Model Retention Analysis — verified 2026-08-26

Purpose: note which downloaded models are likely not useful, so the owner can decide. Nothing is discarded without the owner’s decision. The first model set was downloaded by interest, then benchmarked (see sources/models/_combine-amd-rx-5500.md and sources/models/logs/). GGUF files live in the HuggingFace cache on each host (HF_HUB_CACHE=~/.cache/huggingface/hub on beast; default cache on athena).

Conventions (owner, 2026-09-03)

Full archive: everything ever downloaded is backed up by rsyncing the HF cache to /backups/huggingface/ on each host (large spinning storage, e.g. vg2-backups 4 T); sources/models/backup.sh performs the copy. The archive is not pruned by this project — a discarded model can always be restored from it (verified: bartowski/Qwen3.8-27B-GGUF = 18 G on each host’s backup).
Working cache stays small: the live HF_HUB_CACHE holds only models that make sense to run on that host (roles below).
Download once: a model wanted on both hosts is rsynced host-to-host from the host that has it (or restored from /backups/huggingface/), never downloaded twice from the web.

Discard executed 2026-08-26

The owner approved the discard on 2026-08-26; the files below marked “Discard candidates” were deleted from the HuggingFace cache on beast and athena by scripts/cache-discard.sh. Backups exist at /backups/huggingface/ on both hosts (sources/models/backup.sh) and are not managed by this project; a re-download is also possible if a model is needed later. Beast freed ~163 GB (28 → 13 models); athena freed ~306 GB (32 → 15 models). Services were restarted to rescan the caches. One incident during execution: the blob-cleanup step of the first script run deleted keeper quants (gpt-oss-120b/gpt-oss-20b UD-Q4_K_XL on beast) because stat -c %i did not dereference symlinks; the keepers were restored over LAN from athena before athena’s own discard ran. See records/2026-08-26-03-cache-discard.md.

Classes: keep (preset, has a role), keep-with-tuning (cache entry that should become a preset or serve a role), discard-candidate (redundant or roleless — deleted 2026-08-26).

Beast (28 models)

Keep (presets, have roles)

Model id	ctx	class	role
`Qwen/Qwen2.5-Coder-32B-Instruct-GGUF:Q4_K_M`	16384	GPU	editing, agent
`unsloth/Devstral-Small-2-24B-Instruct-2512-GGUF:Q4_K_M`	32768	GPU	dev workflows
`unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF:Q4_K_XL`	32768	GPU	reasoning Q&A
`unsloth/gemma-4-26B-A4B-it-qat-GGUF:Q4_K_XL`	81920	GPU	large context
`unsloth/gpt-oss-120b-GGUF:Q4_K_XL`	65536	CPU	planning, background
`unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF:Q4_K_XL`	32768	CPU	deep reasoning
`unsloth/Llama-3.3-70B-Instruct-GGUF:Q4_K_XL`	32768	CPU	general 70B

Keep-with-tuning (cache → promote to presets; all fit the MI25 GPU)

Model id	size	note
`bartowski/Qwen2.5-Coder-14B-Instruct-GGUF:Q4_K_M`	~9 GB	14B coder; strong candidate for a GPU preset between 3B and 32B
`unsloth/Qwen3.5-9B-GGUF:Q4_K_M`	~5.5 GB	general 9B; fits GPU
`unsloth/gpt-oss-20b-GGUF:Q4_K_XL`	~13 GB	llama.vscode’s recommended agent model; fits GPU
`unsloth/gemma-4-12B-it-qat-GGUF:Q4_K_XL`	~8 GB	fits GPU
`unsloth/gemma-4-E2B-it-qat-GGUF:Q4_K_XL`	2.43 GB	keep for minerva (travel) decision
`unsloth/gemma-4-E4B-it-qat-GGUF:Q4_K_XL`	3.91 GB	keep for minerva (travel) decision

Discard candidates (redundant or roleless on beast) — deleted 2026-08-26

Model id	reason
`Qwen/Qwen2.5-Coder-1.5B-Instruct-GGUF:Q4_K_M`	athena serves the Q8 preset; beast copy redundant
`Qwen/Qwen2.5-Coder-1.5B-Instruct-GGUF:Q8_0`	same; third 1.5B coder copy
`ggml-org/Qwen2.5-Coder-1.5B-Q8_0-GGUF:Q8_0`	same; third 1.5B coder copy
`Qwen/Qwen2.5-Coder-3B-Instruct-GGUF:Q4_K_M`	athena serves the Q8 preset; beast copy redundant
`Qwen/Qwen2.5-Coder-3B-Instruct-GGUF:Q8_0`	same
`unsloth/Llama-3.2-1B-Instruct-GGUF:Q8_0`	no role
`unsloth/Llama-3.2-3B-Instruct-GGUF:Q8_0`	athena serves the preset; beast copy redundant
`bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M`	no role; Qwen coders dominate for code
`bartowski/Qwen3.8-27B-GGUF:Q4_K_M`	27B family: three near-duplicates, none fits the GPU, CPU-only and slow
`unsloth/Qwen3.5-27B-GGUF:Q4_K_M`	27B family (keep at most one)
`unsloth/Qwen3.5-27B-GGUF:Q4_K_S`	27B family (keep at most one)
`unsloth/gpt-oss-120b-GGUF:Q4_K_M`	the `Q4_K_XL` preset is the keeper; duplicate quant
`unsloth/gpt-oss-20b-GGUF:Q4_K_M`	the `Q4_K_XL` is the keeper (promote it); duplicate quant
`unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF:Q4_K_M`	the `Q4_K_XL` preset is the keeper; duplicate quant
`unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q4_K_M`	athena serves it as CPU fallback preset; beast copy redundant

Athena (32 models)

Keep (presets)

Qwen/Qwen2.5-Coder-1.5B-Instruct-GGUF:Q8_0, Qwen/Qwen2.5-Coder-3B-Instruct-GGUF:Q8_0, unsloth/Llama-3.2-3B-Instruct-GGUF:Q8_0, unsloth/gemma-4-E4B-it-qat-GGUF:Q4_K_XL (GPU), Qwen/Qwen2.5-Coder-32B-Instruct-GGUF:Q4_K_M, unsloth/DeepSeek-R1-Distill-Qwen-32B-GGUF:Q4_K_XL (CPU fallback when beast is offline).

Keep-with-tuning

Model id	note
`ggml-org/Nomic-Embed-Text-V2-GGUF:Q8_0`	embeddings for RAG; needs an embeddings preset (`/v1/embeddings`)
`josephmayo/gemma-4-E4B-it-Coder-GGUF:Q5_K_M`	coding-tuned small model; candidate for athena completion
`unsloth/Qwen3.5-2B-GGUF:Q4_K_M`, `unsloth/Qwen3.5-4B-GGUF:Q4_K_M`	keep for minerva (travel) decision

Discard candidates (redundant or roleless on athena) — deleted 2026-08-26

The 70B/120B/27B and big GPU cache entries on athena run at ~1 t/s on the 5900X and duplicate beast presets:

unsloth/DeepSeek-R1-Distill-Llama-70B-GGUF:Q4_K_XL, unsloth/Llama-3.3-70B-Instruct-GGUF:Q4_K_XL, unsloth/gpt-oss-120b-GGUF:Q4_K_M/Q4_K_XL — beast is their home.
unsloth/gpt-oss-20b-GGUF:Q4_K_M/Q4_K_XL, unsloth/Qwen3.5-27B-GGUF:Q4_K_M/Q4_K_S, bartowski/Qwen3.8-27B-GGUF:Q4_K_M, unsloth/Devstral-Small-2-24B-Instruct-2512-GGUF:Q4_K_M, unsloth/gemma-4-26B-A4B-it-qat-GGUF:Q4_K_XL, unsloth/gemma-4-12B-it-qat-GGUF:Q4_K_XL, unsloth/DeepSeek-R1-Distill-Qwen-14B-GGUF:Q4_K_M — beast presets are authoritative; athena copies run CPU-only.
bartowski/Mistral-Nemo-Instruct-2407-GGUF:Q4_K_M — no role.
unsloth/DeepSeek-R1-Distill-Qwen-1.5B-GGUF:Q4_K_XL, unsloth/Llama-3.2-1B-Instruct-GGUF:Q4_K_M/Q8_0 — no role.
unsloth/Llama-3.2-3B-Instruct-GGUF:Q4_K_M — the Q8 preset is the keeper.

Remaining cleanup candidates — kept by decision (2026-08-26)

on athena: Qwen/Qwen2.5-Coder-1.5B-Instruct-GGUF:Q4_K_M and Qwen/Qwen2.5-Coder-3B-Instruct-GGUF:Q4_K_M cache quants remain in repos whose Q8 presets are kept; and the unused ggml-org/Qwen2.5-Coder-1.5B-Q8_0-GGUF repo remains on disk.
Owner closed the retention question on 2026-08-26: these leftovers are kept as-is (harmless on disk; usable as lighter fallbacks, and the 3B/1.5B Q4_K_M quants are useful for a weak laptop). No further deletion planned.

Benchmark reference

From sources/models/_combine-amd-rx-5500.md (athena, 2026-07-28), generation t/s (tg128):

Model	t/s
Qwen-2.5-Coder-1.5B Q8 (GPU)	98.0
Qwen-2.5-Coder-3B Q8 (GPU)	53.7
Gemma-4-E4B QAT (GPU)	45.7
Llama-3.2-3B Q8 (GPU)	29.8
GPT-OSS-120B (CPU)	8.42
Gemma-4-E2B QAT (CPU)	21.7
Llama-3.3-70B (CPU)	0.91
DeepSeek-R1-Distill-70B (CPU)	0.91

Notes: the 120B at 8.42 t/s on CPU versus 0.91 t/s for the 70B models is the signature of MTP speculative decoding (the gpt-oss architecture has a native MTP head); verify with --spec-type none versus draft-mtp in task 01. See documents/13-mtp-and-rag.md.

Update 2026-09-03 — qwen27 re-promoted on athena

bartowski/Qwen3.8-27B-GGUF:Q4_K_M (discard candidate 2026-08-26: “none fits the GPU, CPU-only and slow”) is re-promoted as the athena qwen27 preset — athena’s GPU was fixed (records/2026-09-03-01-athena-gpu-fixes-qwen27-llamacpp-current.md), making an 8 GB-VRAM GPU + CPU split viable. Correction 2026-09-03 (records/2026-09-03-02-athena-qwen27-load-fixes.md): the preset resolves the batiai repo; the /backups/huggingface/ copy is bartowski (a different repo — the backup stays archive-only and the batiai files were web-downloaded). It remains a discard candidate on beast (beast’s Qwen3.8 role is the qwen38 Flash-Next CPU preset, which has no backup anywhere and needs a first-time ~119 G download).