ariacompute/qwen3.5-0.8b_q8

Model Card for Qwen3.5-0.8B (Aria Quant Bundle, q8)

Model Details

Model Description

Qwen3.5-0.8B is a 0.8-billion-parameter, dense Transformer decoder-only language model developed by the Qwen team at Alibaba Cloud, featuring a 3:1 DeltaNet linear-attention to full-attention layer ratio for efficient long-context processing. Pre-trained on diverse public corpora and aligned via SFT + 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 — on qwen3-0.6B, logprob delta is just +0.00685 above FP16 and exact prefix fraction ties the best at 0.3854 — at 1.8× smaller than BF16 (1.6 GB → ~850 MB). 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, DeltaNet + full-attn hybrid)
• Language(s): English (primary), Chinese, and 20+ additional languages
• License: Apache 2.0
• Finetuned from model: Qwen/Qwen3.5-0.8B

Model Sources

• Original Repository: QwenLM/Qwen3.5
• Original Paper: Qwen3.5 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)	~1.0 GB	✅ Recommended
Mid-range smartphone (4–6 GB)	~1.0 GB	✅
Budget phone (2–3 GB)	~1.0 GB	✅ Functional
Raspberry Pi 5 / SBC (4–8 GB)	~1.0 GB	✅
IoT gateway (1–2 GB)	~1.0 GB	⚠️ Tight
Wearable (1 GB)	~1.0 GB	❌ Insufficient

Memory breakdown (q8, at 4K context): ~850 MB quantized model weights (mmap) + ~40 MB KV cache + ~50 MB runtime overhead + ~60 MB codebook overhead ≈ ~1.0 GB.

Note: KV cache is compact due to GQA depth (20 layers × 2 KV heads × head_dim=128), ~3.5× smaller than Qwen3-1.7B's cache.

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)

• Bundle size: ~850 MB (BF16 original: ~1.6 GB, ~1.8× compression)
• Generation quality: Awaiting gen_quant_eval audit. Method reference (qwen3-0.6b_q8+group): token overlap 0.6429, exact prefix fraction 0.3854 (ties best), logprob delta +0.00685 — near-lossless, statistically 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 Qwen3.5-0.8B. 8-bit per-group codebooks preserve generation fidelity essentially identical to FP16. For a smaller bundle with comparable prefix quality, use qwen3.5-0.8b_q326_channel (recommended generation quality). For the smallest bundle, use qwen3.5-0.8b_q4

Model Architecture

Qwen3.5-0.8B employs a dense Transformer decoder architecture with a 3:1 DeltaNet linear-attention to full-attention layer mix:

Parameter	Qwen3-0.6B	Qwen3.5-0.8B
Layers	28	20
Hidden size	1,024	1,024
FFN intermediate size	2,816	3,584
Attention heads (Query)	16	8
Attention heads (KV)	8 (GQA, group 2)	2 (GQA, group 4)
Head dimension	128	128
Activation	SiLU (SwiGLU-style gating)	SiLU (SwiGLU-style gating)
Position encoding	RoPE (θ = 1,000,000)	RoPE (θ = 1,000,000)
Normalization	RMSNorm (pre-norm)	RMSNorm (pre-norm)
Vocabulary size	151,936	248,320
Max context length	32,768	32,768
Input/output vocab	Tied	Tied

Design highlights (shared with Qwen3 family):

• GQA (Grouped Query Attention): 2 KV heads serving 8 query heads (group 4) — KV Cache memory is 4× smaller than full attention
• RoPE high base frequency (1M): Native support for up to 32K context length
• DeltaNet + full-attn hybrid: 3:1 linear-attention ratio enables efficient sub-quadratic long-context processing without sacrificing short-context quality
• Tied vocab: Input embedding and output projection weights are shared, saving ~255M parameters

Key architecture differences from Qwen3-0.6B:

• Mixed attention architecture (DeltaNet linear + full-attn hybrid vs pure full-attn in Qwen3) — the defining Qwen3.5 innovation for efficient long-context inference
• Larger vocab (248K vs 152K) — improves multilingual tokenization efficiency
• Wider FFN (3,584 vs 2,816) — increased per-layer capacity despite fewer total layers (20 vs 28)
• Fewer query heads (8 vs 16) and KV heads (2 vs 8, group 4) — more aggressive GQA reduces KV cache while maintaining head_dim=128

Bias, Risks, and Limitations

Limitations

• Reasoning depth: Multi-step logical reasoning (≥3 steps) is adequate for a sub-1B model, but falls far short of 7B+ frontier models. Verify outputs in high-stakes scenarios.
• Mathematics: GSM8K and MATH performance is modest, typical for sub-1B models. Use larger models for quantitative tasks requiring precision.
• Code generation: Capable of short function completions and single-line snippets; unreliable for multi-line synthesis or algorithmic problem solving.
• Factual knowledge: World knowledge is limited compared to larger models due to compact parameter count. Always verify factual claims against authoritative sources.
• Instruction following: Handles simple single-constraint prompts reliably. Complex, highly constrained instructions may degrade at longer contexts.
• Quantization drift: As the near-lossless 8-bit recipe, generation consistency is expected to be essentially indistinguishable from FP16. On qwen3-0.6b_q8+group, logprob delta is just +0.00685 and exact prefix fraction ties the best at 0.3854. Greedy sampling can still fork on ambiguous continuations.

Bias and Risks

• Bias: As with all large language models trained on web-scale data, Qwen3.5 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