproduction tests its reasoning across those fields: we remove a result figure from a published paper and ask the investigator to reproduce it. The model has to interpret the method, direct an implementation and assess what the executed experiment actually established.

Across 28 held-out problems, Valar scored 0.831, about 5% above GPT-5.5 and Claude Opus 4.8, both at 0.790. GPT-6 Astra and Claude Opus 5.5 scored 0.945 and 0.936, respectively.2

Scientific reproduction scores on 28 held-out problems: GPT-6 Astra 0.945, Claude Opus 5.5 0.936, Valar-27B 0.831, GPT-5.5 0.790 and Claude Opus 4.8 0.790.
Figure 1 Mean scientific reproduction score across 28 held-out problems. Frontier models ran at high reasoning effort.

Valar produced 10 reproductions scoring at least 0.9, compared with five for GPT-5.5 and three for Opus 4.8. It also exceeded Opus 4.8 on four of the five graded dimensions, including reproducing the scientific claim and the fidelity of the implementation.

In one groundwater-modeling investigation, Valar requested checks of boundary conditions and limiting cases, read the execution log and asked for a second review of the numerical results. The completed experiment recovered approximation errors close to those reported in the paper.

A smaller investigator with a choice of tools

We also evaluated Valar with Claude Sonnet 5.5 as its coding agent. Our strongest Valar configuration for that setup scored 0.748, about 12% above Opus 4.8 at maximum effort and 94% of Opus 5.5’s score at maximum effort. It outscored Opus 4.8 on 18 of 21 problems.

Sonnet configuration on 21 development problems: Claude Opus 5.5 0.794, Valar-27B 0.748 and Claude Opus 4.8 0.671.
Figure 2 Mean scientific reproduction score with Sonnet as the coding agent on 21 development problems.

Teams can run Valar privately on workstations and GPU servers, pairing it with the coding tools that fit their research.

What Actuator makes possible

Valar was trained using Actuator, our patent-pending system for closed-loop model training. Actuator feeds measurements of model behavior back into the training process, adjusting training pressure when protected capabilities start to drift. The same control paradigm spans fine-tuning, alignment, distillation and compression, allowing a model to specialize while protecting the capabilities its application depends on.

Valar shows the value of that approach in a demanding scientific role: a 27B investigator scoring above named frontier models on complete research tasks. For product teams, specialization offers a way to build models around their own workflows and deployment constraints. We built Actuator to make that process repeatable, so teams can develop capabilities specific to their products while retaining control over how those models are trained and run.

Scientific reasoning researchers can trust

For AI to have a meaningful impact on science, researchers need to trust the rigor and quality of its work. As we discussed in our post on vibe science, that requires continued advances in scientific interpretation, hypothesis generation and long-horizon reasoning while reducing hallucinations. These capabilities still need to improve even in the strongest frontier models.

That commitment to scientific rigor is at the heart of Marvin. Valar brings more of that reasoning capability to a 27B model, making it practical on workstations and private GPU servers and expanding where scientific agents can run. Smaller models also allow more investigations to run in parallel, bringing this capability within reach of more research teams. The weights are available on Hugging Face so other researchers can build on Valar and help advance agentic science.

We’ve also used Kev to build Maiar-27B, a fast decision model based on Valar that returns choices and probabilities without generating a full written response. See the October 9 update above for the results.

Learn more about Actuator · Explore Marvin

  1. Accuracy: 684 questions across Evidence Inference, SciFact, ClimateCheck, PubMedQA and SciTab. Maiar is task-trained; frontier models run zero-shot at high reasoning effort. Costs are estimates for different workloads: Maiar uses eight questions per 512-token input; Qwen uses science-question batch timing, both at $2.50/GPU-hour. Frontier costs use benchmark token usage at API list prices. ↩

  2. Valar and Opus 5.5 used GPT-6.1 Sol as the coding agent; the other frontier runs used GPT-5.6 Sol. ↩