ariacompute/qwen3-0.6b_q8

Model Card for Qwen3-0.6B (Aria Quant Bundle, q8)

Model Details

Model Description

Qwen3 0.6B is a 600-million-parameter, dense Transformer decoder-only language model developed by the Qwen team at Alibaba Cloud, pre-trained on diverse public corpora and aligned via supervised fine-tuning (SFT) and direct preference optimization (DPO). This distribution is provided by Aria Compute as an aria-quant-bundle — a uniform 8-bit quantized package using Hadamard rotation + Lloyd-Max codebook quantization with per-group codebooks (group size 32). It delivers near-lossless generation quality — logprob delta just +0.00685 above FP16, exact prefix fraction on par with the best quant recipes — at ~3.5× smaller than FP16. Optimized for CPU-only, on-device inference on mobile phones, edge devices, and single-board computers via the Aria Engine runtime. No GPU or cloud connection is required.

• Developed by: Qwen Team (Alibaba Cloud)
• Quantized and distributed by: Aria Compute
• Model type: Dense Transformer decoder-only (language)
• Language(s): English (primary), Chinese, and 20+ additional languages
• License: Apache 2.0
• Finetuned from model: Qwen/Qwen3-0.6B

Model Sources

• Original Repository: QwenLM/Qwen3
• Original Paper: Qwen3 Technical Report (pending)
• Aria Compute Dashboard: ariacompute.com/dashboard/models
• Aria Engine: ariacompute.com

Uses

Direct Use

This quantized bundle is intended for on-device, offline text-generation tasks on resource-constrained hardware, including:

• On-device chat and conversational assistants
• Real-time text completion and sentence prediction
• Structured tool calling / function calling for mobile and IoT APIs
• Lightweight text embeddings for on-device retrieval and classification
• Short-form summarization of notifications, messages, and local content

All inference runs locally on CPU. No data is sent to external servers.

Target Devices

Platform	Runtime Memory	Feasibility
High-end smartphone (8 GB)	~500 MB	✅ Recommended
Mid-range smartphone (4–6 GB)	~500 MB	✅
Budget phone (2–3 GB)	~500 MB	⚠️ Tight, but functional
Wearable (1 GB)	~500 MB	⚠️ Marginal
Raspberry Pi 5 / SBC (4–8 GB)	~500 MB	✅
IoT gateway (1–2 GB)	~500 MB	⚠️ Marginal

Memory breakdown (q8, at 4K context): ~340 MB quantized model weights (mmap) + 112 MB KV cache + 50 MB runtime overhead ≈ 500 MB.

Out-of-Scope Use

• Long-form creative writing (>2K tokens per generation)
• Mathematical theorem proving or complex multi-step reasoning
• Full program synthesis (reliable for short functions only)
• Multimodal input (this model is text-only)
• Real-time audio/speech processing (use Aria speech models)
• Safety-critical decision systems without human oversight
• Deployment in production when batch inference or GPU acceleration is required (this bundle targets CPU-only, single-prompt inference)

How to Get Started with the Model

Download from Aria Compute

Authenticated dashboard users can download the bundle via: https://ariacompute.com/dashboard/models

Quantization Recipe

This bundle uses the uniform 8-bit quantization recipe — the maximum-fidelity / near-lossless option in the Aria Compute lineup:

Component	Quantization Strategy	Details
Attention Q/K/V/O weights	8-bit	Per-group codebooks (group_size=32), Hadamard pre-processing
FFN up/gate/down weights	8-bit	Per-group codebooks (group_size=32), Hadamard pre-processing
RMSNorm weights	FP16	Preserved at full precision
Embedding table	FP16	Preserved at full precision (tied input/output vocab)

• Compression ratio: ~3.5× smaller than FP16
• Generation quality: Near-lossless — mean token overlap 0.6429, exact prefix fraction 0.3854 (ties q326_channel for best), mean logprob delta +0.00685 (indistinguishable from FP16)
• Calibration-free: Hadamard rotation + Lloyd-Max codebook, no task-specific calibration data required
• Maximum fidelity recipe: This is the highest-quality quantized option for Qwen3-0.6B in the Aria Compute family; for a smaller bundle with comparable prefix quality, consider q326_channel (recommended generation-quality recipe)

Model Architecture

Qwen3-0.6B employs a standard dense Transformer decoder architecture:

Parameter	Value
Layers	28
Hidden size	1,024
FFN intermediate size	2,816
Attention heads (Query)	16
Attention heads (KV)	8 (GQA, group size 2)
Head dimension	128
Activation	SiLU (SwiGLU-style gating)
Position encoding	RoPE (θ = 1,000,000)
Normalization	RMSNorm (pre-norm)
Vocabulary size	~152,000
Input/output vocab	Tied

Design highlights:

• GQA (Grouped Query Attention): 8 KV heads serving 16 query heads — halves KV Cache memory
• RoPE high base frequency (1M): Native support for 32K context length
• Dense FFN + SiLU gating: High inference efficiency, suitable for on-device use
• Tied vocab: Input embedding and output projection weights are shared, saving ~150M parameters

Bias, Risks, and Limitations

Limitations

• Reasoning depth: Multi-step logical reasoning (≥3 steps) degrades noticeably due to the 0.6B scale. Verify outputs in high-stakes scenarios.
• Mathematics: GSM8K and MATH performance is modest. Use larger models for quantitative tasks.
• Code generation: Capable of short function completions and one-liners; unreliable for multi-file synthesis or algorithmic problem solving.
• Factual knowledge: Limited world knowledge compared to larger models. Always verify factual claims against authoritative sources.
• Instruction following: May fail on complex, multi-constraint prompts. Keep instructions short and specific.
• Quantization drift: As a near-lossless 8-bit recipe, generation consistency is essentially indistinguishable from FP16 (logprob delta +0.00685). However, greedy sampling can still fork on ambiguous continuations (e.g., "The capital of France is" forks after " Paris" — same behavior as q326_channel on "Complete: The sky is"). The q8 and q326_channel recipes share the best mean exact prefix fraction at 0.3854.

Bias and Risks

• Bias: As with all large language models trained on web-scale data, Qwen3 may reflect societal biases present in its training corpus. Evaluate outputs before deployment in sensitive domains (hiring, healthcare, law).
• Toxicity: The base model has been safety-aligned with refusal training. However, no safety filter is exhaustive. Consider an additional output classifier in high-risk environments.
• Hallucination: May generate plausible-sounding but factually incorrect information. Implement output verification for critical applications.
• Dual-use risk: Text-generation capabilities could be misused for spam, disinformation, or impersonation. Deploy responsibly and in accordance with the Apache 2.0 license terms.

Recommendations

Users (both direct and downstream) should be made aware of the above risks, biases, limitations, and constraints of the model. We recommend:

• Adding a lightweight output safety classifier for user-facing deployments
• Verifying factual claims with external knowledge bases
• Not using the model for high-stakes decisions without human review