edtorre/ qwen3.6-hermes

27 5 days ago

Qwen 3.6 27B (Q4_K_M) optimized for Hermes Agent — 64K context, 8192 max tokens, MTP for speed, flash attention + Q8 KV cache.

vision tools thinking 
ollama run edtorre/qwen3.6-hermes

Applications

Claude Code
Claude Code ollama launch claude --model edtorre/qwen3.6-hermes
Codex App
Codex App ollama launch codex-app --model edtorre/qwen3.6-hermes
OpenClaw
OpenClaw ollama launch openclaw --model edtorre/qwen3.6-hermes
Hermes Agent
Hermes Agent ollama launch hermes --model edtorre/qwen3.6-hermes
Codex
Codex ollama launch codex --model edtorre/qwen3.6-hermes
OpenCode
OpenCode ollama launch opencode --model edtorre/qwen3.6-hermes

Models

View all →

1 model

qwen3.6-hermes:latest

18GB · 256K context window · Text, Image · 5 days ago

Readme

edtorre/qwen3.6-hermes - Best for 24+ gb vram setups.

A powerful, agent-tuned Qwen 3.6 27B model for local agent harnesses like Hermes Agent. Based on qwen3.6:27b-mtp-q4_K_M with native multi-token prediction (MTP) for faster generation.

Specs

┌──────────────────┬────────────────────────┐
│ Parameter │ Value │
├──────────────────┼────────────────────────┤
│ Base Model │ qwen3.6:27b-mtp-q4_K_M │
├──────────────────┼────────────────────────┤
│ Parameters │ 27.3B │
├──────────────────┼────────────────────────┤
│ Quantization │ Q4_K_M │
├──────────────────┼────────────────────────┤
│ Size │ 17 GB │
├──────────────────┼────────────────────────┤
│ Context Window │ 65,536 (64K) │
├──────────────────┼────────────────────────┤
│ Max Output │ 8,192 tokens │
├──────────────────┼────────────────────────┤
│ Temperature │ 0.7 │
├──────────────────┼────────────────────────┤
│ Top-k / Top-p │ 20 / 0.9 │
├──────────────────┼────────────────────────┤
│ Presence Penalty │ 1.2 │
├──────────────────┼────────────────────────┤
│ Draft Predict │ 4 (MTP enabled) │
└──────────────────┴────────────────────────┘

Capabilities

  • ✅ Tools / function calling
  • ✅ Vision
  • ✅ Audio
  • ✅ Thinking

Performance Notes

At 17 GB + 64K context, this model is tight on 20 GB VRAM. Requires these Ollama server settings:

   OLLAMA_FLASH_ATTENTION=1                                                                                                                                                                   
   OLLAMA_KV_CACHE_TYPE=q8_0

With flash attention and Q8_0 KV cache, total VRAM usage is approximately 17-18 GB — fits on a 20 GB card. Native MTP (draft_num_predict 4) provides faster token generation when available.

Quick Start

   ollama run edtorre/qwen3.6-hermes

With Hermes Agent:

   hermes config set model.default edtorre/qwen3.6-hermes                                                                                                                                     
   hermes chat

License

Apache 2.0

qwen3.6_hermes.jpeg