chore(model-gallery): propose variant groupings for review

[localai-org-maint-bot]

Proposed gallery variant groupings

This is a proposal, not a decision. The gallery agent adds one build per model and never joins an existing family, so entries that are alternative builds of the same weights drift apart as the gallery grows. This job re-applies the grouping heuristics from the manual sweeps and asks a human to confirm.

Each family below lists the parent, the variants, and the evidence that they are the same weights. Reject anything whose evidence you do not believe.

To decline a family permanently, add one line to gallery/variant-exclusions.yaml in this pull request and close it:

pairs:
  - {parent: some-model, variant: some-model-thing, reason: "different finetune"}

Proposed families (4)

kimodo-g1-rp

variant	signals	evidence
kimodo-g1-rp-bf16	name-modulo-quant, weight-filename	same name once quantization markers are stripped: kimodo-g1-rp; same primary weight filename once quantization markers are stripped: kimodo-g1-rp-v1; same upstream repo localai-io/kimodo-g1-rp-v1-ggml; differing quantization tokens: bf16, f32, g1
kimodo-g1-rp-q4_k	name-modulo-quant, weight-filename	same name once quantization markers are stripped: kimodo-g1-rp; same primary weight filename once quantization markers are stripped: kimodo-g1-rp-v1; same upstream repo localai-io/kimodo-g1-rp-v1-ggml; differing quantization tokens: f32, g1, q4_k
kimodo-g1-rp-q4_k_m	name-modulo-quant, weight-filename	same name once quantization markers are stripped: kimodo-g1-rp; same primary weight filename once quantization markers are stripped: kimodo-g1-rp-v1; same upstream repo localai-io/kimodo-g1-rp-v1-ggml; differing quantization tokens: f32, g1, q4_k_m
kimodo-g1-rp-q5_k	name-modulo-quant, weight-filename	same name once quantization markers are stripped: kimodo-g1-rp; same primary weight filename once quantization markers are stripped: kimodo-g1-rp-v1; same upstream repo localai-io/kimodo-g1-rp-v1-ggml; differing quantization tokens: f32, g1, q5_k
kimodo-g1-rp-q6_k	name-modulo-quant, weight-filename	same name once quantization markers are stripped: kimodo-g1-rp; same primary weight filename once quantization markers are stripped: kimodo-g1-rp-v1; same upstream repo localai-io/kimodo-g1-rp-v1-ggml; differing quantization tokens: f32, g1, q6_k

kimodo-g1-seed

variant	signals	evidence
kimodo-g1-seed-bf16	name-modulo-quant, weight-filename	same name once quantization markers are stripped: kimodo-g1-seed; same primary weight filename once quantization markers are stripped: kimodo-g1-seed-v1; same upstream repo localai-io/kimodo-g1-seed-v1-ggml; differing quantization tokens: bf16, f32, g1
kimodo-g1-seed-q4_k	name-modulo-quant, weight-filename	same name once quantization markers are stripped: kimodo-g1-seed; same primary weight filename once quantization markers are stripped: kimodo-g1-seed-v1; same upstream repo localai-io/kimodo-g1-seed-v1-ggml; differing quantization tokens: f32, g1, q4_k
kimodo-g1-seed-q4_k_m	name-modulo-quant, weight-filename	same name once quantization markers are stripped: kimodo-g1-seed; same primary weight filename once quantization markers are stripped: kimodo-g1-seed-v1; same upstream repo localai-io/kimodo-g1-seed-v1-ggml; differing quantization tokens: f32, g1, q4_k_m
kimodo-g1-seed-q5_k	name-modulo-quant, weight-filename	same name once quantization markers are stripped: kimodo-g1-seed; same primary weight filename once quantization markers are stripped: kimodo-g1-seed-v1; same upstream repo localai-io/kimodo-g1-seed-v1-ggml; differing quantization tokens: f32, g1, q5_k
kimodo-g1-seed-q6_k	name-modulo-quant, weight-filename	same name once quantization markers are stripped: kimodo-g1-seed; same primary weight filename once quantization markers are stripped: kimodo-g1-seed-v1; same upstream repo localai-io/kimodo-g1-seed-v1-ggml; differing quantization tokens: f32, g1, q6_k

kimodo-soma-rp

variant	signals	evidence
kimodo-soma-rp-bf16	name-modulo-quant, weight-filename	same name once quantization markers are stripped: kimodo-soma-rp; same primary weight filename once quantization markers are stripped: kimodo-soma-rp-v1.1; same upstream repo localai-io/kimodo-soma-rp-v1.1-ggml; differing quantization tokens: bf16, f32
kimodo-soma-rp-q4_k	name-modulo-quant, weight-filename	same name once quantization markers are stripped: kimodo-soma-rp; same primary weight filename once quantization markers are stripped: kimodo-soma-rp-v1.1; same upstream repo localai-io/kimodo-soma-rp-v1.1-ggml; differing quantization tokens: f32, q4_k
kimodo-soma-rp-q4_k_m	name-modulo-quant, weight-filename	same name once quantization markers are stripped: kimodo-soma-rp; same primary weight filename once quantization markers are stripped: kimodo-soma-rp-v1.1; same upstream repo localai-io/kimodo-soma-rp-v1.1-ggml; differing quantization tokens: f32, q4_k_m
kimodo-soma-rp-q5_k	name-modulo-quant, weight-filename	same name once quantization markers are stripped: kimodo-soma-rp; same primary weight filename once quantization markers are stripped: kimodo-soma-rp-v1.1; same upstream repo localai-io/kimodo-soma-rp-v1.1-ggml; differing quantization tokens: f32, q5_k
kimodo-soma-rp-q6_k	name-modulo-quant, weight-filename	same name once quantization markers are stripped: kimodo-soma-rp; same primary weight filename once quantization markers are stripped: kimodo-soma-rp-v1.1; same upstream repo localai-io/kimodo-soma-rp-v1.1-ggml; differing quantization tokens: f32, q6_k

kimodo-soma-seed

variant	signals	evidence
kimodo-soma-seed-bf16	name-modulo-quant, weight-filename	same name once quantization markers are stripped: kimodo-soma-seed; same primary weight filename once quantization markers are stripped: kimodo-soma-seed-v1.1; same upstream repo localai-io/kimodo-soma-seed-v1.1-ggml; differing quantization tokens: bf16, f32
kimodo-soma-seed-q4_k	name-modulo-quant, weight-filename	same name once quantization markers are stripped: kimodo-soma-seed; same primary weight filename once quantization markers are stripped: kimodo-soma-seed-v1.1; same upstream repo localai-io/kimodo-soma-seed-v1.1-ggml; differing quantization tokens: f32, q4_k
kimodo-soma-seed-q4_k_m	name-modulo-quant, weight-filename	same name once quantization markers are stripped: kimodo-soma-seed; same primary weight filename once quantization markers are stripped: kimodo-soma-seed-v1.1; same upstream repo localai-io/kimodo-soma-seed-v1.1-ggml; differing quantization tokens: f32, q4_k_m
kimodo-soma-seed-q5_k	name-modulo-quant, weight-filename	same name once quantization markers are stripped: kimodo-soma-seed; same primary weight filename once quantization markers are stripped: kimodo-soma-seed-v1.1; same upstream repo localai-io/kimodo-soma-seed-v1.1-ggml; differing quantization tokens: f32, q5_k
kimodo-soma-seed-q6_k	name-modulo-quant, weight-filename	same name once quantization markers are stripped: kimodo-soma-seed; same primary weight filename once quantization markers are stripped: kimodo-soma-seed-v1.1; same upstream repo localai-io/kimodo-soma-seed-v1.1-ggml; differing quantization tokens: f32, q6_k

Declined by the ledger (9)

Candidates the heuristics found and the ledger has already settled. They are listed so the ledger's effect stays visible rather than silently shrinking the job's output.

• kokoros + kokoros-cmn: different languages: one Kokoro release per language, not alternative builds of one model
• kokoros + kokoros-de: different languages: one Kokoro release per language, not alternative builds of one model
• kokoros + kokoros-ja: different languages: one Kokoro release per language, not alternative builds of one model
• kokoros-cmn + kokoros-de: different languages: one Kokoro release per language, not alternative builds of one model
• kokoros-ja + kokoros-cmn: different languages: one Kokoro release per language, not alternative builds of one model
• kokoros-ja + kokoros-de: different languages: one Kokoro release per language, not alternative builds of one model
• llama-3.2-1b-instruct:q4_k_m + bert-embeddings: different parameter sizes (segment "1b")
• mistralai_mistral-small-3.1-24b-instruct-2503 + mistralai_mistral-small-3.1-24b-instruct-2503-multimodal: different capability surface: the multimodal entry adds a vision projector
• whisper-1 + whisper-base: alias: identical payload, whisper-1 is the name OpenAI-compatible clients send

Found but refused (12)

Candidates the heuristics found but the authoring rules would not let this job write. They need a human edit or a rule change.

• gemma-4-12b-it-q4 + gemma-4-12b-it-q5 + gemma-4-12b-it-q6 + gemma-4-12b-it-q8 + gemma-4-12b-it-qat-q4_0: the parent "gemma-4-12b-it-q4" defines YAML anchor &gemma-4-12b-it, and a variants key added there is inherited by the 3 entries that merge it (gemma-4-12b-it-q5, gemma-4-12b-it-q6, gemma-4-12b-it-q8). Grouping this family by hand also means adding an explicit variants: [] to each of those entries
• glm-5.3-flash + glm-5.3-flash-q4 + glm-5.3-flash-q8: "glm-5.3-flash-q4" already offers variants of its own, so it cannot itself be a variant target
• granite-4.2-30b-q4 + granite-4.2-30b-q8 + granite-4.2-30b:vllm: the parent "granite-4.2-30b-q4" defines YAML anchor &granite-4-2-30b, and a variants key added there is inherited by the 1 entries that merge it (granite-4.2-30b-q8). Grouping this family by hand also means adding an explicit variants: [] to each of those entries
• granite-4.2-3b-q4 + granite-4.2-3b-q8 + granite-4.2-3b:vllm: the parent "granite-4.2-3b-q4" defines YAML anchor &granite-4-2-3b, and a variants key added there is inherited by the 1 entries that merge it (granite-4.2-3b-q8). Grouping this family by hand also means adding an explicit variants: [] to each of those entries
• granite-4.2-8b-q4 + granite-4.2-8b-q8 + granite-4.2-8b:vllm: the parent "granite-4.2-8b-q4" defines YAML anchor &granite-4-2-8b, and a variants key added there is inherited by the 1 entries that merge it (granite-4.2-8b-q8). Grouping this family by hand also means adding an explicit variants: [] to each of those entries
• ornith-1.0-9b-q4 + ornith-1.0-9b-q8 + ornith-1.0-9b: "ornith-1.0-9b-q4" already offers variants of its own, so it cannot itself be a variant target
• ornith-1.5-35b-a3b-apex + ornith-1.5-35b-a3b-q4 + ornith-1.5-35b-a3b-q8 + ornith-1.5-35b-a3b-q5 + ornith-1.5-35b-a3b-q6: the parent "ornith-1.5-35b-a3b-q4" defines YAML anchor &ornith-1-5-35b-a3b, and a variants key added there is inherited by the 3 entries that merge it (ornith-1.5-35b-a3b-q8, ornith-1.5-35b-a3b-q5, ornith-1.5-35b-a3b-q6). Grouping this family by hand also means adding an explicit variants: [] to each of those entries
• ornith-1.5-9b-uncensored + ornith-1.5-9b-uncensored-q4 + ornith-1.5-9b-uncensored-q6 + ornith-1.5-9b-uncensored-q8: "ornith-1.5-9b-uncensored-q4" already offers variants of its own, so it cannot itself be a variant target
• qwen3.8-27b + qwen3.8-27b-q4 + qwen3.8-27b-q8: "qwen3.8-27b-q4" already offers variants of its own, so it cannot itself be a variant target
• qwen3.8-27b-uncensored + qwen3.8-27b-uncensored-q4 + qwen3.8-27b-uncensored-q8: "qwen3.8-27b-uncensored-q4" already offers variants of its own, so it cannot itself be a variant target
• spark-x2.5-4b + spark-x2.5-4b-q4 + spark-x2.5-4b-q6 + spark-x2.5-4b-q8: "spark-x2.5-4b-q4" already offers variants of its own, so it cannot itself be a variant target
• whisper-1 + whisper-base-q5_1: the natural parent "whisper-base-q5_1" is already a variant of "whisper-base"; proposing it as a parent would nest variants

---

Opened by .github/ci/variantproposals. Heuristics and the rejection ledger live there and in the ledger file; a wrong proposal is a bug in one of the two.