1bit.MONSTERDocs GitHub ↗

Model Families

1bit auto-detects 552 architecture tokens / 1,774 HF arch strings from GGUF/1BP/safetensors headers — no config files, no per-model glue code. Point the engine at a model and run. Coverage is measured against a full HF census: 317,310 / 317,310 arch-bearing text-gen checkpoints (100.00%) map to an engine token — see the models SSOT.

Every family below has its own page with a full breakdown: parameter sizes, 1BP file size, supported backends, and real measured performance. Numbers trace back to the performance SSOT; model-support detail traces back to wiki/models.md.

Backend legend: 🟢 supported & validated · 🟡 functional, perf pending · 🔴 not yet · 🔬 experimental Pipeline legend: 🧠 LLM · 👁️ vision · 🗣️ voice/TTS · 🧬 EEG · 🏁 end-to-end validated


The flagship ecosystem — a complete stack from EEG → LLM (dense, MoE, Mamba) → TTS → voice cloning, all running on the same binary.

Zyphra family page


All families

Family Type Sizes Page
Zyphra (Zaya1, ZR1, BlackMamba, Zamba/Zamba2) MoE · SSM · dense 1.2B–74B zyphra.md
Qwen (Qwen2/2.5/3/3.5/3.6, Qwen-VL) Dense · MoE · VL 0.5B–72B qwen.md
Llama (3.1 / 3.2, TinyLlama) Dense 1B–8B llama.md
Mistral (Mistral, Ministral, Pixtral) Dense · VL 7B–12B mistral.md
Gemma (Gemma 3/4, MedGemma, Embedding-Gemma) Dense · embedding 300M–4B gemma.md
Phi (Phi-3 / 3.5 / 4-mini) Dense 3.8B phi.md
Falcon (Falcon3) Dense (parallel attn) 1B–40B falcon.md
OLMo (OLMo-2) Dense (LayerNorm) 7B–13B olmo.md
Granite (Granite-3.2) Dense 2B–8B granite.md
SmolLM (SmolLM2) Dense 135M–1.7B smollm.md
DeepSeek (V2/V3/R1 + distills) MoE (MLA) 8B–671B deepseek.md
GPT-OSS MoE 20B gpt-oss.md
Laguna (Poolside) Dense · MoE 3B–7B laguna.md
Moonshot Kimi (Moonlight, Kimi-VL) Gated MLA MoE 16B (3B active) kimi.md
BitNet / Bonsai (Deepgrove) Ternary-native (TQ2) 1.7B–27B bitnet-bonsai.md
Whisper Speech-to-text V3 Turbo whisper.md

The Mage-ViT / Mage-VL vision path and Whisper speech path together feed the JARVIS pipeline.

1BP format note

Dense (non-ternary-trained) models use Q4NX 4-bit; only ternary-native checkpoints (BitNet/Bonsai) use TQ2 2-bit. Converting a dense model to TQ2 is quality-destructive — see the 1BP format policy.