Pure Rust multimodal inference engine based on Qwen3.5-0.8B. No Python, no CUDA, no external ML frameworks. Single executable + model weights = portable AI that runs on any machine.

Designed for mobile and edge devices — only 600 MB model file, loads in under 1 second, and runs at ~4 tok/s on a standard CPU. Smart system awareness — automatically detects your hardware (RAM, CPU threads) on Windows, Linux, and macOS, and adjusts generation parameters so the model runs well even on constrained systems.

License

This project is licensed under Apache 2.0. The base model Qwen3.5-0.8B is released by the Qwen team under Apache 2.0.

What It Does

QORA-0.8B is a 0.8-billion parameter language model with built-in vision. It can:

Text generation — answer questions, write code, summarize text
Image understanding — describe photos, answer questions about images
Video understanding — analyze frame sequences, describe motion and temporal changes
Thinking mode — chain-of-thought reasoning with configurable budget

Architecture

QORA-0.8B uses a hybrid architecture combining two attention mechanisms:

Component	Details
Parameters	0.8B total
Hidden dim	1024
Layers	24 (18 DeltaNet + 6 Full Attention)
Layer pattern	3x DeltaNet + 1x Full Attention, repeated 6 times
Vocabulary	248,320 tokens
Context	262K tokens natively

DeltaNet Layers (18 of 24)

Gated linear attention with delta rule state updates
16 QK heads + 16 V heads, head_dim=128
Causal Conv1d (kernel=4) + SiLU activation
O(1) memory per token (recurrent state, no KV cache needed)

Full Attention Layers (6 of 24)

Grouped Query Attention (8Q / 2KV heads), head_dim=256
QK-norm + partial RoPE (64/256 dims rotated), theta=10M
Output gating (sigmoid gate on attention output)
Standard KV cache

Vision Encoder

12-layer ViT, hidden=768, 12 heads
Conv3d patch embedding [768, 3, 2, 16, 16] (temporal_patch_size=2)
Learned positional embedding with bilinear interpolation from 48x48 grid
2D spatial RoPE (dim=32, theta=10000)
2x2 spatial merger: LayerNorm → concat → MLP(3072 → 1024)
Images: single frame duplicated along temporal axis
Video: actual Conv3d over consecutive frame pairs (N frames → N/2 temporal patches)

Smart System Awareness

QORA-0.8B detects your system at startup and automatically adjusts generation limits:

QORA-0.8B - Pure Rust Multimodal Inference Engine
System: 16384 MB RAM (9856 MB free), 12 threads

Available RAM	Think Budget	Max Tokens	Behavior
< 4 GB	128 (cap 256)	256 (cap 512)	Minimal generation, warning displayed
4-8 GB	256 (cap 1024)	512 (cap 1024)	Constrained, warning displayed
8-12 GB	1024 (cap 2048)	1024 (cap 2048)	Normal operation
>= 12 GB	2048 (cap 8192)	2048 (cap 8192)	Full capability

Hard caps apply even to explicit user values — if you pass --max-tokens 5000 on a system with 6 GB free RAM, it gets clamped to 1024 automatically. This prevents the model from running for too long on weak systems.

Supports Windows (wmic), Linux (/proc/meminfo), and macOS (sysctl/vm_stat).

Weight Format

Format	Size	Quality	Speed
Q4 (default)	~600 MB	Good	~3.9 tok/s

Q4 uses 4-bit symmetric quantization with group_size=32 and LUT-optimized dequantization. Multi-threaded GEMV/GEMM via rayon for large matrices.

AVX-512 SIMD Acceleration

On CPUs with AVX-512 support (Intel 11th gen+, AMD Zen 4+), QORA-0.8B automatically uses hand-written AVX-512 SIMD kernels for significant CPU speedup:

Kernel	Technique	Speedup
Q4 GEMV	`permutexvar_ps` 16-entry LUT lookup, nibble extract via `cvtepu8_epi32`	~2.5x
F16 GEMV	`cvtph_ps` f16→f32 + `fmadd_ps` FMA accumulation	~2.5x
DeltaNet state	Vectorized decay/retrieve/delta/output over 128-dim heads	~3x
Fused gate+up	Parallel gate & up SIMD LUT decode in MLP	~2.5x

Detection is automatic at runtime — falls back to scalar code on non-AVX-512 CPUs with zero overhead.

The model is small enough that Q4 is the only format needed — it loads in under 1 second and uses minimal RAM.

Platform Support

Platform	Binary	Status
Windows x86_64	`qor08b.exe`	Tested
Linux x86_64	`qor08b`	Supported
macOS aarch64	`qor08b`	Supported

CPU-only by design — the 0.8B model is small and fast enough that GPU is not needed. Pre-built binaries are available on the Releases page.

Quick Start

Download from the Releases page:
- model.qor08b (600 MB)
- tokenizer.json (13 MB)
- qor08b.exe (Windows) or build from source (Linux/macOS)
Run:

bash

# Text generation
qor08b --prompt "Explain quantum computing" --max-tokens 500

# Image understanding
qor08b --prompt "What's in this image?" --image photo.jpg

# Video understanding (directory of frame images)
qor08b --prompt "What happens in this video?" --video frames_dir/

# Thinking mode (default, extended reasoning)
qor08b --prompt "What is the capital of France?" --think-budget 512

# No-think mode (faster, direct answers)
qor08b --prompt "What is 2+2?" --no-think

# Greedy decoding (deterministic output)
qor08b --prompt "Hello" --greedy

CLI Flags

Flag	Description
`--prompt TEXT`	Input prompt (default: "Hello, how are you?")
`--image PATH`	Path to an image file (PNG/JPG)
`--video PATH`	Path to directory of frame images (PNG/JPG, sorted by name)
`--max-tokens N`	Max tokens to generate (default: 1024)
`--think-budget N`	Max thinking tokens before forcing answer (default: 1024)
`--no-think`	Disable thinking mode (direct answers)
`--show-think`	Display thinking tokens on stderr
`--greedy`	Greedy decoding (temperature=0, not recommended with thinking mode)

Sampling Defaults

Parameter	Think mode	No-think mode
temperature	1.0	0.7
top_k	20	20
top_p	0.95	0.95
presence_penalty	1.5	1.5

Video Input

Video is provided as a directory of frame images (not a video file). Extract frames however you like:

bash

# Example: extract 4 frames from a video with ffmpeg
ffmpeg -i video.mp4 -vf "select=not(mod(n\,30))" -frames:v 4 frames/frame_%02d.png

# Then run
qor08b --prompt "Describe what happens" --video frames/

Frames are loaded in alphabetical order, resized to uniform dimensions (max 768px, divisible by 32), and processed as temporal pairs via Conv3d. Odd frame counts are padded by duplicating the last frame.

Building from Source

bash

# All platforms (CPU-only, no GPU needed)
cargo build --release

Dependencies

Language: Pure Rust (2024 edition)
cortex — Rust deep learning framework (used for binary format types only)
rayon — Thread pool for parallel GEMV, attention
half — F16 support
image — Image loading (PNG/JPG)
tokenizers — HuggingFace tokenizer
memmap2 — Memory-mapped I/O for converter
serde_json — Config parsing
No ML framework for inference — all matrix ops are hand-written Rust

Cross-Platform Releases

Pre-built binaries are automatically built via GitHub Actions for:

Windows x86_64
Linux x86_64
macOS aarch64

Create a git tag (e.g. v0.1.0) and push to trigger a release build.

File Structure

src/
  main.rs      — CLI entry point, argument parsing
  config.rs    — Model architecture configuration
  gemv.rs      — GEMV/GEMM kernels (F16 + Q4), hybrid forward pass, prefill
  simd.rs      — AVX-512 SIMD kernels (Q4/F16 GEMV, DeltaNet, fused MLP)
  generate.rs  — Text generation loop (text, image, video modes)
  tokenizer.rs — Tokenizer wrapper and chat templates
  vision.rs    — Vision encoder (ViT + merger), image/video loading
  save.rs      — Binary model format (.qor08b) save/load
  convert.rs   — One-time safetensors → .qor08b converter
  system.rs    — System awareness (RAM detection, smart limits)
  lib.rs       — Module exports

Model Binary Format (.qor08b)

Custom binary format for fast loading:

Header:  "QR08" magic + version(u32) + format(u8: 0=F16, 1=Q4)
Config:  Architecture params (vocab, hidden, layers, heads, etc.)
Layers:  24 layers, each with type byte + layer-specific weights
Global:  Embedding + final norm + precomputed RoPE tables
Vision:  Conv3d patch embed + pos_embed + 12 ViT blocks + merger MLP

Loading is ~500ms for the Q4 model (~600 MB) via buffered sequential reads.

Performance

Tested on i5-11500 (6C/12T), 16GB RAM:

Task	Speed
Text decode	~3.9 tok/s
Text prefill	~13 tok/s (batched DeltaNet)
Model load	~500ms (Q4, 600 MB)
RAM usage	~791 MB

CPU-only by design — the 0.8B model is small enough that CPU inference is fast and efficient, making it ideal for mobile and edge deployment without GPU dependencies. Batched DeltaNet prefill processes all GEMM projections in parallel across tokens, with only the lightweight conv1d and recurrent state update running sequentially.

Comparison with QORA-4B

	QORA-0.8B	QORA-4B
Parameters	0.8B	4B
Model size (Q4)	600 MB	2.9 GB
Load time	~500ms	~30s
Decode speed	~3.9 tok/s	~3.3 tok/s (GPU), ~1.3 tok/s (CPU)
RAM usage	~791 MB	~3.5 GB
GPU support	No (not needed)	Yes (Vulkan)
Vision	12L ViT (768)	24L ViT (1024)
Best for	Mobile, edge, quick tasks	Desktop, complex reasoning

qoranet/QORA-LLM-0.8B

QORA-0.8B

License

What It Does

Architecture

DeltaNet Layers (18 of 24)

Full Attention Layers (6 of 24)

Vision Encoder

Smart System Awareness

Weight Format

AVX-512 SIMD Acceleration

Platform Support

Quick Start

CLI Flags

Sampling Defaults

Video Input

Building from Source

Dependencies

Cross-Platform Releases

File Structure

Model Binary Format (.qor08b)

Performance

Comparison with QORA-4B

qoranet/
QORA-LLM-0.8B