Model Catalog — verified 2026-08-26

Derived from the router captures dataflow.in/endpoints/beast-models.json (28 models) and dataflow.in/endpoints/athena-models.json (32 models), captured 2026-08-26. ctx is the preset ctx-size; ngl is the offload depth (99 = GPU-resident, 0 = CPU). The full lists live in the captures; this document records the curated entries that have execution roles.

Beast (MI25 16 GB GPU + dual Xeon 256 GB CPU)

GPU-resident (ngl 99) — fast, interactive

Model id ctx Role
Qwen/Qwen2.5-Coder-32B-Instruct-GGUF:Q4_K_M 16384 default editing and agent work; strong tool calls
unsloth/Devstral-Small-2-24B-Instruct-2512-GGUF:Q4_K_M 32768 dev-workflow and tool use
unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF:Q4_K_XL 32768 reasoning Q&A; weak tool following — not for the agent loop
unsloth/gemma-4-26B-A4B-it-qat-GGUF:Q4_K_XL 81920 very large context; multimodal-capable

CPU-heavy (ngl 0) — slow, background

Model id ctx Role
unsloth/gpt-oss-120b-GGUF:Q4_K_XL 65536 planning, architecture, large repeated background tasks; runs on CPU (numa = distribute) so the MI25 stays free
unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF:Q4_K_XL 32768 deep reasoning on CPU
unsloth/Llama-3.3-70B-Instruct-GGUF:Q4_K_XL 32768 general 70B chat on CPU

Cache entries with roles (no preset yet, CPU, ctx 32768)

The remaining cache entries on beast are the keep-with-tuning candidates from documents/12-model-retention.md: bartowski/Qwen2.5-Coder-14B-Instruct-GGUF:Q4_K_M, unsloth/Qwen3.5-9B-GGUF:Q4_K_M, unsloth/gemma-4-12B-it-qat-GGUF:Q4_K_XL, unsloth/gemma-4-E2B-it-qat-GGUF:Q4_K_XL, unsloth/gemma-4-E4B-it-qat-GGUF:Q4_K_XL, and unsloth/gpt-oss-20b-GGUF:Q4_K_XL. Promote to presets when a role is assigned (see documents/14-service-tuning-recommendations.md).

Athena (RX 5500 XT 8 GB GPU + Ryzen 9 5900X 128 GB CPU)

GPU-resident (ngl 99) — fast, small

Model id ctx Role
Qwen/Qwen2.5-Coder-1.5B-Instruct-GGUF:Q8_0 8192 inline completion (FIM-capable), tiny fast tasks
Qwen/Qwen2.5-Coder-3B-Instruct-GGUF:Q8_0 8192 inline completion, small edits
unsloth/Llama-3.2-3B-Instruct-GGUF:Q8_0 8192 small chat
unsloth/gemma-4-E4B-it-qat-GGUF:Q4_K_XL 8192 small gemma chat

CPU fallback (ngl 0) — used when beast is offline

Model id ctx Role
Qwen/Qwen2.5-Coder-32B-Instruct-GGUF:Q4_K_M 16384 editing fallback on 128 GB RAM
unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q4_K_XL 16384 reasoning fallback

Special-purpose (cache)

Model id ctx Role
ggml-org/Nomic-Embed-Text-V2-GGUF:Q8_0 8192 embeddings (RAG); needs an embeddings preset
josephmayo/gemma-4-E4B-it-Coder-GGUF:Q5_K_M 8192 coding-tuned small model
unsloth/Qwen3.5-2B-GGUF:Q4_K_M, unsloth/Qwen3.5-4B-GGUF:Q4_K_M 8192 minerva travel candidates

The athena cache also retains Qwen/Qwen2.5-Coder-1.5B-Instruct-GGUF:Q4_K_M and Qwen/Qwen2.5-Coder-3B-Instruct-GGUF:Q4_K_M cache quants in the repos whose Q8 presets are kept (remaining cleanup candidates; owner to decide).

Notes