Antares

Cisco Foundation AI · analysed 2026-08-10 · skill v0.4.0

dense full-attention text-only open-weights medium domain-specific

Design overview

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

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

Influence out

No descendants recorded yet.

Notable omissions

Closest relatives in the atlas

Spec sheet

Identity

FamilyAntares[1]
OrganisationCisco Foundation AI
Report URLhttps://cisco-foundation-ai.github.io/antares/technical-report.pdf[2]
Open weightsyes[3]

Variants

Scale

Total parameters1B[4]
Active parameters / token1B
Embedding tyingyes[5]

Per-variant configs

VariantTotalLayersHiddenFFNQ headsKV headsContext
350M350Mnot disclosednot disclosednot disclosednot disclosednot disclosed32,768
1B1Bnot disclosednot disclosednot disclosednot disclosednot disclosed131,072
3B3Bnot disclosednot disclosednot disclosednot disclosednot disclosed131,072

Core block

Block typedense
Attention variantGQA[6]
Attention layer patternuniform
Depth mixingsequential-residual
ActivationSwiGLU[7]
MethodRoPE[7]
TypeRMSNorm[7]

Context

Trained context length131,072[8]
Deployed context length131,072

Tokenizer

Training

Curriculum / stagingnone disclosed[2]
OptimizerAdamW[9]
LR scheduleCosine, peak 5e-5[9]
ParallelismFSDP[9]
HardwareSingle 8×H100 node[9]

Post-training

SFTyes[10]
Preference optimisationGRPO[11]
Reasoning trainingGRPO with verifiable multi-component rewards over terminal vulnerability-localization trajectories[12]
Distillationnone disclosed[2]

Modality

Typetext-only

Inference efficiency

Quantisation shippednone disclosed[2]
Speculative decodingnone disclosed[2]
Serving optimisationsnone disclosed[2]

Evaluation

Benchmarks (report's own numbers only)

BenchmarkValueRef
VLoc Bench File F1 (Antares-1B)0.209eval tables
VLoc Bench File F1 (Antares-3B)0.223eval tables
VLoc Bench File F1 (Antares-350M)0.135eval tables
VLoc Bench File F1 (GPT-5.5 xhigh)0.229eval tables

Provenance references

  1. stated · Title
  2. stated · pipeline
  3. stated · Abstract: 'publicly release Antares-350M and Antares-1B'
  4. stated · name convention (Antares-1B; report leads with 350M and 1B releases)
  5. stated · architecture blueprint (tied embeddings)
  6. stated · architecture blueprint (Granite 4.0: GQA)
  7. stated · architecture blueprint
  8. stated · context section (1B variant: 128K)
  9. stated · training section
  10. stated · post-training section (SFT, 1 epoch, semantic-conditioning auxiliary objective)
  11. stated · post-training section (verifiable multi-component rewards, no reward model, KL vs SFT reference, veRL+vLLM)
  12. stated · post-training section