Skip to main content

Verified Models

The Verified models list in the offline mode selector is curated: GGUF builds we have tested with Ante's agent loop, downloaded from Hugging Face with memory requirements known up front. Each model carries one of two statuses:

  • Evaled — ran the full public benchmark: same Terminal-Bench 2.1 harness, pinned builds, and auditable Harbor runs as our frontier-model results. Numbers live at antigma.ai/eval.
  • Verified — tested for compatibility: the model loads, the chat template resolves, and tool calling works in Ante's agent loop. Not yet through the full eval.

Models graduate from Verified to Evaled as benchmark runs complete.

Current list

ModelQuantDownloadMax contextStatus
Qwen3.6 27BQ4_K_M17 GB256KEvaled56.2% on Terminal-Bench 2.1
MiniMax-M2.7Q3_K_XL102 GB192KVerified
Qwen3.5 122B-A10BQ2_K45 GB256KVerified
Qwen3.6 35B-A3BQ4_K_M22 GB256KVerified
GLM-4.7-FlashQ4_K_M18 GB198KVerified
Qwen3.5 27BQ4_K_M17 GB256KVerified
Gemma 4 26B-A4B-itQ4_K_M16 GB256KVerified
Devstral-Small-2 24BQ4_K_M14 GB384KVerified
Qwen3.5 9BQ4_K_M5.7 GB256KVerified
Gemma 4 E4B-itQ4_K_M4.7 GB128KVerified

"Max context" is what the model supports; Ante caps the default context window at 32K to keep memory predictable, and you can raise it per model in preferences or with ANTE_OFFLINE_CONTEXT.

The one eval so far: Qwen3.6 27B reached 56.2% accuracy across 89 Terminal-Bench 2.1 tasks with 445 trials — a 17 GB download holding its own on the same benchmark we use for frontier models.

Add your own

Ante merges ~/.ante/verified_models.json on top of the built-in list at startup. Entries with a filename matching a built-in override it; new filenames are appended.

{
"models": [
{
"name": "My Custom Model",
"repo": "username/repo-name",
"filename": "model-Q4_K_M.gguf",
"context_window": 32768,
"file_size_mb": 5000,
"kv_cache_bytes_per_token": 131072,
"support_vision": false
}
]
}
FieldDescription
nameDisplay name in the model selector
repoHugging Face repo the file downloads from
filenameGGUF filename within the repo (also the merge key)
context_windowMaximum context the model supports (tokens)
file_size_mbDownload size, used for memory estimation
kv_cache_bytes_per_tokenKV cache growth per context token, used for memory estimation
support_visionWhether the model accepts image input (optional)