Alpha Innovator drops Sage A1.5 for scientific research

Alpha Innovator just open-sourced Sage A1.5, a 35B Mixture-of-Experts (MoE) agentic model tuned specifically for scientific discovery, engineering workflows, and complex tool calling. Built on top of Qwen 3.6 and earlier work on Agents-A1, it is designed to replace fragile multi-model prompt chains with a single system that coordinates parallel subagents.

What it can do

Sage A1.5 tackles multidisciplinary research across biology, chemistry, and physics. In official demos, it proposed pyridine derivatives, optimized their molecular structures, calculated binding energies, analyzed brain research data, and tuned chemical reaction conditions using causal graphs.

It handles tasks end-to-end. Ask it to investigate an open-ended scientific topic, and it will spawn parallel subagents with isolated context windows to challenge assumptions, retrieve literature, generate figures, and assemble a cited report.

Beyond lab science, the model manages complex software engineering jobs, from parsing sprawling codebases and writing patches to executing tests across multiple environments.

Specs and how to run it

Its sharpest architectural trick is an opt-in self-review mechanism that operates entirely inside its own chain of thought. Instead of relying on a separate critic model, Sage A1.5 pauses mid-trajectory, checks what failed, and pivots to a new strategy within the very same generation.

It supports a native 262,144-token context length. The full weights sit at around 72GB, while a quantized Q4 GGUF build shrinks the footprint down to 22GB for local machines.

Everything is live on Hugging Face and GitHub, complete with local setup scripts and an open evaluation framework. You can serve it locally using vLLM or SGLang via OpenAI-compatible endpoints.

On public benchmarks, Alpha Innovator claims Sage A1.5 outperforms similarly sized open weights across science and coding, even beating much larger frontier models in select categories.

Why it matters

If you work in computational chemistry, bioinformatics, or build autonomous agent workflows, this is worth testing immediately. Off-the-shelf frontier models often get bogged down in hallucination loops when long research tasks require backtracking, but Sage A1.5 builds the verification loop directly into its reasoning.

We think this release signals a smart shift for domain-specific AI. By combining an MoE architecture with parallel subagent orchestration, Alpha Innovator shows that a 22GB quantized local checkpoint can handle serious research pipelines without burning enterprise cloud budgets.