MakazhanAlpamys/Soup v0.71.24

v0.71.24 — 2026 model-family recipes (catalog 116 → 133)

on GitHub

What's New

17 new ready-made SFT recipes for the open-weight models released Feb–Jun 2026 — every base repo-ID verified to resolve on Hugging Face before shipping. Grab one with soup recipes use <name> or browse all 133 via soup recipes list.

• Qwen 3.5 (Apache-2.0): qwen3.5-0.8b-sft, qwen3.5-2b-sft, qwen3.5-4b-sft, qwen3.5-9b-sft, qwen3.5-27b-sft, and the MoE sizes qwen3.5-35b-a3b-sft / qwen3.5-122b-a10b-sft / qwen3.5-397b-a17b-sft.
• Qwen 3.6 (Apache-2.0): qwen3.6-27b-sft, qwen3.6-35b-a3b-sft.
• DeepSeek-V4 (MIT): deepseek-v4-flash-sft, deepseek-v4-pro-sft.
• GLM (MIT): glm-5.1-sft.
• Kimi (Modified MIT): kimi-k2.5-sft, kimi-k2.6-sft.
• MiniMax (MiniMax Community License — commercial use needs a separate agreement): minimax-m3-sft.
• Mistral Large 3 (Apache-2.0, 675B / 41B-active multimodal MoE): mistral-large-3-sft.
• Fix: the stale glm-5-sft recipe now points at zai-org/GLM-5 (the org migrated from THUDM).

Install/Upgrade

pip install -U soup-cli
soup recipes list          # 133 recipes
soup recipes use qwen3.5-9b-sft

Security

No new attack surface — this is a pure config/data release. Recipe YAML is parsed via the existing Pydantic + yaml.safe_load chain; new model-id strings are validated for owner/name shape with no path-traversal or shell metacharacters.

Known Limitations

• The multi-GPU giant recipes (122B/397B, DeepSeek-V4-Pro, GLM-5.1, Kimi, MiniMax, Mistral-Large-3) are validated by YAML parse only, not trained end-to-end — they ship the same multi-GPU shape as the existing qwen3-32b-zeropp / deepseek-v3-pipeline entries (advisory; same infra-blocked class as other large-model recipes).
• Repo-IDs were verified at release time; if an upstream org later renames or removes a repo, soup recipes show/use still works and only the eventual soup train download would fail with a friendly Hugging Face error.

🤖 Generated with Claude Code