From 287ef12bf04aad6884e4652ca34b1a81ea600568 Mon Sep 17 00:00:00 2001 From: localai-org-maint-bot Date: Sun, 30 Aug 2026 09:01:05 +0200 Subject: [PATCH 1/3] feat(gallery): add Granite 4.2 variants (#11779) Add the 3B, 8B, and 30B safetensors checkpoints as one vLLM variant family so LocalAI can select the largest build that fits. Configure the parsers and sampling defaults recommended for Granite reasoning and tool calls. Assisted-by: Codex:gpt-5 Co-authored-by: localai-org-maint-bot <306269227+localai-org-maint-bot@users.noreply.github.com> --- gallery/index.yaml | 106 +++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 106 insertions(+) diff --git a/gallery/index.yaml b/gallery/index.yaml index cab4117e092b..202ce3d14e56 100644 --- a/gallery/index.yaml +++ b/gallery/index.yaml @@ -16341,6 +16341,112 @@ - filename: antares-1b-q8_0.gguf sha256: f0b147c42d9056d59eb3f0ab51d4d57a2259bf1b4e37ae986ef96a6c2e0ac6b7 uri: huggingface://mitkox/antares-1b-Q8_0-GGUF/antares-1b-q8_0.gguf +- &granite-4-2-3b-vllm + name: "granite-4.2-3b:vllm" + variants: + - model: granite-4.2-8b:vllm + - model: granite-4.2-30b:vllm + url: "github:mudler/LocalAI/gallery/vllm.yaml@master" + urls: + - https://huggingface.co/ibm-granite/granite-4.2-3b + description: | + Granite 4.2 3B is IBM's compact dense reasoning model for code generation, + tool calling, agentic workflows, multilingual chat, and long-context tasks. + This entry serves the bfloat16 safetensors with vLLM and supports a + 128K-token context. It is the smallest fallback in a family that also offers + the higher-capacity 8B and 30B checkpoints as variants. + license: apache-2.0 + icon: https://cdn-avatars.huggingface.co/v1/production/uploads/639bcaa2445b133a4e942436/CEW-OjXkRkDNmTxSu8Egh.png + tags: + - llm + - vllm + - gpu + - granite + - granite-4.2 + - 3b + - reasoning + - code + - function-calling + - agentic + - multilingual + - long-context + - instruction-tuned + last_checked: "2026-08-29" + overrides: + context_size: 131072 + options: + - reasoning_parser:granite + - tool_parser:qwen3_coder + parameters: + model: ibm-granite/granite-4.2-3b + temperature: 1 + top_p: 0.95 +- !!merge <<: *granite-4-2-3b-vllm + name: "granite-4.2-8b:vllm" + variants: [] + urls: + - https://huggingface.co/ibm-granite/granite-4.2-8b + description: | + Granite 4.2 8B is IBM's mid-sized dense reasoning model for code generation, + tool calling, agentic workflows, multilingual chat, and long-context tasks. + This entry serves the higher-capacity bfloat16 safetensors with vLLM and + supports a 128K-token context. + tags: + - llm + - vllm + - gpu + - granite + - granite-4.2 + - 8b + - reasoning + - code + - function-calling + - agentic + - multilingual + - long-context + - instruction-tuned + overrides: + context_size: 131072 + options: + - reasoning_parser:granite + - tool_parser:qwen3_coder + parameters: + model: ibm-granite/granite-4.2-8b + temperature: 1 + top_p: 0.95 +- !!merge <<: *granite-4-2-3b-vllm + name: "granite-4.2-30b:vllm" + variants: [] + urls: + - https://huggingface.co/ibm-granite/granite-4.2-30b + description: | + Granite 4.2 30B is IBM's largest dense Granite 4.2 reasoning model for code + generation, tool calling, agentic workflows, multilingual chat, and + long-context tasks. This entry serves the bfloat16 safetensors with vLLM and + supports a 128K-token context. + tags: + - llm + - vllm + - gpu + - granite + - granite-4.2 + - 30b + - reasoning + - code + - function-calling + - agentic + - multilingual + - long-context + - instruction-tuned + overrides: + context_size: 131072 + options: + - reasoning_parser:granite + - tool_parser:qwen3_coder + parameters: + model: ibm-granite/granite-4.2-30b + temperature: 1 + top_p: 0.95 - name: ibm-granite_granite-4.0-h-small url: github:mudler/LocalAI/gallery/granite4.yaml@master urls: From 2a13b6e1f27b9169af6ba696926a96cf4c6c8d45 Mon Sep 17 00:00:00 2001 From: localai-org-maint-bot Date: Sun, 30 Aug 2026 09:33:24 +0200 Subject: [PATCH 2/3] feat(gallery): add Qwen3.8 Cold Fusion (#11754) Add Q4_K_M and Q8_0 MTP variants with the shared vision projector. The publisher recommends these builds for faster Qwen3.8 generation. Assisted-by: Codex:gpt-5 Co-authored-by: localai-org-maint-bot <306269227+localai-org-maint-bot@users.noreply.github.com> --- gallery/index.yaml | 97 ++++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 97 insertions(+) diff --git a/gallery/index.yaml b/gallery/index.yaml index 202ce3d14e56..d9a689db3ffd 100644 --- a/gallery/index.yaml +++ b/gallery/index.yaml @@ -2600,6 +2600,103 @@ - filename: llama-cpp/mmproj/qwen3.5-9b-defiant-fable/mmproj-BF16.gguf uri: huggingface://DavidAU/Qwen3.5-9B-The-Defiant-Fable-Uncensored-Heretic-NEO-IMATRIX-MAX-MTP-GGUF/mmproj-BF16.gguf sha256: 853698ce7aa6c7ba732478bad280240969ddf7b0fcbf93900046f63903a83383 +- &qwen3-8-27b-cold-fusion + name: "qwen3.8-27b-cold-fusion-q4-mtp" + variants: + - model: qwen3.8-27b-cold-fusion-q8-mtp + url: "github:mudler/LocalAI/gallery/virtual.yaml@master" + urls: + - https://huggingface.co/DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1 + - https://huggingface.co/DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NM-DAU-NEO-MAX-MTP-GGUF + description: | + Qwen3.8 27B Cold Fusion is an Apache-2.0 multimodal fine-tune for + reasoning, coding, creative writing, and roleplay. This entry uses the + publisher's NEO-imatrix Q4_K_M GGUF with multi-token prediction enabled. + It supports vision through the shared BF16 projector and a native 256K + context window. + license: apache-2.0 + tags: + - llm + - gguf + - cpu + - gpu + - qwen3.8 + - reasoning + - coding + - creative-writing + - vision + - multimodal + - mtp + last_checked: "2026-08-27" + overrides: + backend: llama-cpp + function: + automatic_tool_parsing_fallback: true + grammar: + disable: true + known_usecases: + - chat + - vision + mmproj: llama-cpp/mmproj/qwen3.8-27b-cold-fusion/mmproj-BF16.gguf + options: + - use_jinja:true + - spec_type:draft-mtp + - spec_n_max:6 + - spec_p_min:0.75 + parameters: + min_p: 0.0 + model: llama-cpp/models/qwen3.8-27b-cold-fusion/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NM-DAU-NEO-MAX-NEO-MTP-Q4_K_M.gguf + repeat_penalty: 1.0 + temperature: 1.0 + top_k: 20 + top_p: 0.95 + template: + use_tokenizer_template: true + files: + - filename: llama-cpp/models/qwen3.8-27b-cold-fusion/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NM-DAU-NEO-MAX-NEO-MTP-Q4_K_M.gguf + uri: huggingface://DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NM-DAU-NEO-MAX-MTP-GGUF/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NM-DAU-NEO-MAX-NEO-MTP-Q4_K_M.gguf + sha256: db466a9432a52b87a7b7560f432f0e1caafeb111dbe3d168acf74dfe143a637c + - filename: llama-cpp/mmproj/qwen3.8-27b-cold-fusion/mmproj-BF16.gguf + uri: huggingface://DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NM-DAU-NEO-MAX-MTP-GGUF/mmproj-BF16.gguf + sha256: 734916669e61e798d2af7cdbff50ef7b1520b66993b69089bb883048cc048929 +- !!merge <<: *qwen3-8-27b-cold-fusion + name: "qwen3.8-27b-cold-fusion-q8-mtp" + variants: [] + description: | + Qwen3.8 27B Cold Fusion in the higher-quality NEO-imatrix Q8_0 GGUF + format. Multi-token prediction is enabled, and the shared BF16 projector + provides vision support. + overrides: + backend: llama-cpp + function: + automatic_tool_parsing_fallback: true + grammar: + disable: true + known_usecases: + - chat + - vision + mmproj: llama-cpp/mmproj/qwen3.8-27b-cold-fusion/mmproj-BF16.gguf + options: + - use_jinja:true + - spec_type:draft-mtp + - spec_n_max:6 + - spec_p_min:0.75 + parameters: + min_p: 0.0 + model: llama-cpp/models/qwen3.8-27b-cold-fusion/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NM-DAU-NEO-MAX-NEO-MTP-Q8_0.gguf + repeat_penalty: 1.0 + temperature: 1.0 + top_k: 20 + top_p: 0.95 + template: + use_tokenizer_template: true + files: + - filename: llama-cpp/models/qwen3.8-27b-cold-fusion/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NM-DAU-NEO-MAX-NEO-MTP-Q8_0.gguf + uri: huggingface://DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NM-DAU-NEO-MAX-MTP-GGUF/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NM-DAU-NEO-MAX-NEO-MTP-Q8_0.gguf + sha256: e78d3087059cc7280fbae60b551f6be0a2fdcd0afea8d5d28283b06878172a81 + - filename: llama-cpp/mmproj/qwen3.8-27b-cold-fusion/mmproj-BF16.gguf + uri: huggingface://DavidAU/Qwen3.8-27B-Cold-Fusion-GAIN-V1.1-NM-DAU-NEO-MAX-MTP-GGUF/mmproj-BF16.gguf + sha256: 734916669e61e798d2af7cdbff50ef7b1520b66993b69089bb883048cc048929 - &qwen3-8-2b-distill name: "qwen3.8-2b-distill-q4" variants: From 9c7c9974ebbebd111791e98421a194194df6fae8 Mon Sep 17 00:00:00 2001 From: localai-org-maint-bot Date: Sun, 30 Aug 2026 09:33:39 +0200 Subject: [PATCH 3/3] feat(gallery): add LFM2.5 8B DSpark variants (#11751) Add Q4 and Q8 target builds with LiquidAI DSpark draft sidecars. The variants expose mainline llama.cpp speculative decoding for the existing LFM2.5 8B family. Assisted-by: Codex:gpt-5 Co-authored-by: localai-org-maint-bot <306269227+localai-org-maint-bot@users.noreply.github.com> --- gallery/index.yaml | 105 +++++++++++++++++++++++++++++++++++++++++++++ 1 file changed, 105 insertions(+) diff --git a/gallery/index.yaml b/gallery/index.yaml index d9a689db3ffd..e46620e729c1 100644 --- a/gallery/index.yaml +++ b/gallery/index.yaml @@ -8666,6 +8666,8 @@ - model: lfm2.5-8b-a1b-ud-q5-k-m - model: lfm2.5-8b-a1b-ud-q6-k - model: lfm2.5-8b-a1b-q8-0 + - model: lfm2.5-8b-a1b-dspark + - model: lfm2.5-8b-a1b-q8-dspark url: "github:mudler/LocalAI/gallery/virtual.yaml@master" urls: - https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-GGUF @@ -8698,6 +8700,109 @@ - filename: llama-cpp/models/LFM2.5-8B-A1B-GGUF/LFM2.5-8B-A1B-Q8_0.gguf uri: https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-GGUF/resolve/main/LFM2.5-8B-A1B-Q8_0.gguf sha256: 33ab3b8ce6a964fb8ebac89360c9b3cf72c4fa418d5e4c0a94d46883124d5c02 +- name: "lfm2.5-8b-a1b-dspark" + url: "github:mudler/LocalAI/gallery/virtual.yaml@master" + urls: + - https://huggingface.co/LiquidAI/LFM2.5-8B-A1B + - https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-GGUF + - https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-DSpark-GGUF + description: | + LFM2.5-8B-A1B with LiquidAI's DSpark speculative drafter. This build pairs + the Q4_K_M target with the compact Q4_K_M draft sidecar for lower-memory + hosts. DSpark proposes blocks of tokens that the target model verifies, + which preserves the target model's output while accelerating generation. + license: "other" + tags: + - llm + - gguf + - reasoning + - cpu + - gpu + - dspark + icon: https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/qUZVGkns1bg3sZUShBbhv.png + overrides: + backend: llama-cpp + context_size: 131072 + draft_model: llama-cpp/models/LFM2.5-8B-A1B-DSpark-GGUF/LFM2.5-8B-A1B-DSpark-Q4_K_M.gguf + function: + automatic_tool_parsing_fallback: true + grammar: + disable: true + known_usecases: + - chat + options: + - use_jinja:true + - spec_type:draft-dspark + - spec_n_max:10 + - spec_n_min:0 + parameters: + min_p: 0.15 + model: llama-cpp/models/LFM2.5-8B-A1B-GGUF/LFM2.5-8B-A1B-Q4_K_M.gguf + repeat_penalty: 1.05 + temperature: 0.1 + top_k: 50 + top_p: 0.1 + template: + use_tokenizer_template: true + files: + - filename: llama-cpp/models/LFM2.5-8B-A1B-GGUF/LFM2.5-8B-A1B-Q4_K_M.gguf + uri: https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-GGUF/resolve/main/LFM2.5-8B-A1B-Q4_K_M.gguf + sha256: 4923ec14f06b968b74d663e5949867d2d9c3bf13a20b8be1a9f9af39989b2bb0 + - filename: llama-cpp/models/LFM2.5-8B-A1B-DSpark-GGUF/LFM2.5-8B-A1B-DSpark-Q4_K_M.gguf + uri: https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-DSpark-GGUF/resolve/main/LFM2.5-8B-A1B-DSpark-Q4_K_M.gguf + sha256: 017278ec4409671890f2722a690a52dba59f5f02eff604f0c1e79cf52f7c4ae8 +- name: "lfm2.5-8b-a1b-q8-dspark" + url: "github:mudler/LocalAI/gallery/virtual.yaml@master" + urls: + - https://huggingface.co/LiquidAI/LFM2.5-8B-A1B + - https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-GGUF + - https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-DSpark-GGUF + description: | + LFM2.5-8B-A1B with LiquidAI's DSpark speculative drafter. This build pairs + the higher-quality Q8_0 target with the recommended F16 draft sidecar for + the best acceptance length. DSpark proposes blocks of tokens that the + target model verifies, which preserves the target model's output while + accelerating generation. + license: "other" + tags: + - llm + - gguf + - reasoning + - cpu + - gpu + - dspark + icon: https://cdn-uploads.huggingface.co/production/uploads/61b8e2ba285851687028d395/qUZVGkns1bg3sZUShBbhv.png + overrides: + backend: llama-cpp + context_size: 131072 + draft_model: llama-cpp/models/LFM2.5-8B-A1B-DSpark-GGUF/LFM2.5-8B-A1B-DSpark-F16.gguf + function: + automatic_tool_parsing_fallback: true + grammar: + disable: true + known_usecases: + - chat + options: + - use_jinja:true + - spec_type:draft-dspark + - spec_n_max:10 + - spec_n_min:0 + parameters: + min_p: 0.15 + model: llama-cpp/models/LFM2.5-8B-A1B-GGUF/LFM2.5-8B-A1B-Q8_0.gguf + repeat_penalty: 1.05 + temperature: 0.1 + top_k: 50 + top_p: 0.1 + template: + use_tokenizer_template: true + files: + - filename: llama-cpp/models/LFM2.5-8B-A1B-GGUF/LFM2.5-8B-A1B-Q8_0.gguf + uri: https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-GGUF/resolve/main/LFM2.5-8B-A1B-Q8_0.gguf + sha256: 33ab3b8ce6a964fb8ebac89360c9b3cf72c4fa418d5e4c0a94d46883124d5c02 + - filename: llama-cpp/models/LFM2.5-8B-A1B-DSpark-GGUF/LFM2.5-8B-A1B-DSpark-F16.gguf + uri: https://huggingface.co/LiquidAI/LFM2.5-8B-A1B-DSpark-GGUF/resolve/main/LFM2.5-8B-A1B-DSpark-F16.gguf + sha256: a73be221783d3e832016c73945640fd81376f5de7a250025fd67a536435ce4b6 - name: "qwopus3.5-9b-coder-mtp" url: "github:mudler/LocalAI/gallery/virtual.yaml@master" urls: