Antares is not a new architecture: it is a post-training recipe applied to IBM Granite 4.0 checkpoints at 350M, 1B, and 3B, released openly for agentic vulnerability localization. The report states the inherited blueprint — GQA attention, SwiGLU, RMSNorm, RoPE, tied embeddings — but discloses no layer counts, dimensions, or head numbers, so the scale section records only what the report says (sizes, context lengths of 32K/128K/128K). The contribution is entirely in the recipe: one epoch of SFT with a semantic-conditioning auxiliary objective on terminal trajectories, then GRPO with verifiable multi-component rewards (no reward model) using veRL+vLLM under FSDP on a single 8×H100 node, with KL regularization against the SFT reference. Evaluation is on the in-house VLoc Bench file-F1 metric, where Antares-1B reaches 0.209 against 0.229 for a much larger frontier model.
What the report claims is novel
Agentic vulnerability localization with open 350M/1B models [Abstract]
Semantic-conditioning auxiliary objective during SFT [post-training section]
GRPO with verifiable multi-component rewards, no reward model [post-training section]
Atlas assessment: Architecture novelty ≈ 0 — Antares is an SFT+GRPO post-training of IBM Granite 4.0 checkpoints (the block section records only the Granite blueprint names, everything else unknown). The contribution is the recipe: terminal-trajectory fine-tuning with semantic conditioning and verifiable-reward GRPO for security tasks. The atlas's thinnest architecture record, honestly so.
Relation to the atlas
Lineage in
Base checkpoints (350M/1B/3B from Granite 4.0) from IBM Granite 4.0
GRPO with verifiable rewards from GRPO line (DeepSeekMath et al.)
Influence out
No descendants recorded yet.
Notable omissions
No architecture table: layers, hidden dims, heads, vocab, tokenizer all undisclosed