Zen 6 Nano (zenlm/zen6-nano)

Zen 6 Nano is an open-weight Apache-2.0 1.58-bit ternary multimodal & agentic language model (9B default + 4B & 2B CPU/CI tiers) developed by Hanzo AI and Zen LM in collaboration with Zoo Labs Foundation.

Post-trained directly in the Sylvester-Walsh-Hadamard ternary weight space via Straight-Through Estimator (STE) Quantization-Aware Training (QAT) with Hybrid Muon + AdamW, zen6-nano combines:

  • Authentic Weight-Space Identity: Native zen6-nano provenance with zero system-prompt wrapper overhead, while strictly respecting custom developer personas (e.g., Godot game NPCs, JSON extractors).
  • Agentic Tool Calling & Hanzo MCP: Multi-step <tool_call> execution, <tool_response> synthesis, error recovery, and calibrated tool abstention across fs, exec, code, git, test, hanzo, and game-engine director tools.
  • Runtimes: PQ2_0 / PTQ1_0 / Q2_0 run natively on hanzo-engine (CUDA, Metal, ROCm, Vulkan, CPU) with MTP, DFlash2, SA_SPEC suffix-automaton lookup, and GuardedDraft, as well as llama.cpp (Q2_0 on upstream b11556+). MLX 2-bit packs run with mlx-lm on Apple Silicon.
  • Zero-Token-Burn Cloud CI/CD: Load zen6-nano-2b-PTQ1_0.gguf (~429 MB) on CPU runners for end-to-end LLM and tool-calling integration tests with no API keys or per-token cost.

Weight Artifacts

File Format Description
zen6-nano-9b-PQ2_0.gguf PQ2_0 (2.13 bpw) 2.42 GB โ€” Default zen6-nano 9B model (GPU & desktop)
zen6-nano-9b-PTQ1_0.gguf PTQ1_0 (1.75 bpw) 2.00 GB โ€” Compact 1.75-bit 9B ternary pack
zen6-nano-9b-Q2_0.gguf Q2_0 (2.25 bpw) 2.56 GB โ€” Group-64 9B ternary band
zen6-nano-9b-PQ2_0-MTP-Q8_0.gguf PQ2_0 + MTP 2.68 GB โ€” 9B with Multi-Token Prediction speculative head
zen6-nano-9b-DFlash2-Q8_0.gguf Q8_0 Draft 1.57 GB โ€” 9B DFlash2 block-diffusion speculative draft
zen6-nano-9b-mmproj-Q8_0.gguf Q8_0 624 MB โ€” 9B Multimodal vision projector
zen6-nano-9b-mmproj-BF16.gguf BF16 922 MB โ€” 9B Full-precision vision projector
zen6-nano-4b-PQ2_0.gguf PQ2_0 (2.13 bpw) 1.15 GB โ€” 4B laptop / mobile / game NPC tier
zen6-nano-4b-PTQ1_0.gguf PTQ1_0 (1.75 bpw) 950 MB โ€” 4B sub-1GB ternary pack
zen6-nano-2b-PQ2_0.gguf PQ2_0 (2.13 bpw) 517 MB โ€” 2B ultra-fast CPU & edge tier
zen6-nano-2b-PTQ1_0.gguf PTQ1_0 (1.75 bpw) 429 MB โ€” Default zero-token-burn Cloud CI/CD CPU tier

Quickstart โ€” Local Stack & Game Backend (hanzo up)

# Pull & serve zenlm/zen6-nano locally on http://127.0.0.1:36900/v1
hanzo pull zenlm/zen6-nano
hanzo engine serve zenlm/zen6-nano

# Or bring up the full local Hanzo stack + open-weight AI
hanzo up

OpenAI-Compatible /v1/chat/completions + MCP Tools

from openai import OpenAI

client = OpenAI(base_url="http://127.0.0.1:36900/v1", api_key="hanzo-local")
resp = client.chat.completions.create(
    model="zen6-nano",
    messages=[{"role": "user", "content": "Check Cargo.toml and run the workspace unit tests."}],
    tools=[{
        "type": "function",
        "function": {
            "name": "fs",
            "description": "Read or list workspace files",
            "parameters": {
                "type": "object",
                "properties": {"action": {"type": "string"}, "path": {"type": "string"}},
                "required": ["action", "path"],
            },
        },
    }],
)
print(resp.choices[0].message)

Zero-Token-Burn GitHub Actions CPU CI

- name: Serve open-weight Zen 6 Nano on CPU
  run: |
    hf download hanzoai/zen6-nano zen6-nano-2b-PTQ1_0.gguf --local-dir ./models
    llama-server -m ./models/zen6-nano-2b-PTQ1_0.gguf -ngl 0 --port 36900 --alias zen6-nano &

Architecture & Hadamard Ternary Specification

  • Hybrid Backbone: Interleaved Gated-DeltaNet (GDN) linear-attention blocks (3x) and Grouped-Query Full Attention blocks (1x) with SwiGLU MLPs and multimodal vision projector (mmproj).
  • Hadamard Outlier Rotation: Linear weights and token embeddings are stored in an orthonormal block Sylvester-Walsh-Hadamard basis ($H_B[r,c] = (-1)^{\text{popcount}(r & c)}/\sqrt{B}$ with explicit per-channel $\pm 1$ sign vectors embedded in GGUF KV metadata prism.hadamard.*), eliminating activation outliers so group-128 ternary {-d, 0, +d} weights retain full dense precision.
  • Production Formats:
    • PQ2_0 (GGML type 142, 34 B / 128 weights = 2.125 bpw): Fastest SIMD / CUDA dp4a / ROCm / Metal execution (zen6-nano-9b-PQ2_0.gguf).
    • PTQ1_0 (GGML type 143, 28 B / 128 weights = 1.750 bpw): Smallest disk & RAM footprint (5 trits/byte base-3 packing).
    • Q2_0 (GGML type 42, 18 B / 64 weights = 2.250 bpw): Group-64 compatibility band.

License & Attribution

Released under the Apache License 2.0 (LICENSE). Preserves full upstream attribution to Qwen (Alibaba Cloud), Bonsai-2 (Prism ML, Inc.), and DFlash / ProCreations (NOTICE).

Downloads last month
1,744
GGUF
Model size
2B params
Architecture
qwen35
Hardware compatibility
Log In to add your hardware

1-bit

2-bit

8-bit

Inference Providers NEW
This model isn't deployed by any Inference Provider. ๐Ÿ™‹ Ask for provider support