Skip to content

Recipe catalog

Soup ships 181 ready-made configs at v0.78.0, and this page lists every one of them. The Recipes page explains how to use them; this one is the catalog, grouped by training task. v0.78.0 added eight recipes, see What changed in v0.78.0; v0.77.2 added two recipes, see What changed in v0.77.2; v0.77.0 added two recipes, removed one and marked two unavailable, see What changed in v0.77.0; v0.76.0 added nine, removed six and changed 38 more, see What changed in v0.76.0.

bash
soup recipes list                      # all of them
soup recipes search qwen --task grpo   # keyword, --task and --size filters
soup recipes show llama3.1-8b-sft      # print the YAML
soup recipes use llama3.1-8b-sft -o soup.yaml   # write it, then: soup train --config soup.yaml

What a recipe is, and is not

A recipe is a complete soup.yaml for one base model and one task. At v0.78.0, 179 of the 181 load through the config schema with no unknown keys and no refusal, and the other two, kimi-k2-sft and kimi-k2-thinking-grpo, are refused when their config loads, and soup recipes use exits 3 on them (see What changed in v0.77.0 and Recipes): that was checked at v0.78.0 by loading every one through the schema, and by running soup recipes use and soup doctor --config on each of the eight new ones. That says the file is valid. It does not say the run works end to end, and it does not say it was trained to a quality target: the data paths are placeholders you replace, a few recipes stop at trainer setup or carry a key the trainer ignores (the list below names them, as checked at v0.77.0, plus the v0.78.0 additions), and this page loaded every recipe through the schema without training any of them. Treat every hyperparameter as a starting point, and read the base model's license before you train on it.

The Size column is the base model's total parameter count as the catalog states it (for a mixture-of-experts model, all resident parameters, not the active ones), not a memory figure. Since v0.76.0 a test compares the label of every recipe whose base has a recorded parameter count (all but four bases the Hub has no count for) with that count and allows 0.70x to 1.15x (#1161), so the labels are approximate rather than exact. Nothing checked them before that: through v0.75.2 deepseek-v3-7b-sft listed 7B for DeepSeek-V3-0324 (it now lists 685B, and the 7b in its name still does not describe the size), and glm-4.6-sft, glm-5-sft and minimax-m2-sft listed 9B. N/A means the catalog gives no size; the test pins those 13 recipes by name. Check the model card before you size a GPU. Quantization is the training.quantization the recipe sets (4bit is 4-bit NF4 loading, the QLoRA setup; none means no quantization; 8bit is 8-bit bitsandbytes loading; not set leaves the schema default, which is 4bit). bitnet_1.58 no longer appears: it is refused at config load and its only recipe was removed. Recipes that run on Apple Silicon say mlx in the notes.

What changed in v0.78.0

  • Eight recipes added. qwen3.5-4b-grpo (#2522), qwen3.5-27b-grpo (#2605), qwen3.5-122b-a10b-grpo (#2660), qwen3.5-397b-a17b-grpo (#2718), qwen3.5-122b-a10b-dpo (#2606), minimax-m3-dpo (#2719), qwen3.5-9b-pretrain (#2720) and qwen3.5-2b-pretrain (#2721), by @kok-o, @nazek52 and @abdumutalipabay0. Upstream says the Qwen3.5 and MiniMax recipes use inherited settings and are not tuned or measured (tracked as #275). None was trained for this page; all eight write with soup recipes use and pass soup doctor --config at v0.78.0 (for these tasks that pass means "not reviewed": the command prints None of the 0 setting(s) known to be unread where it has no table). Nothing was removed or edited: the catalog was 173 and is 181.
  • What upstream checked, and what it did not. For seven of the eight (all but minimax-m3-dpo), the pinned public configuration's text architecture and LoRA attachment were checked on the meta device, and a tiny random decoder's forward and backward pass on CPU. For minimax-m3-dpo upstream's tests run one DPO step on a tiny LLaVA that has the wrapper's shape, not on the 428B weights, NF4 or a multi-GPU setup. For all eight, full-checkpoint training, NF4 execution, GPU memory use and training quality are not checked. The three large MoE recipes (qwen3.5-122b-a10b-dpo, qwen3.5-122b-a10b-grpo, qwen3.5-397b-a17b-grpo) say in the catalog that multi-GPU or multi-node memory fit is unverified, and upstream's docs add that no GPU count or memory-fit recommendation has been measured and that ten (seventeen) billion active parameters do not mean ten (seventeen) billion resident parameters.
  • minimax-m3-dpo exists now. It is MiniMax-M3 preference training through DPO's vision path (modality: vision, attention-only LoRA on the language tower, text-only prompt, chosen and rejected rows). The processor is custom code, so pass --trust-remote-code, and the MiniMax Community License needs a separate agreement for commercial use. (This site said through v0.77.2 that no minimax-m3-dpo recipe was in the catalog; that was true then, and since v0.78.0 it is not.)
  • Search by size changed. soup recipes search --size matches the size a recipe declares, as a whole token (#2930, #2932): see Recipes.

What changed in v0.77.2

  • Two recipes added. qwen3.5-0.8b-dpo (#2184) and qwen3.5-27b-dpo (#2445), both by @kok-o, are text DPO recipes for Qwen/Qwen3.5-0.8B and Qwen/Qwen3.5-27B (modality: text, LoRA, 4-bit). Upstream says they use inherited Qwen 3.5 DPO settings, not hyperparameters tuned for either size, and that they are not tuned or measured (tracked as #275). The 0.8B sets batch_size: auto; the 27B sets gradient_accumulation_steps: 8. None was trained for this page; both write with soup recipes use and pass soup doctor --config at v0.77.2.
  • Nothing else changed. No recipe was removed or edited, and kimi-k2-sft and kimi-k2-thinking-grpo were still the two that were refused at load. The catalog was 173, of which the DPO group was then 21.

What changed in v0.77.0

  • Two recipes added. qwen3.5-4b-dpo and qwen3.5-2b-dpo are text DPO recipes for Qwen/Qwen3.5-4B and Qwen/Qwen3.5-2B (modality: text, so the language decoder only). Upstream says their settings are inherited from the existing Qwen 3.5 DPO recipes, and it claims no training or performance result for either (#2129, #2160). None was trained for this page.
  • One recipe removed. pixtral-dpo: its llava rows had no chosen / rejected columns, and v0.77.0 refuses llava and sharegpt4v rows on preference tasks at config load (#1867, #2064). soup recipes show pixtral-dpo and soup recipes use pixtral-dpo now exit 1 with Recipe not found.
  • Two recipes marked unavailable. kimi-k2-sft and kimi-k2-thinking-grpo stay in the catalog, and at v0.77.0 a resident Kimi K2 load was refused at config load on Transformers 5.x (#2104, #2130). soup recipes show still prints their YAML; soup recipes use prints the reason, writes nothing and exits 3. At v0.77.0 Kimi-K2-Base and -Instruct were admitted only under training.stream_layers: true, which upstream verified on CPU with a tiny K2-shaped checkpoint only (nothing has been trained on K2); Kimi-K2-Thinking has no supported route. (Since v0.78.0 a resident route for Kimi-K2-Base and -Instruct is admitted for task: sft and task: grpo at quantization: none, with the same thin evidence; the two recipes stay unavailable because kimi-k2-sft sets quantization: 4bit and Thinking has no route.) See Recipes.
  • 22 recipes gained training.gradient_checkpointing: true. Since v0.77.0 the key reaches the TRL trainers, and with its default false the grpo, ipo, bco, reward_model and online_dpo tasks train with checkpointing off, where TRL's own default (on) used to apply. Every grpo, ipo, reward_model and online_dpo recipe in the catalog now sets it, so they keep the memory they had (#1923).

What changed in v0.76.0

  • Nine recipes added. kimi-k2.5-dpo and kimi-k2.5-grpo, deepseek-v4-pro-dpo and deepseek-v4-pro-grpo, qwen3.5-35b-a3b-grpo, qwen3.6-35b-a3b-dpo and qwen3.6-35b-a3b-grpo, qwen3.6-27b-dpo and qwen3.6-27b-grpo. Upstream's changelog says the Kimi K2.5 pair is not trained and takes its hyperparameters from the K2.6 recipes. None of the nine was trained for this page either.
  • Six recipes removed. soup recipes show and soup recipes use of any of them now exit 1 with Recipe not found. cogito-v2-sft and gemma3-9b-sft: no compatible replacement base exists for their shipped definitions (#1122). mistral-large-3-sft: Mistral-Large-3 is published in Mistral's native params.json layout with no config.json, which AutoConfig cannot read (#1398). falcon-e-bitnet-sft: quantization: bitnet_1.58 is refused at config load because BitNet training is not implemented (#825). sesame-csm-tts and spark-tts: their raw-audio configs are refused at config load (#265; since v0.77.0 Sesame CSM trains natively, #1769, and the recipe was not restored). See Removed in v0.76.0.
  • Seven base model IDs fixed. devstral-sft, glm-4.6-sft, granite-4-sft, internvl-3-5-sft, kimi-k2-sft, mistral-medium-3-5-sft and voxtral-sft named invalid Hugging Face IDs; the table below shows the IDs they use now (#1122).
  • Two recipes changed path. minimax-m3-sft and mistral-medium-3-5-sft now set modality: vision, because their bases are vision-language wrappers. oute-tts now reads data.format: chatml (pre-encoded codec tokens) instead of raw audio.
  • 23 existing recipes gained lora.dropout: 0.0 (22 MoE recipes that keep moe_lora, and minimax-m3-sft, which kept the line when it dropped moe_lora). peft refuses dropout on the fused 3-D expert parameters of a Qwen3-MoE under transformers 5.x, so with the schema default of 0.05 the LoRA attach fails (#798). If you turn moe_lora on for a MoE base of your own, set lora.dropout: 0.0 as these recipes do.

Before you run one

Loading as a config and running are different things. These are the recipes where the YAML is valid and the run still stops, needs something you have to supply first, or carries a setting that does not do what its name says. Each was checked against the v0.76.0 source, and the points below that name v0.77.0 were checked again at that tag; none was trained to confirm it.

soup recipes verify is new in v0.76.0 and checks one narrower thing for every recipe at once: whether the LoRA adapter the recipe asks for attaches to its base model. It builds the base's architecture on the meta device from its config.json, so no weights are downloaded, but it fetches each config.json from the Hub and needs the training stack installed. It does not look at your data, your memory or whether training converges, and its exit 0 can mean every config was unverified, so read the summary line. Details are on Recipes; soup recipes verify --config soup.yaml checks a config of your own before a long download.

  • Sharding is a launch flag, not a recipe key. Where a description names DeepSpeed ZeRO-2 or ZeRO-3 or FSDP2 (for example llama3.1-70b-sft, qwen2.5-72b-sft, gemma3-27b-sft, qwen3-32b-sft, qwen3.6-27b-dpo, qwen3.6-27b-grpo, deepseek-r1-32b-grpo and llama3-70b-fsdp2) or says multi-node (the DeepSeek V4 Pro, Kimi K2.5 and Kimi K2.6 recipes, among others), the YAML does not select it. Pass --deepspeed zero2, --deepspeed zero3 or --fsdp full_shard, plus --gpus N, to soup train, as the multi-GPU page shows; for more than one machine v0.76.0 adds --nodes, --node-rank, --master-addr and --master-port, see Multi-node launch. llama3-70b-fsdp2 sets use_fsdp2_compile: true, and soup train --config soup.yaml without --fsdp exits 1.
  • Some recipes need something first. llama3.1-8b-ppo loads training.reward_model: ./output_rm, a reward model you train beforehand. llama3.1-8b-longctx sets use_flash_attn: true, which exits 1 without CUDA and the flash-attn package. baichuan-sft sets training.hub: modelscope, so the base model is fetched through ModelScope into ./.soup_hub_cache. minimax-m3-sft loads a processor that is custom code, so it needs soup train --trust-remote-code.
  • MoE recipes and moe_lora. Through v0.75.2 only the SFT and pretrain trainers and the layer-streaming setup read moe_lora. The eight DPO and GRPO MoE recipes in the v0.75.2 catalog (qwen3.5-35b-a3b-dpo, deepseek-v4-flash-dpo, glm-5.1-dpo, kimi-k2.6-dpo, qwen3-30b-a3b-reasoning-grpo, deepseek-v4-flash-grpo, glm-5.1-grpo and kimi-k2.6-grpo) set it anyway. Upstream says they got attention-only LoRA, not what their YAML said, and that on transformers 5.x (which v0.75.2 requires) peft refused the LoRA attach outright for qwen3_moe, so those runs stopped there (#798, #1074). If you ran one of them, re-run. Since v0.76.0 the DPO, GRPO and the other preference and RL trainers read it, so the 15 DPO and GRPO recipes that set it now pass it on, with limits upstream states: since #1421 expert adapters attach through peft's fused-parameter path (target_parameters) on the qwen3_moe, deepseek_v3, glm4_moe, olmoe, qwen2_moe, qwen3_5_moe_text, minimax and mixtral model types, minimax-m3-sft is attention-only LoRA for a different reason (it no longer sets moe_lora, and its YAML comment says it now targets no experts), the Kimi K2.x recipes are untested rather than known-good because no tiny stand-in for that configuration builds in the installed transformers, and upstream's checks ran on tiny random stand-ins on CPU, never on a real 30B-plus checkpoint. moe_aux_loss_coeff: 0.01 in the DPO and GRPO recipes is the default value and at v0.77.0 is still read only by SFT, pretrain and tts: on a DPO or GRPO recipe it does nothing. (Since v0.77.0 a non-default value is refused at load on backend: unsloth and with stream_layers, which apply no MoE auxiliary loss either.)
  • A reward domain that does nothing. deepseek-v3-reasoning, kimi-k2-thinking-grpo and qwq-32b-grpo set verifiable_domain: math beside a reward_fn that is not verifiable, and verifiable_domain is read only by the verifiable reward (still so at v0.77.0), so it does nothing there.
  • Vision-language bases. minimax-m3-sft and mistral-medium-3-5-sft set modality: vision, so SFT loads the vision-language wrapper and LoRA stays on the language tower. The minimax-m3-sft YAML says in a comment that the vision collator supervises every token, prompt included, and that train_on_responses_only is not applied on this path. pixtral-dpo was not usable as shipped through v0.76.1, because its data.format: llava rows carry no prompt, chosen or rejected, which is what DPO reads (#1867). (Since closed in v0.77.0: the recipe was removed and vision row formats are refused on preference tasks at config load, #2064.) The GRPO trainer is still text-only at v0.77.0, so llama3.2-vision-grpo does not use images.
  • Speech and BitNet. Three TTS recipes remain. orpheus-tts-sft and llasa-tts set data.format: audio, the live-codec path (Orpheus needs pip install snac, Llasa the [audio] extra); oute-tts trains on pre-encoded chatml rows, because a raw-audio Oute config is refused at config load (still so at v0.77.0). Sesame CSM trains natively since v0.77.0, but there is no Sesame CSM recipe; see Modality II. There is no BitNet recipe: falcon-e-bitnet-sft was removed, and the export path (soup export --format tq1_0) is the BitNet path that works, see BitNet. whisper-large-v3-ft goes through the generic audio SFT path; for Whisper speech-to-text use whisper-large-v3-asr. deepseek-v3-pipeline validates its pipeline setting and then runs data-parallel.
  • The v0.78.0 recipes. minimax-m3-dpo loads a custom processor, so it needs soup train --trust-remote-code, and sets modality: vision (DPO builds the vision-language wrapper; the rows are text only). The GRPO recipes score with the built-in accuracy reward, so every prompt needs a reference answer, either an answer column or the final assistant message. Upstream's docs say GRPO promotes a configured batch size of 1 to at least 4 for four generations, so the 122B and 397B GRPO recipes are not one-completion workloads, and that the whole 122B (397B) base, the rollout state and the reference model matter, not only the active parameters. None of these was trained or fitted to a card by anyone here. These points come from upstream's docs and the YAML at the tag; the recipes were loaded and written, not run.

SFT: supervised fine-tuning

RecipeBase modelSizeQuantizationWhat it is
baichuan-sftbaichuan-inc/Baichuan2-13B-Chat13B4bitBaichuan 2 13B chat SFT
biomistral-7b-sftBioMistral/BioMistral-7B7B4bitBioMistral 7B medical/biomedical domain SFT
codellama-13b-sftcodellama/CodeLlama-13b-Instruct-hf13B4bitCode Llama 13B code-specialist SFT
codellama-70b-sftcodellama/CodeLlama-70b-Instruct-hf70B4bitCode Llama 70B code-specialist SFT (multi-GPU)
deepseek-ocr-sftdeepseek-ai/DeepSeek-OCRN/A4bitDeepSeek-OCR vision OCR SFT
deepseek-r1-distill-llama-8b-sftdeepseek-ai/DeepSeek-R1-Distill-Llama-8B8B4bitDeepSeek-R1-Distill Llama 8B reasoning SFT
deepseek-r1-distill-qwen-1.5b-sftdeepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B1.5B4bitDeepSeek-R1-Distill Qwen 1.5B reasoning SFT
deepseek-r1-distill-qwen-7b-sftdeepseek-ai/DeepSeek-R1-Distill-Qwen-7B7B4bitDeepSeek-R1-Distill Qwen 7B reasoning SFT
deepseek-v3-7b-sftdeepseek-ai/DeepSeek-V3-0324685B4bitDeepSeek V3 SFT with MoE LoRA. The base is DeepSeek-V3-0324, so the 7b in the name does not describe the size
deepseek-v3-pipelinedeepseek-ai/DeepSeek-V3671B4bitDeepSeek V3 SFT scaffold. Its parallelism: pipeline (4 stages) is checked (it needs CUDA and at least 4 GPUs) and then the run is data-parallel at v0.77.0, see Pipeline parallelism
deepseek-v4-flash-sftdeepseek-ai/DeepSeek-V4-FlashN/A4bitDeepSeek V4 Flash MoE SFT (MIT, efficiency-tier)
deepseek-v4-pro-sftdeepseek-ai/DeepSeek-V4-ProN/A4bitDeepSeek V4 Pro flagship MoE SFT (MIT, 1.6T-class). Requires multi-node DeepSpeed.
devstral-sftmistralai/Devstral-Small-250724B4bitDevstral Small code/agent SFT
gemma2-2b-sftgoogle/gemma-2-2b-it2B4bitGemma 2 2B SFT (edge-friendly)
gemma3-12b-sftgoogle/gemma-3-12b-it12B4bitGemma 3 12B instruction tuning
gemma3-27b-sftgoogle/gemma-3-27b-it27B4bitGemma 3 27B SFT with DeepSpeed ZeRO-2
gemma3-4b-sft-mlxmlx-community/gemma-3-4b-it-4bit4B4bitGemma 3 4B SFT on Apple Silicon via MLX (M1+ 16GB) (mlx)
glm-4.6-sftzai-org/GLM-4.6357B4bitGLM 4.6 MoE instruction tuning with LoRA
glm-5-sftzai-org/GLM-5754B4bitGLM 5 MoE SFT (next-gen GLM family)
glm-5.1-sftzai-org/GLM-5.1754B4bitGLM 5.1 MoE SFT (MIT, 754B). Multi-GPU / multi-node recommended.
gpt-oss-120b-sftopenai/gpt-oss-120b120B4bitGPT-OSS 120B SFT (multi-GPU recommended)
gpt-oss-20b-sftopenai/gpt-oss-20b20B4bitGPT-OSS 20B SFT (reasoning_effort=medium)
granite-4-sftibm-granite/granite-4.0-h-tiny-base7B4bitIBM Granite 4.0 tiny SFT
internvl-2.5-8b-sftOpenGVLab/InternVL2_5-8B8B4bitInternVL 2.5 8B vision-language SFT
internvl-3-5-sftOpenGVLab/InternVL3_5-8B8B4bitInternVL 3.5 vision SFT
kimi-k2-sftmoonshotai/Kimi-K2-BaseN/A4bitUnavailable since v0.77.0. The catalog now describes it as "Kimi K2 SFT: resident unsupported; guarded streaming only (see recipes use)." soup recipes use exits 3 and writes nothing: since v0.78.0 the message says the resident Kimi K2 route refuses training.quantization: 4bit, which this recipe sets (the route takes quantization: none; 4-bit goes through training.stream_layers: true). Evidence for either route is CPU parity on a tiny K2-shaped checkpoint; nothing trained on K2
kimi-k2.5-sftmoonshotai/Kimi-K2.51T4bitKimi K2.5 MoE SFT (Modified MIT, ~1T / 32B active). Requires multi-node DeepSpeed.
kimi-k2.6-sftmoonshotai/Kimi-K2.61T4bitKimi K2.6 MoE SFT (Modified MIT, ~1T / 32B active). Requires multi-node DeepSpeed.
lfm2-sftLiquidAI/LFM2-1.2B1.2B4bitLiquid LFM2 1.2B SFT (edge-optimised)
llama2-13b-finance-sftmeta-llama/Llama-2-13b-hf13B4bitLlama 2 13B finance-domain SFT (FinGPT-style starter recipe)
llama3-70b-fsdp2meta-llama/Llama-3.1-70B-Instruct70B4bitLlama 3.1 70B SFT with FSDP2 full shard + torch.compile. Requires 8 x A100/H100 80GB.
llama3.1-70b-sftmeta-llama/Llama-3.1-70B-Instruct70B4bitLlama 3.1 70B SFT with DeepSpeed ZeRO-3
llama3.1-8b-longctxmeta-llama/Llama-3.1-8B-Instruct8B4bitLlama 3.1 8B long-context (32k) with YaRN RoPE scaling
llama3.1-8b-sftmeta-llama/Llama-3.1-8B-Instruct8B4bitLlama 3.1 8B instruction tuning with LoRA
llama3.1-8b-sft-mlxmlx-community/Llama-3.1-8B-Instruct-4bit8B4bitLlama 3.1 8B SFT on Apple Silicon via MLX (M2+ 16GB) (mlx)
llama3.2-11b-visionmeta-llama/Llama-3.2-11B-Vision-Instruct11B4bitLlama 3.2 11B Vision multimodal fine-tuning
llama3.2-1b-sftmeta-llama/Llama-3.2-1B-Instruct1B8bitLlama 3.2 1B instruction tuning (mobile-friendly)
llama3.2-3b-sftmeta-llama/Llama-3.2-3B-Instruct3B4bitLlama 3.2 3B instruction tuning (edge-friendly)
llama3.2-vision-90b-sftmeta-llama/Llama-3.2-90B-Vision-Instruct90B4bitLlama 3.2 90B Vision multimodal SFT (8 x A100/H100 80GB)
llama4-scout-17b-sftmeta-llama/Llama-4-Scout-17B-16E-Instruct109B4bitLlama 4 Scout 17B active SFT with LoRA (4bit)
llama4-scout-toolsmeta-llama/Llama-4-Scout-17B-16E-Instruct109B4bitLlama 4 Scout 17B active tool-calling / function-calling SFT
llava-next-sftllava-hf/llava-v1.6-mistral-7b-hf7B4bitLLaVA-Next 7B vision SFT
magicoder-7b-sftise-uiuc/Magicoder-S-DS-6.7B6.7B4bitMagicoder S-DS 6.7B code-specialist SFT
magistral-small-sftmistralai/Magistral-Small24B4bitMagistral Small reasoning SFT
mathstral-7b-sftmistralai/Mathstral-7B-v0.17B4bitMathstral 7B math/STEM-specialist SFT
medgemma-sftgoogle/medgemma-4b-it4B4bitMedGemma 4B medical SFT
meditron-7b-sftepfl-llm/meditron-7b7B4bitMeditron 7B medical / clinical domain SFT
minicpm-v-2.6-sftopenbmb/MiniCPM-V-2_68B4bitMiniCPM-V 2.6 vision-language SFT (edge-friendly multimodal)
minimax-m2-sftMiniMaxAI/MiniMax-M2230B4bitMiniMax M2 MoE SFT instruction tuning
minimax-m3-sftMiniMaxAI/MiniMax-M3428B4bitMiniMax M3 MoE SFT (428B / 23B active). MiniMax Community License - commercial use requires a separate agreement. Multi-GPU recommended. Sets modality: vision, so SFT loads the vision-language wrapper, and its processor is custom code: run soup train --trust-remote-code. It no longer sets moe_lora, so no expert is adapted; target_modules: auto attaches to the language tower alone. The recipe's own comment says the vision collator supervises every token, prompt included, so train_on_responses_only is not applied on this path
ministral-sftmistralai/Ministral-8B-Instruct-24108B4bitMinistral 8B SFT
mistral-7b-sftmistralai/Mistral-7B-Instruct-v0.37B4bitMistral 7B instruction tuning
mistral-medium-3-5-sftmistralai/Mistral-Medium-3.5-128B128B4bitMistral Medium 3.5 SFT. Sets modality: vision, so SFT loads the vision-language wrapper (Mistral3ForConditionalGeneration) and target_modules: auto reaches the language model's attention projections, not the vision tower (#1395)
mistral-small-3-sftmistralai/Mistral-Small-24B-Instruct-250124B4bitMistral Small 3 24B SFT
nemotron-4-340b-sftnvidia/Nemotron-4-340B-Instruct340B4bitNemotron-4 340B SFT (massive multi-node deployment)
paddle-ocr-sftPaddlePaddle/PaddleOCR-VLN/A4bitPaddle-OCR-VL OCR SFT
phi3.5-mini-sftmicrosoft/Phi-3.5-mini-instruct3.8B4bitPhi-3.5-mini 3.8B SFT (small / edge)
phi4-14b-sftmicrosoft/phi-414B4bitPhi-4 14B instruction tuning
pixtral-12b-sftmistralai/Pixtral-12B-240912B4bitPixtral 12B vision-language SFT with LoRA
qvq-72b-sftQwen/QVQ-72B-Preview72B4bitQVQ 72B vision-reasoning SFT
qwen-image-sftQwen/Qwen-ImageN/A4bitQwen-Image image-output multimodal SFT
qwen2-audio-7b-sftQwen/Qwen2-Audio-7B-Instruct7B4bitQwen2-Audio 7B audio-language SFT
qwen2-vl-72b-sftQwen/Qwen2-VL-72B-Instruct72B4bitQwen2-VL 72B vision-language SFT (multi-GPU recommended)
qwen2-vl-7b-sftQwen/Qwen2-VL-7B-Instruct7B4bitQwen2-VL 7B vision-language SFT
qwen2.5-0.5b-sftQwen/Qwen2.5-0.5B-Instruct0.5B8bitQwen 2.5 0.5B SFT (mobile / edge)
qwen2.5-1.5b-sftQwen/Qwen2.5-1.5B-Instruct1.5B8bitQwen 2.5 1.5B SFT (edge-friendly)
qwen2.5-3b-sftQwen/Qwen2.5-3B-Instruct3B4bitQwen 2.5 3B SFT (small / edge)
qwen2.5-72b-sftQwen/Qwen2.5-72B-Instruct72B4bitQwen 2.5 72B SFT with DeepSpeed ZeRO-3
qwen2.5-7b-sftQwen/Qwen2.5-7B-Instruct7B4bitQwen 2.5 7B instruction tuning with LoRA
qwen2.5-coder-1.5b-sftQwen/Qwen2.5-Coder-1.5B-Instruct1.5B4bitQwen 2.5 Coder 1.5B instruction tuning with LoRA
qwen2.5-coder-14b-sftQwen/Qwen2.5-Coder-14B-Instruct14B4bitQwen 2.5 Coder 14B instruction tuning with LoRA
qwen2.5-coder-32b-sftQwen/Qwen2.5-Coder-32B-Instruct32B4bitQwen 2.5 Coder 32B instruction tuning with LoRA
qwen2.5-coder-7b-sftQwen/Qwen2.5-Coder-7B-Instruct7B4bitQwen 2.5 Coder 7B instruction tuning with LoRA
qwen2.5-math-1.5b-sftQwen/Qwen2.5-Math-1.5B-Instruct1.5B4bitQwen 2.5 Math 1.5B instruction tuning with LoRA
qwen2.5-math-7b-sftQwen/Qwen2.5-Math-7B-Instruct7B4bitQwen 2.5 Math 7B instruction tuning with LoRA
qwen3-14b-sftQwen/Qwen3-14B14B4bitQwen 3 14B instruction tuning with LoRA
qwen3-30b-a3b-sftQwen/Qwen3-30B-A3B30B4bitQwen 3 30B-A3B MoE instruction tuning
qwen3-32b-sftQwen/Qwen3-32B32B4bitQwen 3 32B SFT with DeepSpeed ZeRO-2
qwen3-32b-zeroppQwen/Qwen3-32B32B4bitQwen3 32B SFT with DeepSpeed ZeRO++ (quantized gradients + hierarchical partitioning). Launch with --deepspeed zero++.
qwen3-8b-sftQwen/Qwen3-8B8B4bitQwen 3 8B instruction tuning
qwen3-8b-sft-mlxmlx-community/Qwen3-8B-4bit8B4bitQwen 3 8B SFT on Apple Silicon via MLX (M2+ 16GB) (mlx)
qwen3-8b-toolsQwen/Qwen3-8B8B4bitQwen 3 8B tool-calling / function-calling SFT
qwen3-coder-30b-sftQwen/Qwen3-Coder-30B-A3B-Instruct30B4bitQwen3-Coder 30B (A3B MoE) code-specialist SFT
qwen3.5-0.8b-sftQwen/Qwen3.5-0.8B0.8B4bitQwen 3.5 0.8B SFT (Apache-2.0, tiny / mobile)
qwen3.5-122b-a10b-sftQwen/Qwen3.5-122B-A10B122B4bitQwen 3.5 122B-A10B MoE SFT (Apache-2.0, 10B active). Multi-GPU recommended (8 x A100/H100 80GB).
qwen3.5-27b-sftQwen/Qwen3.5-27B27B4bitQwen 3.5 27B SFT (Apache-2.0) with DeepSpeed ZeRO-2
qwen3.5-2b-sftQwen/Qwen3.5-2B2B4bitQwen 3.5 2B SFT (Apache-2.0, small / edge)
qwen3.5-35b-a3b-sftQwen/Qwen3.5-35B-A3B35B4bitQwen 3.5 35B-A3B MoE SFT (Apache-2.0, 3B active)
qwen3.5-397b-a17b-sftQwen/Qwen3.5-397B-A17B397B4bitQwen 3.5 397B-A17B flagship MoE SFT (Apache-2.0, 17B active). Requires multi-node DeepSpeed / FSDP.
qwen3.5-4b-sftQwen/Qwen3.5-4B4B4bitQwen 3.5 4B SFT (Apache-2.0, 262K context)
qwen3.5-9b-sftQwen/Qwen3.5-9B9B4bitQwen 3.5 9B SFT (Apache-2.0, 262K context)
qwen3.6-27b-sftQwen/Qwen3.6-27B27B4bitQwen 3.6 27B SFT (Apache-2.0) with DeepSpeed ZeRO-2
qwen3.6-35b-a3b-sftQwen/Qwen3.6-35B-A3B35B4bitQwen 3.6 35B-A3B MoE SFT (Apache-2.0, 3B active)
qwen3.8-27b-sftQwen/Qwen3.8-27B27B4bitQwen 3.8 27B text-only SFT (Apache-2.0) with DeepSpeed ZeRO-2
ra-dit-llama3-8bmeta-llama/Llama-3.1-8B-Instruct8B4bitRA-DIT stage 2 (Meta 2023) — RAFT-style SFT on the generator. Pairs with ra-dit-retriever (stage 1). Uses RAFT data format.
raft-llama3-8bmeta-llama/Llama-3.1-8B-Instruct8B4bitRAFT (Retrieval-Augmented Fine-Tuning, Stanford 2024) — train an 8B Llama 3.1 to answer queries given a golden doc + distractor docs. Rows: {query, golden_doc, distractor_docs, answer}.
seamlessm4t-v2-sftfacebook/seamless-m4t-v2-large2.3B4bitSeamlessM4T v2 multilingual speech-to-text SFT
smollm2-1.7b-sftHuggingFaceTB/SmolLM2-1.7B-Instruct1.7B8bitSmolLM2 1.7B SFT (small / edge)
smollm2-135m-sftHuggingFaceTB/SmolLM2-135M-Instruct135MnoneSmolLM2 135M SFT (ultra-tiny / mobile)
smollm2-360m-sftHuggingFaceTB/SmolLM2-360M-Instruct360MnoneSmolLM2 360M SFT (tiny / mobile)
smollm3-3b-sftHuggingFaceTB/SmolLM3-3B3B8bitSmolLM3 3B SFT (small / edge)
smolvlm-256m-sftHuggingFaceTB/SmolVLM-256M-Instruct256MnoneSmolVLM 256M vision SFT (llava format) — a tiny VLM. SmolVLM uses an Idefics3 processor; Soup mirrors its nested tokenizer surface and performs processor-aware vision collation so image-token expansion and pixel_values reach the model. target_modules are pinned to q_proj/v_proj (auto cannot infer them for Idefics3).
voxtral-sftmistralai/Voxtral-Mini-3B-25074.7B4bitVoxtral Mini 3B audio SFT
whisper-large-v3-ftopenai/whisper-large-v31.5B8bitWhisper Large v3 through the generic audio SFT path (modality: audio), which loads the model with AutoModel, the bare Whisper encoder-decoder without a language-model head. For Whisper speech-to-text use the ASR task, see ASR fine-tuning and whisper-large-v3-asr below

DPO: preference optimization

RecipeBase modelSizeQuantizationWhat it is
deepseek-r1-distill-llama-8b-dpodeepseek-ai/DeepSeek-R1-Distill-Llama-8B8B4bitDeepSeek-R1-Distill Llama 8B DPO alignment
deepseek-r1-distill-qwen-1.5b-dpodeepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B1.5B4bitDeepSeek-R1-Distill Qwen 1.5B DPO alignment
deepseek-r1-distill-qwen-7b-dpodeepseek-ai/DeepSeek-R1-Distill-Qwen-7B7B4bitDeepSeek-R1-Distill Qwen 7B DPO alignment
deepseek-v4-flash-dpodeepseek-ai/DeepSeek-V4-FlashN/A4bitDeepSeek V4 Flash MoE DPO alignment (MIT, efficiency-tier)
deepseek-v4-pro-dpodeepseek-ai/DeepSeek-V4-ProN/A4bitDeepSeek V4 Pro MoE DPO alignment (MIT, 1.6T-class). Sets moe_lora: true. The catalog says it requires multi-node DeepSpeed, which is a launch flag, see Before you run one
gemma3-27b-dpogoogle/gemma-3-27b-it27B4bitGemma 3 27B DPO alignment
glm-5.1-dpozai-org/GLM-5.1754B4bitGLM 5.1 MoE DPO alignment (MIT, 754B). Multi-GPU / multi-node recommended.
kimi-k2.5-dpomoonshotai/Kimi-K2.51T4bitKimi K2.5 MoE DPO alignment (Modified MIT, ~1T / 32B active). Sets moe_lora: true. The catalog says it requires multi-node DeepSpeed; its hyperparameters are inherited from the Kimi K2.6 recipe, not tuned for K2.5
kimi-k2.6-dpomoonshotai/Kimi-K2.61T4bitKimi K2.6 MoE DPO alignment (Modified MIT, ~1T / 32B active). Requires multi-node DeepSpeed.
llama3.1-8b-dpometa-llama/Llama-3.1-8B-Instruct8B4bitLlama 3.1 8B DPO alignment
llama4-scout-17b-dpometa-llama/Llama-4-Scout-17B-16E-Instruct109B4bitLlama 4 Scout 17B active DPO alignment
minimax-m3-dpoMiniMaxAI/MiniMax-M3428B4bitNew in v0.78.0 (#2719, by @abdumutalipabay0). MiniMax M3 MoE DPO alignment (428B / 23B active), a vision-language checkpoint trained through DPO's vision path on text rows (modality: vision, LoRA r 32 / alpha 64, lora.dropout: 0.0, batch 1, 16 accumulation steps); the processor is custom code, so pass --trust-remote-code. MiniMax Community License: commercial use needs a separate agreement. Multi-GPU recommended. Inherited settings, not tuned or measured (#275); not trained here
mistral-7b-dpomistralai/Mistral-7B-Instruct-v0.37B4bitMistral 7B DPO alignment
qwen2.5-7b-dpoQwen/Qwen2.5-7B-Instruct7B4bitQwen 2.5 7B DPO alignment
qwen3.5-0.8b-dpoQwen/Qwen3.5-0.8B0.8B4bitQwen 3.5 0.8B text DPO alignment (Apache-2.0, tiny / mobile). New in v0.77.2 (#2184, by @kok-o); sets modality: text and batch_size: auto. Settings inherited from the other Qwen 3.5 DPO recipes, not tuned or measured for 0.8B (#275)
qwen3.5-122b-a10b-dpoQwen/Qwen3.5-122B-A10B122B4bitNew in v0.78.0 (#2606, by @kok-o). Qwen 3.5 122B-A10B MoE text DPO alignment (Apache-2.0, 10B active). Sets modality: text, moe_lora: true, lora.dropout: 0.0, LoRA r 32 / alpha 64, batch 1, 16 accumulation steps and dpo_beta: 0.1. Inherited starting settings; the catalog says multi-GPU memory fit is unverified. Not trained here
qwen3.5-27b-dpoQwen/Qwen3.5-27B27B4bitQwen 3.5 27B text DPO alignment (Apache-2.0). New in v0.77.2 (#2445, by @kok-o); sets modality: text and gradient_accumulation_steps: 8. Settings inherited from the other Qwen 3.5 DPO recipes, not tuned or measured (#275)
qwen3.5-2b-dpoQwen/Qwen3.5-2B2B4bitQwen 3.5 2B text DPO alignment (Apache-2.0, small / edge). New in v0.77.0; sets modality: text, so it trains the language decoder on text preference pairs. Settings inherited from the other Qwen 3.5 DPO recipes; no training result is claimed
qwen3.5-35b-a3b-dpoQwen/Qwen3.5-35B-A3B35B4bitQwen 3.5 35B-A3B MoE DPO alignment (Apache-2.0, 3B active)
qwen3.5-4b-dpoQwen/Qwen3.5-4B4B4bitQwen 3.5 4B text DPO alignment (Apache-2.0). New in v0.77.0; sets modality: text. Settings inherited from the 9B DPO recipe; no training result is claimed
qwen3.5-9b-dpoQwen/Qwen3.5-9B9B4bitQwen 3.5 9B DPO alignment (Apache-2.0)
qwen3.6-27b-dpoQwen/Qwen3.6-27B27B4bitQwen 3.6 27B DPO alignment (Apache-2.0). The catalog says DeepSpeed ZeRO-2, which is a launch flag, not a key in the YAML. Sets modality: text
qwen3.6-35b-a3b-dpoQwen/Qwen3.6-35B-A3B35B4bitQwen 3.6 35B-A3B MoE DPO alignment (Apache-2.0, 3B active). Sets moe_lora: true and modality: text

GRPO: reinforcement learning with a verifier

RecipeBase modelSizeQuantizationWhat it is
deepseek-r1-32b-grpodeepseek-ai/DeepSeek-R1-Distill-Qwen-32B32B4bitDeepSeek R1 32B GRPO with DeepSpeed
deepseek-r1-8b-grpodeepseek-ai/DeepSeek-R1-Distill-Llama-8B8B4bitDeepSeek R1 8B GRPO reasoning
deepseek-v3-reasoningdeepseek-ai/DeepSeek-V3N/A4bitDeepSeek V3 (671B-parameter MoE) GRPO reasoning recipe. Requires multi-node DeepSpeed. Scores with the accuracy and format rewards; its verifiable_domain: math is only read by reward_fn: verifiable and has no effect here
deepseek-v4-flash-grpodeepseek-ai/DeepSeek-V4-FlashN/A4bitDeepSeek V4 Flash MoE GRPO reasoning training (MIT, efficiency-tier)
deepseek-v4-pro-grpodeepseek-ai/DeepSeek-V4-ProN/A4bitDeepSeek V4 Pro MoE GRPO reasoning training (MIT, 1.6T-class). Sets moe_lora: true and reward_fn: accuracy. The catalog says it requires multi-node DeepSpeed, which is a launch flag
glm-5.1-grpozai-org/GLM-5.1754B4bitGLM 5.1 MoE GRPO reasoning training (MIT, 754B). Multi-GPU / multi-node recommended.
grpo-env-calculatorHuggingFaceTB/SmolLM2-135M-Instruct135Mnot setSmolLM2-135M GRPO on the bundled calculator env (openenv rollout)
grpo-env-guess-numberHuggingFaceTB/SmolLM2-135M-Instruct135Mnot setSmolLM2-135M GRPO on the bundled number-deduction env (openenv rollout)
grpo-env-retrieval-qaHuggingFaceTB/SmolLM2-135M-Instruct135Mnot setSmolLM2-135M GRPO on the bundled retrieval-QA env (openenv rollout)
kimi-k2-thinking-grpomoonshotai/Kimi-K2-ThinkingN/A4bitUnavailable since v0.77.0. The catalog now describes it as "Kimi K2 Thinking GRPO: unsupported resident route; no supported route yet." soup recipes use exits 3 and writes nothing; Kimi-K2-Thinking has no supported route
kimi-k2.5-grpomoonshotai/Kimi-K2.51T4bitKimi K2.5 MoE GRPO reasoning (Modified MIT, ~1T / 32B active). Sets moe_lora: true and reward_fn: accuracy. The catalog says it requires multi-node DeepSpeed; its hyperparameters are inherited from the Kimi K2.6 recipe, not tuned for K2.5
kimi-k2.6-grpomoonshotai/Kimi-K2.61T4bitKimi K2.6 MoE GRPO reasoning (Modified MIT, ~1T / 32B active). Requires multi-node DeepSpeed.
llama3.1-8b-grpometa-llama/Llama-3.1-8B-Instruct8B4bitLlama 3.1 8B GRPO reasoning training
llama3.2-vision-grpometa-llama/Llama-3.2-11B-Vision-Instruct11B4bitLlama 3.2 11B Vision GRPO. The GRPO trainer is text-only at v0.77.0, so the images are not used, see GRPO Plus
llama4-scout-17b-grpometa-llama/Llama-4-Scout-17B-16E-Instruct109B4bitLlama 4 Scout 17B active GRPO reasoning training
phi4-reasoning-grpomicrosoft/phi-414B4bitPhi-4 14B GRPO reasoning training
qwen2.5-7b-grpoQwen/Qwen2.5-7B-Instruct7B4bitQwen 2.5 7B GRPO reasoning training
qwen3-30b-a3b-reasoning-grpoQwen/Qwen3-30B-A3B30B4bitQwen3 30B-A3B GRPO reasoning training (MoE thinking model)
qwen3-8b-grpoQwen/Qwen3-8B8B4bitQwen 3 8B GRPO reasoning training
qwen3.5-0.8b-grpoQwen/Qwen3.5-0.8B0.8B4bitQwen 3.5 0.8B GRPO reasoning training (Apache-2.0, tiny / mobile)
qwen3.5-122b-a10b-grpoQwen/Qwen3.5-122B-A10B122B4bitNew in v0.78.0 (#2660, by @nazek52). Qwen 3.5 122B-A10B text MoE GRPO (Apache-2.0, 10B active). Sets moe_lora: true, reward_fn: accuracy, num_generations: 4, batch 1 and 16 accumulation steps; each prompt needs a reference answer (an answer column or the final assistant message). The catalog calls it an untuned multi-GPU starting configuration with unverified memory fit. Not trained here
qwen3.5-2b-grpoQwen/Qwen3.5-2B2B4bitQwen 3.5 2B GRPO reasoning training (Apache-2.0, small / edge)
qwen3.5-27b-grpoQwen/Qwen3.5-27B27B4bitNew in v0.78.0 (#2605, by @kok-o). Qwen 3.5 27B text GRPO reasoning training (Apache-2.0). Same settings as the 4B recipe (batch_size: auto, 8 accumulation steps, four generations, accuracy reward, gradient checkpointing). Inherited, not tuned or measured; not trained here
qwen3.5-397b-a17b-grpoQwen/Qwen3.5-397B-A17B397B4bitNew in v0.78.0 (#2718, by @abdumutalipabay0). Qwen 3.5 397B-A17B text MoE GRPO (Apache-2.0, 17B active). Sets moe_lora: true, reward_fn: accuracy, batch 1 and 32 accumulation steps. The catalog calls it an untuned multi-node starting configuration with unverified memory fit; its SFT sibling already needs multi-node DeepSpeed or FSDP. Not trained here
qwen3.5-4b-grpoQwen/Qwen3.5-4B4B4bitNew in v0.78.0 (#2522, by @kok-o). Qwen 3.5 4B text GRPO reasoning training (Apache-2.0). Inherits the 0.8B, 2B and 9B GRPO starting settings: batch_size: auto, 8 accumulation steps, four generations, reward_fn: accuracy, gradient checkpointing, 4096 tokens. Each prompt needs a reference answer. Not trained here
qwen3.5-35b-a3b-grpoQwen/Qwen3.5-35B-A3B35B4bitQwen 3.5 35B-A3B MoE GRPO reasoning training (Apache-2.0, 3B active). Completes the SFT, DPO and GRPO set for this base. Sets moe_lora: true and modality: text
qwen3.5-9b-grpoQwen/Qwen3.5-9B9B4bitQwen 3.5 9B GRPO reasoning training (Apache-2.0, 262K context)
qwen3.6-27b-grpoQwen/Qwen3.6-27B27B4bitQwen 3.6 27B GRPO reasoning training (Apache-2.0). The catalog says DeepSpeed ZeRO-2, which is a launch flag, not a key in the YAML. Sets reward_fn: accuracy and modality: text
qwen3.6-35b-a3b-grpoQwen/Qwen3.6-35B-A3B35B4bitQwen 3.6 35B-A3B MoE GRPO reasoning training (Apache-2.0, 3B active). Sets moe_lora: true and modality: text
qwq-32b-grpoQwen/QwQ-32B32B4bitQwQ 32B GRPO reasoning training
r1-distill-llama-70b-grpodeepseek-ai/DeepSeek-R1-Distill-Llama-70B70B4bitDeepSeek-R1-Distill Llama 70B GRPO reasoning (multi-GPU)
r1-distill-qwen-1.5b-grpodeepseek-ai/DeepSeek-R1-Distill-Qwen-1.5B1.5B4bitDeepSeek-R1-Distill Qwen 1.5B GRPO reasoning training
r1-distill-qwen-14b-grpodeepseek-ai/DeepSeek-R1-Distill-Qwen-14B14B4bitDeepSeek-R1-Distill Qwen 14B GRPO reasoning training
r1-distill-qwen-7b-grpodeepseek-ai/DeepSeek-R1-Distill-Qwen-7B7B4bitDeepSeek-R1-Distill Qwen 7B GRPO reasoning training

KTO

RecipeBase modelSizeQuantizationWhat it is
llama3.1-8b-ktometa-llama/Llama-3.1-8B-Instruct8B4bitLlama 3.1 8B KTO unpaired preference alignment

ORPO

RecipeBase modelSizeQuantizationWhat it is
llama3.1-8b-orpometa-llama/Llama-3.1-8B-Instruct8B4bitLlama 3.1 8B ORPO reference-free alignment

SimPO

RecipeBase modelSizeQuantizationWhat it is
llama3.1-8b-simpometa-llama/Llama-3.1-8B-Instruct8B4bitLlama 3.1 8B SimPO length-normalized alignment

IPO

RecipeBase modelSizeQuantizationWhat it is
llama3.1-8b-ipometa-llama/Llama-3.1-8B-Instruct8B4bitLlama 3.1 8B IPO regularized preference alignment

PPO

RecipeBase modelSizeQuantizationWhat it is
llama3.1-8b-ppometa-llama/Llama-3.1-8B-Instruct8B4bitLlama 3.1 8B PPO (RLHF stage 3)

Online DPO

RecipeBase modelSizeQuantizationWhat it is
online-dpo-smollm2-135mHuggingFaceTB/SmolLM2-135M-Instruct135MnoneSmolLM2 135M Online DPO — on-policy generation judged by a pairwise LLM judge (the recipe sets training.online_dpo_judge: ollama://llama3.1, which needs a local Ollama server; point it at your own judge)

Reward model

RecipeBase modelSizeQuantizationWhat it is
qwen2.5-7b-rewardQwen/Qwen2.5-7B-Instruct7B4bitQwen 2.5 7B reward model (RLHF stage 2)

Embedding

RecipeBase modelSizeQuantizationWhat it is
embedding-gemma-sftgoogle/embeddinggemma-300m300M4bitEmbeddingGemma 300M sentence-embedding SFT
llama3.1-8b-embedmeta-llama/Llama-3.1-8B8B4bitLlama 3.1 8B sentence embedding with cosine loss
ra-dit-retrieversentence-transformers/all-mpnet-base-v2N/Anot setRA-DIT stage 1 (Meta 2023) — contrastive retriever training. Pairs with ra-dit-llama3-8b (stage 2) for the full pipeline.

Continued pre-training

RecipeBase modelSizeQuantizationWhat it is
qwen2.5-7b-pretrainQwen/Qwen2.5-7B7B4bitQwen 2.5 7B continued pre-training
qwen3.5-2b-pretrainQwen/Qwen3.5-2B-Base2B4bitNew in v0.78.0 (#2721, by @abdumutalipabay0). Qwen 3.5 2B Base continued pre-training (Apache-2.0) on plaintext rows, with the qwen3.5-4b-pretrain settings, inherited and not tuned for 2B. Not trained here
qwen3.5-4b-pretrainQwen/Qwen3.5-4B-Base4B4bitQwen 3.5 4B Base continued pre-training (Apache-2.0)
qwen3.5-9b-pretrainQwen/Qwen3.5-9B-Base9B4bitNew in v0.78.0 (#2720, by @abdumutalipabay0). Qwen 3.5 9B Base continued pre-training (Apache-2.0) on plaintext rows. Takes the qwen3.5-4b-pretrain settings (one epoch, lr: 1e-4, batch_size: auto, 8 accumulation steps, LoRA r 32 / alpha 64, 4096 tokens), inherited and not tuned for 9B. Not trained here

ASR: speech recognition

RecipeBase modelSizeQuantizationWhat it is
whisper-base-asropenai/whisper-base74MnoneWhisper base (74M) ASR fine-tune with LoRA on q/v projections — fits a 4 GB GPU
whisper-large-v3-asropenai/whisper-large-v31.5BnoneWhisper large-v3 (1.5B) ASR fine-tune with LoRA — needs a larger GPU (>= 16 GB); the tiny/base recipes fit a 4 GB card
whisper-tiny-asropenai/whisper-tiny39MnoneWhisper tiny (39M) ASR fine-tune with LoRA on q/v projections — fits a 4 GB GPU. Rows: {"audio": path, "text": transcript}

TTS: text to speech

RecipeBase modelSizeQuantizationWhat it is
llasa-ttsHKUSTAudio/Llasa-1B1Bnot setLlasa-TTS. Sets data.format: audio, the live-codec path: raw audio is encoded with XCodec2 at training time and needs pip install "soup-cli[audio]", see TTS
orpheus-tts-sftcanopylabs/orpheus-3b-0.1-ft3Bnot setOrpheus emotional TTS. Its data.format: audio is the live-codec path and needs pip install snac; the pre-encoded path needs no extra package, see TTS
oute-ttsOuteAI/OuteTTS-0.3-500M0.5Bnot setOuteTTS v0.3 on pre-encoded data. Sets data.format: chatml, so your rows must already be codec tokens as text; a raw-audio (data.format: audio) Oute config is refused at config load, see TTS

See also

  • Recipes: how to list, search, show and use them, and soup recipes verify to check that a recipe's adapter attaches before you download weights.
  • Configuration: every key a recipe can set.
  • Quant Menu: what each quantization value does.
  • Config refusals: the settings v0.77.0 refuses at config load; run soup doctor --config soup.yaml on a recipe you have edited.

Soup is free and Apache-2.0. If it saved you a training run, starring the repo costs nothing and helps most. You can also fund the GPU time behind the work a 4 GB laptop cannot reach.