ARGUS AI TEAM

Building persistent AI systems that extend human capability

Long-horizon agents · AI systems in silicon · Model deployment · Mathematical research

Argus AI Team is an independent research and engineering team building AI systems that can pursue substantial work beyond a single model response.

Our work spans the full path from autonomous agent runtimes to verified RTL, practical deployment of frontier multimodal models, and inspectable mathematical research. Across these domains, we focus on one principle: capability should be accompanied by evidence.

We believe AI should not diminish human agency. It should expand what researchers, engineers, and creators are able to understand, build, and verify.

Argus separates campaign control, planning, execution, and acceptance across distinct agent roles. Durable project state preserves tasks, checkpoints, decisions, skills, and evidence across sessions and runtime changes.

The accompanying technical report records approximately 78% on SWE-Bench Pro, compared with 59% for Direct Copilot, at 1.41× aggregate token use. Six paper pipelines completed 254 missions, including 16 evidence-driven stage rollbacks.

ACE-2 is an evidence-first Qwen2.5-0.5B W4A8 accelerator. Its public record includes:

• 18/18 validated Layer-0 fixed-point operator boundaries;
• 13,914/13,914 runtime commands across a demonstrated 24-layer, two-token path;
• a mapped SKY130 result of 62,283 cells and 0.614 mm² non-SRAM area;
• a 100 MHz target with positive setup slack.

ACE-3 advances the architecture toward native asymmetric AWQ W4A16 execution. The accepted 24-layer fixture consumes all 624/624 official decoder tensors. Its current public scope is pre-synthesis RTL evidence; measured silicon, FPGA, area, power, and latency claims are deliberately left outside that boundary.

Our MiniMax-H3 projects make complete video-and-stereo-audio generation available on hardware outside datacenter-scale deployments:

• MiniMax-H3 for Mac runs the documented generation path on an Apple M4 Pro with 24 GB unified memory through streamed model execution.
• MiniMax-H3 Desktop runs the full FL2VA stack on one RTX A6000 and publishes prompts, outputs, matched benchmarks, structural audio/video checks, and system telemetry.
• ComfyUI MiniMax-H3 MLX exposes model loading, generation, conversion, and direct MP4 output through a visual node workflow on Apple Silicon.

Argus Mathematics currently preserves nine public result packages across Riemannian and algebraic geometry, graph theory, convex geometry, arithmetic dynamics, braid-group algebra, and smooth four-manifold topology.

The archive includes five independently runnable verification paths, one Lean-checked logical composition, technical reports, review packages, certificates, and checksums for 45 public artifacts. It explicitly distinguishes original constructions from literature reconstructions, historical negative results, scope corrections, and claims whose novelty has not yet been certified.

We aim to make ambitious work easy to inspect and difficult to overstate.

• Evidence before headline. Results are linked to reports, commands, certificates, measurements, or reproducible artifacts.
• Scope before scale. We state what a result demonstrates and what it does not yet establish.
• Review separate from execution. Acceptance is based on artifacts and checks rather than the producing agent's confidence.
• Negative results remain visible. Failed routes and bounded evidence help prevent repeated mistakes and unsupported claims.
• Corrections are part of the record. Public work should become more precise as stronger evidence arrives.