Antares Models Released: 3B Parameter Rivals GPT-5.5 with Fast Inference on Single H100
aminkarbasi · x · 2026-08-06
Cisco and other institutions have released Antares, a family of foundation models (350M, 1B, and 3B parameters) designed for agentic vulnerability localization.
- Training Pipeline: Based on IBM Granite base models, it uses a two-stage pipeline combining supervised fine-tuning on cybersecurity reasoning and repository exploration data, followed by reinforcement learning (RL) with verifiable rewards over vulnerable repositories.
- Performance: Antares-3B approaches GPT-5.5 and outperforms open-weight models over 200x larger.
- Inference Cost: Enables fast, low-cost local inference, completing a 500-task evaluation sweep in about 15 minutes on a single H100 GPU, costing less than $0.002 per task.
More from Models
- Zuck Teases He Will 'Share More on Open Source' Soon — realmvp77 · 2026-08-06
- Anthropic Pauses Plan to Move Third-Party Apps Off Subscription Limits for 7 Weeks — Deep_Ad1959 · 2026-08-06
- Rant: Models Wasting Tokens on Security Hacks Ruin the Actual Work Experience — mattrickard · 2026-08-06
- Meta's AI Model Accidentally Hacked Another Company During Testing — Simon Willison · 2026-08-06
- Benchmarking Fallback Models for Agents: Why Failure Visibility Beats Raw Quality — AccomplishedLab3697 · 2026-08-06
- Zuckerberg Hints Meta's Muse Code Model Could Be Open-Sourced Soon — ns123abc · 2026-08-06