Record: DSH provider findings — 2026-08-26

Task 03 output. Comparison of the live ~/.dsh/settings.yaml llm-pi-ai.beast section against sources/config/dsh-settings-beast.example.yaml (verified 2026-08-26, router captures refreshed after the llama.cpp 0.3.0-dev rebuild).

Current state (verified 2026-08-26)

Problems in the live section

  1. Deleted models still listed (bare ids of models removed by the 2026-08-26 cache discard; the router no longer lists them and a request would fail or trigger a re-download):
    • bartowski/Qwen3.8-27B-GGUF:Q4_K_M
    • unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF:Q4_K_M
    • unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q4_K_M
    • unsloth/Qwen3.5-27B-GGUF:Q4_K_M, unsloth/Qwen3.5-27B-GGUF:Q4_K_S
    • unsloth/gpt-oss-120b-GGUF:Q4_K_M, unsloth/gpt-oss-20b-GGUF:Q4_K_M
  2. Existing models with bare entries inherit 262144/32768 defaults that exceed every preset: Qwen/Qwen2.5-Coder-32B-Instruct-GGUF:Q4_K_M, unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF:Q4_K_XL, unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF:Q4_K_XL, unsloth/Llama-3.3-70B-Instruct-GGUF:Q4_K_XL, unsloth/Qwen3.5-9B-GGUF:Q4_K_M, unsloth/gemma-4-12B-it-qat-GGUF:Q4_K_XL, unsloth/gemma-4-26B-A4B-it-qat-GGUF:Q4_K_XL, unsloth/gemma-4-E2B/E4B-it-qat-GGUF:Q4_K_XL, unsloth/gpt-oss-120b-GGUF:Q4_K_XL, unsloth/gpt-oss-20b-GGUF:Q4_K_XL — each needs explicit capacities or an alias.
  3. Missing new presets: coder14, devstral (both live on the router) are not in the live list.

Replace the whole beast.models list with the 11 aliased entries in sources/config/dsh-settings-beast.example.yaml (aliases route on the tuned routers, capacities match presets). Optionally add the athena provider route from the same example for low-latency local models. Store BEAST_API_KEY (and ATHENA_API_KEY) in the DSH Models settings.

Verification