Architecture: Qwen3.5 (hybrid LinearAttention + standard attention)
Parameters: 27B (4-bit quantized)
Format: MNN (Alibaba Mobile Neural Network)
Quantization: W4A16 (4-bit weights, 16-bit activations)
Attention: 48 LinearAttention layers (gated_delta_rule) + 16 standard attention layers

Files

Website: <a href="https://tokforge.ai" target="_blank" rel="noopener noreferrer" class="text-blue-400 hover:text-blue-300 underline underline-offset-2 transition-colors">https://tokforge.ai
Discord: <a href="https://discord.gg/Acv3CBtfVm" target="_blank" rel="noopener noreferrer" class="text-blue-400 hover:text-blue-300 underline underline-offset-2 transition-colors">https://discord.gg/Acv3CBtfVm
Google Play: <a href="https://play.google.com/store/apps/details?id=dev.tokforge" target="_blank" rel="noopener noreferrer" class="text-blue-400 hover:text-blue-300 underline underline-offset-2 transition-colors">https://play.google.com/store/apps/details?id=dev.tokforge
iOS TestFlight: <a href="https://testflight.apple.com/join/jnufjzRr" target="_blank" rel="noopener noreferrer" class="text-blue-400 hover:text-blue-300 underline underline-offset-2 transition-colors">https://testflight.apple.com/join/jnufjzRr

File	Size	Description
`llm.mnn`	8.5MB	Model graph
`llm.mnn.weight`	14GB	Quantized weights
`embeddings_bf16.bin`	2.4GB	BF16 embedding table (required)
`llm_config.json`	4.4KB	Model config with jinja chat template
`tokenizer.txt`	2.9MB	Tokenizer

Requirements

RAM: 24GB+ device recommended
Backend: CPU recommended (LinearAttention models have reduced OpenCL performance)

Optimized for TokForge — an Android app for on-device LLM inference.

Safety filters removed for unrestricted conversation. Use responsibly.

Intended for TokForge / MNN power users with very high-memory devices.
Qwen3.5 hybrid LinearAttention models are usually better on CPU than OpenCL in current TokForge routing.
This is a runtime bundle for on-device inference, not a standard Transformers training repo.
Large-model mobile performance depends heavily on device RAM, swap behavior, and OEM background-killer policies.