Clover 2 open-sourced: 27.8B base with per-token routing over 7 LoRA experts
bakatristan · reddit · 2026-11-17
Kitani AI's team released Clover 2 with everything public on Hugging Face (kitaniai/clover-2).
- Architecture: a dense 27.8B base plus 7 trained LoRA experts (roleplay, general assistant, security, math, long-document, efficiency, knowledge), with a trained router picking expert(s) per token at inference; a null slot falls back to the plain base — not a static merge
- Benchmarks: meaningful gains over the untouched base on GSM8K/MATH-500/GSM-Hard, a bump on LongBench-v2, and fewer generic AI-writing tics in creative writing
- Open assets: base model, all 7 PEFT-format adapters, the trained router, and a vLLM plugin for true per-token behavior
- Self-hosting: no GGUF yet; hybrid-attention llama.cpp support is bleeding-edge, so vLLM is the supported path. Free trial for a week, ZDR supported
More from Models
- Sage releases open S2.1 Flash science multimodal model, tops MLE-bench Lite at 90.3 — QuixiAI · 2026-11-17
- Amp vs Claude Code hands-on: user finds Amp gets more done even as $20-tier Fable 5.2 access ends — bytebot · 2026-11-17
- Skeptic Pushes Back on Qwen4 Science SOTA Claim: Not Open Source, Marginal Edge — teortaxesTex · 2026-11-17
- miHoYo unveils Gamer, a generalist AI agent that plays games by watching the screen — minchoi · 2026-11-17
- Codex Pro's $200 plan includes $200/month in API credits, saving devs real money — burkov · 2026-11-17
- Claim: Claude covertly halted a game dev's GPU deployments after refusing a clone job — RexDouglass · 2026-11-17