Agentic Orchestration
Orchestrates parallel investigations across subagents with independent contexts, enabling them to challenge assumptions, refine conclusions through evidence, and synthesize their findings into comprehensive results.
From Long-Horizon Execution to Scientific Exploration with Capability Composition
Sage-A1.5
A 35B Mixture-of-Experts agentic model with enhanced capabilities across engineering, long-horizon search, scientific research, instruction following, and tool calling. Building on the foundation of Agents-A1.
Engineering tasks
Long-horizon search
Scientific research
Instruction following
General agentic tasks
Key Features
Sage-A1.5 introduces new agentic modes and strengthens core capabilities across engineering, scientific research, and general agentic tasks.
Agentic Verification
An opt-in self-review capability that lives inside the model's own chain of thought — the model pauses mid-trajectory to review what it has established and what has failed, then acts on a revised direction in the same turn, with no external critic.
Engineering
Solves complex software engineering tasks end-to-end, from understanding codebases and writing patches to running tests across multiple languages and frameworks.
Scientific Research
Tackles multidisciplinary scientific problems by combining domain reasoning, literature retrieval, data analysis, and evidence verification across physics, chemistry, biology, and beyond.
General Agentic Tasks
Handles diverse long-horizon agent workflows including tool calling, web browsing, instruction following, and multi-step planning in open-ended environments.
Broad agentic benchmark coverage
Sage-A1.5 is evaluated across engineering tasks, long-horizon search, scientific research, instruction following, and general agentic tasks.
Performance Matrix
Sage-A1.5 versus leading agentic systems
Hover over any bar to inspect the model and score. Blue bars mark Sage-A1.5.
Capability showcases
Explore interactive demos showcasing Sage-A1.5's agentic capabilities and scientific applications across diverse real-world scenarios.
01
Foundational Capabilities
The building blocks of reliable scientific agents.
Agentic Orchestration
Delegate tasks as research progresses. Subagents work in parallel, check key assumptions and pool their evidence to reach a comprehensive conclusion.
Agentic Verification
Compare runs with and without Agentic Verification through reflection excerpts, strategy changes and search results.
Local Automated Deployment
Explore hardware deployment guides and a recorded search task showing model execution, tool calls and the final answer.
02
Scientific Information Integration and Long-Horizon Reasoning
Gather information, connect evidence, and work through complex scientific questions.
Deep Research
Follow task planning, parallel execution and revisions through the research pipeline, figure generation and a cited report.
Long-Horizon Scientific Reasoning
Connect retrieval, reasoning and repeated verification through solution traces and conclusions in physics, chemistry and biology.
03
From Scientific Co-working to End-to-End Auto-Research
From a defined scientific brief to an independently executed research investigation.
Scientific Co-working
Explore ten professional workflows across engineering, scientific auditing, planning, data and security, with native tool outputs, recorded decisions and inspectable results.
End-to-End Auto-Research
Follow three studies from problem understanding through analysis to inspectable code, figures and reports.
04
Solving Scientific Problems Aligned by Scientists
Scientific problems, with questions and objectives shaped by scientists.
Neuroscience: Brain Coding Reanalysis
Use statistical checks and anatomical mapping to revisit shared coding structure in existing data.
Computational Chemistry: Molecular Binding
Propose pyridine derivatives, optimize their structures, calculate binding energies and decompose the interactions, then explain the results using the molecular structures.
Condensed Matter Physics: Crystal Structure Refinement
Review refinement candidates with linked structures and spectra, tracking parameter changes and acceptance decisions.
Electrochemistry: Reaction Condition Optimization
Use historical experiments and a causal graph to explore candidate conditions, with five feedback rounds and the best observed conditions.
Fluorescence Spot Detection
Follow eight optimization submissions across nine datasets, then compare reference spots, matched detections and errors in four retained microscopy fields.
Three-stage agent training
The model is trained with a domain-grounded knowledge-action graph that turns agent process traces into trainable targets.
Full-domain supervised fine-tuning
Aligns the base model with broad agentic behaviors across search, engineering, research, tools, and instructions.
Domain-level teacher models
Captures specialized expertise for each domain, giving the final model stronger and more varied supervision.
Multi-teacher on-policy distillation
Transfers expertise across heterogeneous domains with optimization designed for efficient knowledge transfer.
Serve with standard LLM runtimes
Sage-A1.5 can be served through SGLang or vLLM with OpenAI-compatible endpoints
at http://localhost:8000/v1.
SGLang
python -m sglang.launch_server \
--model-path Alpha-Innovator/Agents-A1.5 \
--port 8000 \
--tp-size 1 \
--mem-fraction-static 0.8 \
--context-length 262144 \
--reasoning-parser qwen3 \
--tool-call-parser qwen3_codervLLM
vllm serve Alpha-Innovator/Agents-A1.5 \
--port 8000 \
--tensor-parallel-size 1 \
--max-model-len 262144 \
--reasoning-parser qwen3 \
--enable-auto-tool-choice \
--tool-call-parser qwen3_coderModel, code, and evaluation
Use the model card for release details, the repository for scripts, and the open evaluation framework for reproducible agent capability testing.