# self model registry. # # Public, human-edited list of the models `self` can serve. Every entry: # # :: # description: one line (shown by `self ls-remote` and on the overview page) # readme: model card, Markdown file relative to this file (readmes//.md) # type: decision | audio | text | embedding | image # default: quant used when --quant is not given # capabilities: text | vision | multi-image | choice | score | noul | number | excerpt | cite | audio | embedding | multi-embedding | image | image-edit # info: descriptive metadata (family, parameters, architecture, # base_model, source, license, languages, context_length, # max_options, homepage). max_options also caps requests. # settings: engine parameters for the adapter, e.g. context_size, # threads, temperature (see dev-docs/11-onboarding.md). Can be # overridden per quant (`settings:` under the quant) and # at run time with `self serve --set key=value`. # : # adapter: engine adapter (ggmlc-laya | ggmlc-custom-decider | ggmlc-audio | llama-server | sd-server) # repo: Hugging Face repo (owner/name) # files: - {file, size, sha256[, role: model|mmproj|voice][, repo]} # role voice (audio only): .wav/.mp3 reference voice used # when a request has none. repo: per-file repo override. # role mmproj (text and decision): vision projector. # settings: optional per-quant settings (override the model's) # # size (bytes) and sha256 pin the exact upstream file. A download that does # not match is deleted and never used; if the upstream file changes, `self` # refuses to download it until this file is updated. # # Image pipelines keep diffusion weights in GGUF. Image-only vae and # text-encoder roles also accept safetensors; mmproj remains GGUF. version: 1 models: # --- Images (official upstream stable-diffusion.cpp sd-server) --- qwen-image-2.1:7b: description: "Qwen Image 2.1 7B: generation, transparent images and multi-reference editing" readme: readmes/qwen-image-2.1/7b.md type: image default: int8 capabilities: input: [text, image, multi-image] output: [image, image-edit] max_images: 8 info: family: qwen-image-2.1 parameters: 7B architecture: qwen-image-2.1 base_model: Qwen/Qwen-Image-2.1 source: Qwen/Qwen-Image-2.1 license: qwen-research homepage: "https://huggingface.co/Qwen/Qwen-Image-2.1" settings: cfg_scale: 6.0 sampling_method: euler flash_attention: true offload_to_cpu: true steps: 40 int8: adapter: sd-server repo: Comfy-Org/Qwen-Image-2.1 files: - {file: diffusion_models/qwen_image_2.1_int8_convrot.safetensors, revision: cb504a4090723e43f17ad01cec0359490e2de613, size: 7256783064, sha256: cb74113cb03faecd79611b01fd7fd642f0aa60d6f0b95086abee214d75eaa57d} - {file: vae/qwen_image_2.1_vae_bf16.safetensors, role: vae, repo: Comfy-Org/Qwen-Image-2.1, revision: cb504a4090723e43f17ad01cec0359490e2de613, size: 675509688, sha256: bb21f7473051e1ac368515dd3f2e15cd44d7a11748ee8823e1ddca3e4876b7c9} - {file: Qwen3VL-8B-Instruct-Q4_K_M.gguf, role: text-encoder, repo: Qwen/Qwen3-VL-8B-Instruct-GGUF, revision: f982a07559d4a2f6c8744d840bf6fccab30eea96, size: 5027784800, sha256: 67d1659bfe71b89d50b45a4ad1a9e5b997e5bb16ce5da66a6a6167abd569e9e2} - {file: mmproj-Qwen3VL-8B-Instruct-F16.gguf, role: mmproj, repo: Qwen/Qwen3-VL-8B-Instruct-GGUF, revision: f982a07559d4a2f6c8744d840bf6fccab30eea96, size: 1159029824, sha256: ca524100ebf825c9a870db1c580d03879e0da0ab2541697e2458e64891cf9d38} q4: adapter: sd-server repo: unsloth/Qwen-Image-2.1-GGUF files: - {file: qwen-image-2.1-Q4_K_M.gguf, revision: 2c31ccd392b367a6637841a143813320a02dff55, size: 4199565024, sha256: 631d532e7ca71e8d90a87c71d3699761a812039d22e3370e87498d87754660fe} - {file: vae/qwen_image_2.1_vae_bf16.safetensors, role: vae, repo: Comfy-Org/Qwen-Image-2.1, revision: cb504a4090723e43f17ad01cec0359490e2de613, size: 675509688, sha256: bb21f7473051e1ac368515dd3f2e15cd44d7a11748ee8823e1ddca3e4876b7c9} - {file: Qwen3VL-8B-Instruct-Q4_K_M.gguf, role: text-encoder, repo: Qwen/Qwen3-VL-8B-Instruct-GGUF, revision: f982a07559d4a2f6c8744d840bf6fccab30eea96, size: 5027784800, sha256: 67d1659bfe71b89d50b45a4ad1a9e5b997e5bb16ce5da66a6a6167abd569e9e2} - {file: mmproj-Qwen3VL-8B-Instruct-F16.gguf, role: mmproj, repo: Qwen/Qwen3-VL-8B-Instruct-GGUF, revision: f982a07559d4a2f6c8744d840bf6fccab30eea96, size: 1159029824, sha256: ca524100ebf825c9a870db1c580d03879e0da0ab2541697e2458e64891cf9d38} q8: adapter: sd-server repo: unsloth/Qwen-Image-2.1-GGUF files: - {file: qwen-image-2.1-Q8_0.gguf, revision: 2c31ccd392b367a6637841a143813320a02dff55, size: 7640860384, sha256: c0ed4b2ffd56cbe9c3df1e4a4098045256484ebe94ba5a7e4338d35ec046baa5} - {file: vae/qwen_image_2.1_vae_bf16.safetensors, role: vae, repo: Comfy-Org/Qwen-Image-2.1, revision: cb504a4090723e43f17ad01cec0359490e2de613, size: 675509688, sha256: bb21f7473051e1ac368515dd3f2e15cd44d7a11748ee8823e1ddca3e4876b7c9} - {file: Qwen3VL-8B-Instruct-Q4_K_M.gguf, role: text-encoder, repo: Qwen/Qwen3-VL-8B-Instruct-GGUF, revision: f982a07559d4a2f6c8744d840bf6fccab30eea96, size: 5027784800, sha256: 67d1659bfe71b89d50b45a4ad1a9e5b997e5bb16ce5da66a6a6167abd569e9e2} - {file: mmproj-Qwen3VL-8B-Instruct-F16.gguf, role: mmproj, repo: Qwen/Qwen3-VL-8B-Instruct-GGUF, revision: f982a07559d4a2f6c8744d840bf6fccab30eea96, size: 1159029824, sha256: ca524100ebf825c9a870db1c580d03879e0da0ab2541697e2458e64891cf9d38} qwen-image-2.1-turbo:7b: description: "Qwen Image 2.1 Turbo 7B: eight-step generation and multi-reference image editing" readme: readmes/qwen-image-2.1-turbo/7b.md type: image default: q4 capabilities: input: [text, image, multi-image] output: [image, image-edit] max_images: 8 info: family: qwen-image-2.1 parameters: 7B architecture: qwen-image-2.1 base_model: Qwen/Qwen-Image-2.1 source: Qwen/Qwen-Image-2.1-Turbo license: qwen-research homepage: "https://huggingface.co/Qwen/Qwen-Image-2.1-Turbo" settings: cfg_scale: 1.0 sampling_method: euler flash_attention: true offload_to_cpu: true steps: 8 sigmas: "1.0,0.978453,0.95418,0.926626,0.89508,0.845148,0.704534,0.414568,0.0" q4: adapter: sd-server repo: unsloth/Qwen-Image-2.1-Turbo-GGUF files: - {file: qwen-image-2.1-turbo-Q4_K_M.gguf, revision: 5d1f9f759ea877ae556cda88dbec958b62ad3809, size: 4199565120, sha256: 8940e8b522af886e0c0f5afddf720e82dd223cb9a43580b5f61c4d740452ac20} - {file: vae/qwen_image_2.1_vae_bf16.safetensors, role: vae, repo: Comfy-Org/Qwen-Image-2.1, revision: cb504a4090723e43f17ad01cec0359490e2de613, size: 675509688, sha256: bb21f7473051e1ac368515dd3f2e15cd44d7a11748ee8823e1ddca3e4876b7c9} - {file: Qwen3VL-8B-Instruct-Q4_K_M.gguf, role: text-encoder, repo: Qwen/Qwen3-VL-8B-Instruct-GGUF, revision: f982a07559d4a2f6c8744d840bf6fccab30eea96, size: 5027784800, sha256: 67d1659bfe71b89d50b45a4ad1a9e5b997e5bb16ce5da66a6a6167abd569e9e2} - {file: mmproj-Qwen3VL-8B-Instruct-F16.gguf, role: mmproj, repo: Qwen/Qwen3-VL-8B-Instruct-GGUF, revision: f982a07559d4a2f6c8744d840bf6fccab30eea96, size: 1159029824, sha256: ca524100ebf825c9a870db1c580d03879e0da0ab2541697e2458e64891cf9d38} q8: adapter: sd-server repo: unsloth/Qwen-Image-2.1-Turbo-GGUF files: - {file: qwen-image-2.1-turbo-Q8_0.gguf, revision: 5d1f9f759ea877ae556cda88dbec958b62ad3809, size: 7640860480, sha256: 399d2ddc3636f36456e2f5bfa57187da4e38721bc046430317f42a9eab5b1712} - {file: vae/qwen_image_2.1_vae_bf16.safetensors, role: vae, repo: Comfy-Org/Qwen-Image-2.1, revision: cb504a4090723e43f17ad01cec0359490e2de613, size: 675509688, sha256: bb21f7473051e1ac368515dd3f2e15cd44d7a11748ee8823e1ddca3e4876b7c9} - {file: Qwen3VL-8B-Instruct-Q4_K_M.gguf, role: text-encoder, repo: Qwen/Qwen3-VL-8B-Instruct-GGUF, revision: f982a07559d4a2f6c8744d840bf6fccab30eea96, size: 5027784800, sha256: 67d1659bfe71b89d50b45a4ad1a9e5b997e5bb16ce5da66a6a6167abd569e9e2} - {file: mmproj-Qwen3VL-8B-Instruct-F16.gguf, role: mmproj, repo: Qwen/Qwen3-VL-8B-Instruct-GGUF, revision: f982a07559d4a2f6c8744d840bf6fccab30eea96, size: 1159029824, sha256: ca524100ebf825c9a870db1c580d03879e0da0ab2541697e2458e64891cf9d38} flux2-klein:9b: description: "FLUX.2-klein 9B: four-step generation and multi-reference image editing" readme: readmes/flux2-klein/9b.md type: image default: q4 capabilities: input: [text, image, multi-image] output: [image, image-edit] max_images: 8 info: family: flux2-klein parameters: 9B architecture: flux2 base_model: black-forest-labs/FLUX.2-klein-9B source: black-forest-labs/FLUX.2-klein-9B license: flux-non-commercial-license homepage: "https://huggingface.co/black-forest-labs/FLUX.2-klein-9B" settings: steps: 40 cfg_scale: 1.0 sampling_method: euler flash_attention: true q4: adapter: sd-server repo: unsloth/FLUX.2-klein-9B-GGUF files: - {file: flux-2-klein-9b-Q4_K_M.gguf, revision: fde8634245fe6b749a221c25b34672b5b8fbd079, size: 5909829920, sha256: 5489463ed96056b0bb5472abb5d1bba7055e48d574e37877acb43b407465e26f} - {file: split_files/vae/flux2-vae.safetensors, role: vae, repo: Comfy-Org/flux2-klein, revision: 5f526678002e43af5551dadb73ce2e8c91b43afe, size: 336211292, sha256: 868fe7b343cc8f3a19dbcfcafbc3d5f888802be3f89bd81b65b3621a066ce8f3} - {file: Qwen3-8B-Q4_K_M.gguf, role: text-encoder, repo: unsloth/Qwen3-8B-GGUF, revision: a6adef130ffb23ddaf1a62fec9dced968c9bc482, size: 5027784512, sha256: 120307ba529eb2439d6c430d94104dabd578497bc7bfe7e322b5d9933b449bd4} flux3-klein:5.6b: description: "FLUX.3 Image Klein 5.6B: sixteen-step generation and multi-reference image editing" readme: readmes/flux3-klein/5.6b.md type: image default: q4 capabilities: input: [text, image, multi-image] output: [image, image-edit] max_images: 8 info: family: flux3-klein parameters: 5.6B architecture: flux3 base_model: black-forest-labs/flux-3-image-klein source: black-forest-labs/flux-3-image-klein license: flux-kommunity-license homepage: "https://huggingface.co/black-forest-labs/flux-3-image-klein" settings: steps: 16 cfg_scale: 1.0 sampling_method: euler flash_attention: true offload_to_cpu: true q4: adapter: sd-server repo: Abiray/flux-3-image-klein-sd-GGUF files: - {file: flux3_image_klein_sd_Q4_K_M.gguf, revision: 7a1154a7fcfa71eb16bfd896be942d0c2127e15b, size: 3269858080, sha256: c0dc60488f1d73517e60252d0d1ce24a8a9a19275ea669c0b96408a10fc871ac} - {file: split_files/vae/flux2-vae.safetensors, role: vae, repo: Comfy-Org/flux2-klein, revision: 5f526678002e43af5551dadb73ce2e8c91b43afe, size: 336211292, sha256: 868fe7b343cc8f3a19dbcfcafbc3d5f888802be3f89bd81b65b3621a066ce8f3} - {file: Qwen3VL-4B-Instruct-Q4_K_M.gguf, role: text-encoder, repo: Qwen/Qwen3-VL-4B-Instruct-GGUF, revision: 1cd86afb9a95c410a6038ab3b40d8b578c892266, size: 2497281664, sha256: 66358cb18bb6b3b1b6675aa412c7a88ef01d228f481184d13668e5201c730a0a} # --- Text generation (chat completions; upstream llama.cpp llama-server) --- # Every model lists its multimodal projector with role: mmproj (image input). # settings carry each model card's recommended sampling; `thinking` is the model's own default. winnow:12b: description: "Winnow 12B: Gemma 4 typed decisions about text and images" readme: readmes/winnow/12b.md type: decision default: q8 capabilities: input: [text, vision] output: [choice, score, noul] info: family: winnow parameters: 12B architecture: gemma4 base_model: google/gemma-4-12B-it source: EldanRing/Winnow-12B license: apache-2.0 languages: [multilingual] context_length: 65536 max_options: 64 homepage: "https://huggingface.co/EldanRing/Winnow-12B" settings: context_size: 8192 protocol: winnow temperature: 1.0 q8: adapter: ggmlc-custom-decider repo: EldanRing/Winnow-12B files: - {file: gguf/Winnow-12B-Q8_0.gguf, revision: 859a2cb24921d8b579badce6caa469f9290f3bc1, size: 12669646592, sha256: b710efc4c0d048ee61eed92c5fef5ce323a4d17e7c51f9f0533cc72ae50818ea} - {file: gguf/mmproj-Winnow-12B.gguf, role: mmproj, revision: 859a2cb24921d8b579badce6caa469f9290f3bc1, size: 175115840, sha256: 91f086971e56d7a7d8d39e271873fccdb49541bd259d6e02c401a4f1cb7a219e} winnow:4b: description: "Winnow E4B: efficient Gemma 4 typed decisions about text and images" readme: readmes/winnow/4b.md type: decision default: q8 capabilities: input: [text, vision] output: [choice, score, noul] info: family: winnow parameters: 4.5B effective (8B with embeddings) architecture: gemma4 base_model: google/gemma-4-E4B-it source: EldanRing/Winnow-E4B license: apache-2.0 languages: [multilingual] context_length: 65536 max_options: 64 homepage: "https://huggingface.co/EldanRing/Winnow-E4B" settings: context_size: 8192 protocol: winnow temperature: 1.2574172017327816 q8: adapter: ggmlc-custom-decider repo: EldanRing/Winnow-E4B files: - {file: gguf/Winnow-E4B-Q8_0.gguf, revision: ed891438b340bb60e2962e335fe100b2796c2124, size: 8005437472, sha256: 840e3f50e5a9c218727f44e121d1b37cc9e2c3b318c8eb422ba6ef2e27b618a2} - {file: gguf/mmproj-Winnow-E4B.gguf, role: mmproj, revision: ed891438b340bb60e2962e335fe100b2796c2124, size: 990372672, sha256: ddf46c21d7078e95338cfc22306b19b276a29a5ad089023449dd54d4b6170a51} qwen3.5:9b: description: "Qwen3.5 9B: multilingual chat model with reasoning and image input" readme: readmes/qwen3.5/9b.md type: text default: q4 capabilities: input: [text, vision, multi-image] output: [text] info: family: qwen3.5 parameters: 9B architecture: qwen3.5 base_model: Qwen/Qwen3.5-9B source: Qwen/Qwen3.5-9B license: apache-2.0 languages: [multilingual] context_length: 262144 homepage: "https://huggingface.co/Qwen/Qwen3.5-9B" settings: context_size: 16384 temperature: 1.0 top_p: 0.95 top_k: 20 min_p: 0.0 presence_penalty: 1.5 thinking: "on" q4: adapter: llama-server repo: unsloth/Qwen3.5-9B-GGUF files: - {file: Qwen3.5-9B-Q4_K_M.gguf, size: 5680522464, sha256: 03b74727a860a56338e042c4420bb3f04b2fec5734175f4cb9fa853daf52b7e8} - {file: mmproj-F16.gguf, role: mmproj, size: 918166080, sha256: f70dc3509053962b0d0d3ee8a7eacebf5d60aa560cad78254ae8698516ae029f} q8: adapter: llama-server repo: unsloth/Qwen3.5-9B-GGUF files: - {file: Qwen3.5-9B-Q8_0.gguf, size: 9527502048, sha256: 809626574d0cb43d4becfa56169980da2bb448f2299270f7be443cb89d0a6ae4} - {file: mmproj-F16.gguf, role: mmproj, size: 918166080, sha256: f70dc3509053962b0d0d3ee8a7eacebf5d60aa560cad78254ae8698516ae029f} gemma4:4b: description: "Gemma 4 E4B: efficient multilingual chat model with reasoning and image input" readme: readmes/gemma4/4b.md type: text default: q4 capabilities: input: [text, vision, multi-image] output: [text] info: family: gemma4 parameters: 4.5B effective (8B with embeddings) architecture: gemma4 base_model: google/gemma-4-E4B source: google/gemma-4-E4B-it license: apache-2.0 languages: [multilingual] context_length: 131072 homepage: "https://huggingface.co/google/gemma-4-E4B-it" settings: context_size: 16384 temperature: 1.0 top_p: 0.95 top_k: 64 min_p: 0.0 thinking: "off" q4: adapter: llama-server repo: google/gemma-4-E4B-it-qat-q4_0-gguf files: - {file: gemma-4-E4B_q4_0-it.gguf, size: 5154941280, sha256: 676c35070db6dbe52f93e9c864ee0fba4eddea94b9c875d9cb10daff453fbaee} - {file: gemma-4-E4B-it-mmproj.gguf, role: mmproj, size: 991552256, sha256: 7498a37cb619e55f2fcf87eb931f56e99389ed6d432e4c5c66110694c0d65578} q8: adapter: llama-server repo: ggml-org/gemma-4-E4B-it-GGUF files: - {file: gemma-4-E4B-it-Q8_0.gguf, size: 8031242688, sha256: 34be82b17b4942d389b9b527170c4b058027abdd32531fda063d3d97dd8ce80a} - {file: mmproj-gemma-4-E4B-it-Q8_0.gguf, role: mmproj, size: 559874816, sha256: 197f49a93027f9843772bd24a6a9e0be2a32a788de5a3def330e9c585d86edd1} gemma4:12b: description: "Gemma 4 12B: multilingual chat model with reasoning and image input" readme: readmes/gemma4/12b.md type: text default: q4 capabilities: input: [text, vision, multi-image] output: [text] info: family: gemma4 parameters: 11.95B architecture: gemma4 (unified) base_model: google/gemma-4-12B source: google/gemma-4-12B-it license: apache-2.0 languages: [multilingual] context_length: 262144 homepage: "https://huggingface.co/google/gemma-4-12B-it" settings: context_size: 16384 temperature: 1.0 top_p: 0.95 top_k: 64 min_p: 0.0 thinking: "off" q4: adapter: llama-server repo: google/gemma-4-12B-it-qat-q4_0-gguf files: - {file: gemma-4-12b-it-qat-q4_0.gguf, size: 6975879296, sha256: 93567e57a8fe10b23569b9d9ec38cd005deedf71e29477c421a4b83f418a538b} - {file: mmproj-gemma-4-12b-it-qat-q4_0.gguf, role: mmproj, size: 175115616, sha256: cb018338a7538a9814d994bfe54644c71eb7ed54e31eae2f721e45fd3c260da7} q8: adapter: llama-server repo: ggml-org/gemma-4-12B-it-GGUF files: - {file: gemma-4-12B-it-Q8_0.gguf, size: 12669646976, sha256: abfc3044b93795f6a446fc1a04468ac2d5fc5e9c92e237f09a5b3384c5a433e4} - {file: mmproj-gemma-4-12B-it-Q8_0.gguf, role: mmproj, size: 158987616, sha256: 59e62255435dda870e2d1de97cc031330b31a898bac12b38a182cecff9cd3738} # --- Embeddings (llama.cpp llama-server) --- clm-v0.1:8b: description: "CLM v0.1 8B: contrastive encoder for text ranking and embeddings" readme: readmes/clm-v0.1/8b.md type: embedding default: q4 capabilities: input: [text] output: [embedding] info: family: clm parameters: 8B architecture: qwen3 base_model: Contrastive-LM/CLM-v0.1-8B source: Contrastive-LM/CLM-v0.1-8B license: apache-2.0 languages: [en] context_length: 32768 homepage: "https://huggingface.co/Contrastive-LM/CLM-v0.1-8B" settings: context_size: 8192 pooling: last q4: adapter: llama-server repo: czl/CLM-v0.1-8B-GGUF files: - {file: Qwen3-8B-Q4_K_M.gguf, size: 5489669760, sha256: 373e464e5310bda8fc54c1b67b36c2b860be5f03e5e0d3d046f5ef8da6b7e7d6} nomic-embed-text-v1.5:137m: description: "Nomic Embed v1.5: English text embeddings" readme: readmes/nomic-embed-text-v1.5/137m.md type: embedding default: q4 capabilities: input: [text] output: [embedding] info: family: nomic-embed-text parameters: 137M architecture: nomic-bert base_model: nomic-ai/nomic-embed-text-v1.5 source: nomic-ai/nomic-embed-text-v1.5 license: apache-2.0 languages: [en] context_length: 8192 homepage: "https://huggingface.co/nomic-ai/nomic-embed-text-v1.5" settings: context_size: 8192 pooling: mean q4: adapter: llama-server repo: nomic-ai/nomic-embed-text-v1.5-GGUF files: - {file: nomic-embed-text-v1.5.Q4_K_M.gguf, size: 84106624, sha256: d4e388894e09cf3816e8b0896d81d265b55e7a9fff9ab03fe8bf4ef5e11295ac} nomic-embed-text-v2-moe:475m: description: "Nomic Embed v2 MoE: multilingual mixture-of-experts embeddings" readme: readmes/nomic-embed-text-v2-moe/475m.md type: embedding default: q4 capabilities: input: [text] output: [embedding] info: family: nomic-embed-text parameters: 475M active (305M activated) architecture: nomic-bert base_model: nomic-ai/nomic-embed-text-v2-moe source: nomic-ai/nomic-embed-text-v2-moe license: apache-2.0 languages: [multilingual] context_length: 8192 homepage: "https://huggingface.co/nomic-ai/nomic-embed-text-v2-moe" settings: context_size: 8192 pooling: mean q4: adapter: llama-server repo: nomic-ai/nomic-embed-text-v2-moe-GGUF files: - {file: nomic-embed-text-v2-moe.Q4_K_M.gguf, size: 344120288, sha256: b5fb2811647b8ef461519a68a3bf67014a84a66a130c8a2af9413ff9f06d3f22} embeddinggemma-2:740m: description: "EmbeddingGemma 2 740M: multilingual text, image and audio embeddings" readme: readmes/embeddinggemma-2/740m.md type: embedding default: q8 capabilities: input: [text, vision, audio] output: [embedding] info: family: embeddinggemma parameters: 740M architecture: gemma-embedding2 base_model: google/embeddinggemma-2 source: google/embeddinggemma-2 license: apache-2.0 languages: [multilingual] context_length: 8192 homepage: "https://huggingface.co/google/embeddinggemma-2" settings: context_size: 8192 pooling: mean q8: adapter: llama-server repo: ggml-org/embeddinggemma-2-GGUF files: - {file: embeddinggemma-2-Q8_0.gguf, size: 309855456, sha256: 2188ac1deca4b77dffefd603c2776a9d76d9d74ec01841392982ebb840b09135} - {file: mmproj-embeddinggemma-2-Q8_0.gguf, role: mmproj, size: 554821024, sha256: c4a8a52691ecef40618438928bdf9e68379b854e24166f292592353db0aab64f} # --- Text to speech --- qwen3-tts:1.7b: description: "Qwen3-TTS 1.7B Base: multilingual text-to-speech with reference audio" readme: readmes/qwen3-tts/1.7b.md type: audio default: q4 capabilities: input: [text] output: [audio] info: family: qwen3-tts parameters: 1.7B architecture: qwen3tts base_model: Qwen/Qwen3-TTS-12Hz-1.7B-Base source: Qwen/Qwen3-TTS-12Hz-1.7B-Base license: apache-2.0 q4: adapter: ggmlc-audio repo: ggml-org/Qwen3-TTS-12Hz-1.7B-Base-GGUF files: - {file: Qwen3-TTS-12Hz-1.7B-Base-Q4_K_M.gguf, size: 1035965280, sha256: 8d18c94acb2addd042f97da63c98be144eafa76d0d9495177eab65130cf85129} - {file: mmproj-Qwen3-TTS-12Hz-1.7B-Base-Q8_0.gguf, role: mmproj, size: 446422912, sha256: 6fd65188839bcd6ecc91b277ad471e22a0edfada4699a0fe82f1165c18cfcce2} q8: adapter: ggmlc-audio repo: ggml-org/Qwen3-TTS-12Hz-1.7B-Base-GGUF files: - {file: Qwen3-TTS-12Hz-1.7B-Base-Q8_0.gguf, size: 1847874400, sha256: ac7931aeb2e7aad1a6ed6602d353a5679c9d096b18ce8204ac730a8408d572e1} - {file: mmproj-Qwen3-TTS-12Hz-1.7B-Base-Q8_0.gguf, role: mmproj, size: 446422912, sha256: 6fd65188839bcd6ecc91b277ad471e22a0edfada4699a0fe82f1165c18cfcce2} bf16: adapter: ggmlc-audio repo: ggml-org/Qwen3-TTS-12Hz-1.7B-Base-GGUF files: - {file: Qwen3-TTS-12Hz-1.7B-Base-bf16.gguf, size: 3472593760, sha256: 0322c634ad5d3282524bc45bff030ab3f6f2a32ba14d5e7dace7eed75ecede46} - {file: mmproj-Qwen3-TTS-12Hz-1.7B-Base-bf16.gguf, role: mmproj, size: 669081472, sha256: b9503e95e44705739cf82c15ce909b040b96a31ad79d02cf4dc5e1484399265d} # --- Pocket TTS (Kyutai; one model per language, llama.cpp GGUFs by EryriLabs) --- # The model needs a reference voice; the `voice` file is used when a request has none. pocket-tts-en:100m: description: "Pocket TTS English: small CPU-friendly text-to-speech with voice cloning" readme: readmes/pocket-tts/en.md type: audio default: bf16 capabilities: input: [text] output: [audio] info: {family: pocket-tts, parameters: 80M, architecture: pockettts, base_model: kyutai/pocket-tts, source: kyutai/pocket-tts, license: cc-by-4.0, languages: [en]} bf16: adapter: ggmlc-audio repo: EryriLabs/pocket-tts-GGUF files: - {file: pocket-tts-en.gguf, size: 159390816, sha256: 9cca37ab8f3da366a1864919214fbd545f45609d8f1167d1fb302cebea515d63} - {file: mmproj-pocket-tts-en.gguf, role: mmproj, size: 59858080, sha256: ccb1b6a55a7df0d4f96adff04ae3b9cd99c070edc98f6c84af5adb6d46a64cfc} - {file: alba-mackenna/casual.wav, role: voice, repo: kyutai/tts-voices, size: 958542, sha256: 46264e83cb99115c3d210260e029117566d9c64f20266d10daa78107759ede3e} pocket-tts-de:100m: description: "Pocket TTS German: small CPU-friendly text-to-speech with voice cloning" readme: readmes/pocket-tts/de.md type: audio default: bf16 capabilities: input: [text] output: [audio] info: {family: pocket-tts, parameters: 80M, architecture: pockettts, base_model: kyutai/pocket-tts, source: kyutai/pocket-tts, license: cc-by-4.0, languages: [de]} bf16: adapter: ggmlc-audio repo: EryriLabs/pocket-tts-GGUF files: - {file: german/pocket-tts-german.gguf, size: 159391264, sha256: 7b797a22f993183fc5f496d4bf8207c03c6e527e8909b998fb35580ee49458fe} - {file: german/mmproj-pocket-tts-german.gguf, role: mmproj, size: 59858080, sha256: 6ef66b0cf7a68f5973a2a4a8ded1a9c2532ecddd8b445a00e5fe3469cca11ca3} - {file: unmute-prod-website/default_voice.wav, role: voice, repo: kyutai/tts-voices, size: 480044, sha256: 27157be84054d6ce8af7a1ebca827c73d1d8fc38842f979d3fae207c92b1a9ae} pocket-tts-es:100m: description: "Pocket TTS Spanish: small CPU-friendly text-to-speech with voice cloning" readme: readmes/pocket-tts/es.md type: audio default: bf16 capabilities: input: [text] output: [audio] info: {family: pocket-tts, parameters: 80M, architecture: pockettts, base_model: kyutai/pocket-tts, source: kyutai/pocket-tts, license: cc-by-4.0, languages: [es]} bf16: adapter: ggmlc-audio repo: EryriLabs/pocket-tts-GGUF files: - {file: spanish/pocket-tts-spanish.gguf, size: 159392320, sha256: aa823005551f4600e3ab7a871aa023cfbb268aaef2bfb812654ad349da351d51} - {file: spanish/mmproj-pocket-tts-spanish.gguf, role: mmproj, size: 59858080, sha256: 3fe3df5e6a569eb0a85cb6adaf6526652f8308ad9c011044777f547891ddb8c5} - {file: unmute-prod-website/default_voice.wav, role: voice, repo: kyutai/tts-voices, size: 480044, sha256: 27157be84054d6ce8af7a1ebca827c73d1d8fc38842f979d3fae207c92b1a9ae} pocket-tts-fr:100m: description: "Pocket TTS French (24 layers): text-to-speech with voice cloning" readme: readmes/pocket-tts/fr.md type: audio default: bf16 capabilities: input: [text] output: [audio] info: {family: pocket-tts, parameters: 300M, architecture: pockettts, base_model: kyutai/pocket-tts, source: kyutai/pocket-tts, license: cc-by-4.0, languages: [fr]} bf16: adapter: ggmlc-audio repo: EryriLabs/pocket-tts-GGUF files: - {file: french/pocket-tts-french.gguf, size: 612681760, sha256: 73db75f9ead052c3b488ff1e3aea4a1558ce9d22ebcb7bd6489687fb89c62a1c} - {file: french/mmproj-pocket-tts-french.gguf, role: mmproj, size: 59858080, sha256: ef157baf2b7c422caa2f66bee1f0b14829348ede347e1c756cb306925c23b8b7} - {file: unmute-prod-website/developpeuse-3.wav, role: voice, repo: kyutai/tts-voices, size: 1282324, sha256: d14dc373ead723bc818d8970d271b876aa02833673cabd97aea859444a8cdc3a} pocket-tts-it:100m: description: "Pocket TTS Italian: small CPU-friendly text-to-speech with voice cloning" readme: readmes/pocket-tts/it.md type: audio default: bf16 capabilities: input: [text] output: [audio] info: {family: pocket-tts, parameters: 80M, architecture: pockettts, base_model: kyutai/pocket-tts, source: kyutai/pocket-tts, license: cc-by-4.0, languages: [it]} bf16: adapter: ggmlc-audio repo: EryriLabs/pocket-tts-GGUF files: - {file: italian/pocket-tts-italian.gguf, size: 159391488, sha256: 3e1b54accfefc85952c782aa3aedd5396cf5fcfb979ac1ab3f66e6c39e11b722} - {file: italian/mmproj-pocket-tts-italian.gguf, role: mmproj, size: 59858080, sha256: 46f0a614c23c2fac6bd3dcf93202b227333b2fd67096abbfb4fb3d4ec78f4434} - {file: unmute-prod-website/default_voice.wav, role: voice, repo: kyutai/tts-voices, size: 480044, sha256: 27157be84054d6ce8af7a1ebca827c73d1d8fc38842f979d3fae207c92b1a9ae} pocket-tts-pt:100m: description: "Pocket TTS Portuguese: small CPU-friendly text-to-speech with voice cloning" readme: readmes/pocket-tts/pt.md type: audio default: bf16 capabilities: input: [text] output: [audio] info: {family: pocket-tts, parameters: 80M, architecture: pockettts, base_model: kyutai/pocket-tts, source: kyutai/pocket-tts, license: cc-by-4.0, languages: [pt]} bf16: adapter: ggmlc-audio repo: EryriLabs/pocket-tts-GGUF files: - {file: portuguese/pocket-tts-portuguese.gguf, size: 159392416, sha256: 9095778496f6fc09b1c9b537b20b3ecbd7ec2820db658410e5a7901170131207} - {file: portuguese/mmproj-pocket-tts-portuguese.gguf, role: mmproj, size: 59858080, sha256: 0f17b5c44e5800e5c61bf6188eed66d1f668c913fb6ff58a160d7e4fc57e0db2} - {file: unmute-prod-website/default_voice.wav, role: voice, repo: kyutai/tts-voices, size: 480044, sha256: 27157be84054d6ce8af7a1ebca827c73d1d8fc38842f979d3fae207c92b1a9ae} clef-flash:9b: description: "Clef Flash 9B: multimodal typed decisions with a joint schema head" readme: readmes/clef-flash/9b.md type: decision default: q4 capabilities: input: [text, vision, multi-image] output: [choice, score, noul] max_images: 4 info: family: clef parameters: 9B architecture: qwen3.5 + joint schema head base_model: Qwen/Qwen3.5-9B source: Cloudflare/clef-flash license: apache-2.0 languages: [en] context_length: 16384 max_options: 32 homepage: "https://huggingface.co/Cloudflare/clef-flash" settings: context_size: 16384 q4: adapter: clef repo: bartowski/Cloudflare_clef-flash-GGUF files: - {file: Cloudflare_clef-flash-Q4_K_M.gguf, size: 6040541344, sha256: cf747351878a0e569fd9d6681395bd723e111f63a4f6fe6e26e7563dfdfa9b0a} - {file: mmproj-Cloudflare_clef-flash-f16.gguf, role: mmproj, size: 918165984, sha256: c8a6e9e00e1baf6433c4874dbc73bed2a56e887bcbb582fffcb00b533bc638dd} - {file: joint_head.safetensors, role: head, repo: Cloudflare/clef-flash, size: 243538016, sha256: 19cdcec8c81dc9212be320fff47462ab342fbc1278be4368fb3da71241cf5ba0} # --- Kev (Qwen-based System One decision models, compiled with ggmlc) --- kev:0.5b: description: "Kev 0.5b: Qwen2.5-0.5B-based System One decision model (ggmlc)" readme: readmes/kev/0.5b.md type: decision default: q4 capabilities: input: [text] output: [choice, score, noul] info: family: kev parameters: 0.5B architecture: qwen2.5 base_model: Qwen/Qwen2.5-0.5B source: jaredpalmer/kev-0.5b license: apache-2.0 languages: [en] context_length: 2048 max_options: 16 settings: threads: 4 q4: adapter: ggmlc-laya repo: mys/kev-0.5b-GGUF files: - {file: kev_0.5b_ud_q4_k_m.gguf, size: 570955168, sha256: 154db89455fb1871b93e4343a2954331c2469d718426c893b4d1bd6f0e1b40bf} q8: adapter: ggmlc-laya repo: mys/kev-0.5b-GGUF files: - {file: kev_0.5b_q8_0.gguf, size: 548153728, sha256: 5c747e0da78745e49609d6989a0964ce1688fa4741a49ae13f17256acda13280} fp16: adapter: ggmlc-laya repo: mys/kev-0.5b-GGUF files: - {file: kev_0.5b_f16.gguf, size: 1011668480, sha256: 06d9735d7adfa71b0bad812a712f18d56fc6e8c0b1761d9898561d5110c2470d} kev:0.8b: description: "Kev 0.8b: Qwen3.5-0.8B-based System One decision model (ggmlc)" readme: readmes/kev/0.8b.md type: decision default: q4 capabilities: input: [text] output: [choice, score, noul] info: family: kev parameters: 0.8B architecture: qwen3.5 base_model: Qwen/Qwen3.5-0.8B-Base source: jaredpalmer/kev-0.8b license: apache-2.0 languages: [en] context_length: 2048 max_options: 16 settings: threads: 4 q4: adapter: ggmlc-laya repo: mys/kev-0.8b-GGUF files: - {file: kev_0.8b_ud_q4_k_m.gguf, size: 958685280, sha256: 410d9c071a72435d5579b88af98b53fe520260ae61e4154c234e6a28a4ae9623} q8: adapter: ggmlc-laya repo: mys/kev-0.8b-GGUF files: - {file: kev_0.8b_q8_0.gguf, size: 827809856, sha256: 12f77506dbfaa3c1d12f7e3f8da04feceda276f2d5aa322cd0858a8ce11e6c6c} fp16: adapter: ggmlc-laya repo: mys/kev-0.8b-GGUF files: - {file: kev_0.8b_f16.gguf, size: 1533202496, sha256: 6b15f6bd031e1f0f3b751692aaff75b99d1dbb5362f4cda207b3d4d26e5e11eb} kev:4b: description: "Kev 4b: Qwen3.5-4B-based System One decision model (ggmlc)" readme: readmes/kev/4b.md type: decision default: q4 capabilities: input: [text] output: [choice, score, noul] info: family: kev parameters: 4B architecture: qwen3.5 base_model: Qwen/Qwen3.5-4B-Base source: jaredpalmer/kev-4b license: apache-2.0 languages: [en] context_length: 2048 max_options: 16 settings: threads: 4 q4: adapter: ggmlc-laya repo: mys/kev-4b-GGUF files: - {file: kev_4b_ud_q4_k_m.gguf, size: 4211035872, sha256: f2b31aa5a3c0f75f22f034f0452a8d7553d583c77b2783926f3ba1682143b92b} q8: adapter: ggmlc-laya repo: mys/kev-4b-GGUF files: - {file: kev_4b_q8_0.gguf, size: 4498493152, sha256: 51c9353a5ff0b8e2f7cd7262359c0b26ea54b327214fe26a5da52b505b2978e8} fp16: adapter: ggmlc-laya repo: mys/kev-4b-GGUF files: - {file: kev_4b_f16.gguf, size: 8441712352, sha256: 72fa1c764c99b9683d82333943224fceba094a3bce2426a651f2e23c34a9127e} # --- Laya (ModernBERT / mmBERT encoders, compiled with ggmlc) --- laya-english:421m: description: "Laya English: ModernBERT encoder, English (ggmlc)" readme: readmes/laya/english.md type: decision default: q4 capabilities: input: [text] output: [choice, score, noul] info: family: laya architecture: modernbert source: convaiinnovations/laya license: apache-2.0 languages: [en] context_length: 512 max_options: 16 settings: threads: 4 q4: adapter: ggmlc-laya repo: mys/laya-GGUF files: - {file: laya_english_ud_q4_k_m.gguf, size: 419907712, sha256: e504de8fcac28e5dae1cf531fe1e1c0912703de5b82719ad84bd616e1257d835} q8: adapter: ggmlc-laya repo: mys/laya-GGUF files: - {file: laya_english_q8_0.gguf, size: 451505440, sha256: 6243305fb16cf22e53b932c220bd029be2adbfbec6ce120356517b8ae2f38382} fp16: adapter: ggmlc-laya repo: mys/laya-GGUF files: - {file: laya_english_f16.gguf, size: 846137888, sha256: ebeb7c65a3284bc07729179a4c5f19dfa2c3181b7519fd2beb02f4b0d37e8b5c} laya-multilingual:322m: description: "Laya Multilingual: mmBERT encoder, multilingual (ggmlc)" readme: readmes/laya/multilingual.md type: decision default: q8 capabilities: input: [text] output: [choice, score, noul] info: family: laya architecture: mmbert source: convaiinnovations/laya-multilingual license: apache-2.0 languages: [multilingual] context_length: 1024 max_options: 16 settings: threads: 4 q4: adapter: ggmlc-laya repo: mys/laya-multilingual-GGUF files: - {file: laya_multilingual_ud_q4_k_m.gguf, size: 523916512, sha256: 7c0e47f3dfaf1242929d69e91afb8868fcc5ee78092e7784e1aa2ee85adc9840} q8: adapter: ggmlc-laya repo: mys/laya-multilingual-GGUF files: - {file: laya_multilingual_q8_0.gguf, size: 361712736, sha256: 757c1a4b1f0f41824113dde76d6cd06b0881d37b796103a338de77f9f9b935c3} laya-typed-decisions:421m: description: "Laya Typed Decisions: encoder tuned for typed decisions (ggmlc)" readme: readmes/laya/typed-decisions.md type: decision default: q4 capabilities: input: [text] output: [choice, score, noul] info: family: laya source: convaiinnovations/laya-typed-decisions license: apache-2.0 languages: [en] context_length: 1024 max_options: 16 settings: threads: 4 q4: adapter: ggmlc-laya repo: mys/laya-typed-decisions-GGUF files: - {file: laya_typed_decisions_ud_q4_k_m.gguf, size: 423581888, sha256: 42919cae9c11f744a16b8f3136a3356c586a558108080050e832ba14bbb0010c} q8: adapter: ggmlc-laya repo: mys/laya-typed-decisions-GGUF files: - {file: laya_typed_decisions_q8_0.gguf, size: 455179648, sha256: 6eb6ef58bc4f99bd9db40381604108af782ffc314aa00da6641f429991ca1f18} # --- Decider (Mapika/decider, llama.cpp GGUFs) --- # Served by our own engine (engines/ggmlc-custom-decider, llama.cpp + mtmd). # Vision models list the multimodal projector with role: mmproj. decider:0.8b: description: "Decider 0.8B: small Qwen3.5-based one-pass decision model" readme: readmes/decider/0.8b.md type: decision default: q4 capabilities: input: [text] output: [choice, score, noul] info: family: decider parameters: 0.8B architecture: qwen3.5 base_model: Qwen/Qwen3.5-0.8B-Base source: Mapika/decider-0.8b license: apache-2.0 languages: [en] context_length: 262144 max_options: 10 homepage: "https://github.com/Mapika/decider" settings: context_size: 8192 temperature: 1.0 q4: adapter: ggmlc-custom-decider repo: mradermacher/decider-0.8b-GGUF files: - {file: decider-0.8b.Q4_K_M.gguf, size: 529297856, sha256: ce48f94d8f176e6433dc09ee34fd4a39ac16b29043beaeb7b81d98373b27cd67} q8: adapter: ggmlc-custom-decider repo: mradermacher/decider-0.8b-GGUF files: - {file: decider-0.8b.Q8_0.gguf, size: 811844032, sha256: 2665d08c1052b4e01dabcb08771d25579f6776f7066355e4a0ccc5f74e32d4a2} decider:4b: description: "Decider 4B: Qwen3.5-based one-pass decision model (v2.1)" readme: readmes/decider/4b.md type: decision default: q4 capabilities: input: [text] output: [choice, score, noul] info: family: decider parameters: 4B architecture: qwen3.5 base_model: Qwen/Qwen3.5-4B-Base source: Mapika/decider-4b license: apache-2.0 languages: [en] context_length: 262144 max_options: 10 homepage: "https://github.com/Mapika/decider" settings: context_size: 8192 temperature: 1.0 q4: adapter: ggmlc-custom-decider repo: Mapika/decider-4b-GGUF files: - {file: decider-4b-v2.1-Q4_K_M.gguf, size: 2708804640, sha256: c7083fcfc93f650cd66caeade4a3840ef9eeeeb1e80920907c554f70619d7c56} q8: adapter: ggmlc-custom-decider repo: Mapika/decider-4b-GGUF files: - {file: decider-4b-v2.1-Q8_0.gguf, size: 4482403360, sha256: 74614659ab849bc0ec4f0431e1f84ae8bc11fe30b42c3a235a8b0049d36cbc3d} "decider-vision:2b": description: "Decider 2B Vision: one-pass decisions about text and one image" readme: readmes/decider-vision/2b.md type: decision default: q4 capabilities: input: [text, vision] output: [choice, score, noul] info: family: decider parameters: 2B architecture: "qwen3.5 + qwen3-vl projector" base_model: Qwen/Qwen3.5-2B-Base source: Mapika/decider-2b-vision license: apache-2.0 languages: [en] context_length: 262144 max_options: 10 homepage: "https://github.com/Mapika/decider" settings: context_size: 8192 temperature: 1.0 q4: adapter: ggmlc-custom-decider repo: mradermacher/decider-2b-vision-GGUF files: - {file: decider-2b-vision.Q4_K_M.gguf, size: 1274397056, sha256: b3e38aa90856d6b93df0d2b186e941e57fa6fd4b022a3a89fe538ff01ebd9985} - {file: decider-2b-vision.mmproj-Q8_0.gguf, size: 364664448, sha256: 79e636085e90f1b4bdc5df22003cde0dd0d88086a04b7c101e6bd7d1cc9f2c91, role: mmproj} q8: adapter: ggmlc-custom-decider repo: mradermacher/decider-2b-vision-GGUF files: - {file: decider-2b-vision.Q8_0.gguf, size: 2012012928, sha256: 11e07e4c6a56856e05ec5b4c98308a8e869607bfaad9e70b94c9a6439c595b9c} - {file: decider-2b-vision.mmproj-Q8_0.gguf, size: 364664448, sha256: 79e636085e90f1b4bdc5df22003cde0dd0d88086a04b7c101e6bd7d1cc9f2c91, role: mmproj}