DEV.co
Open-Source LLM · cyankiwi

Qwen3-Next-80B-A3B-Thinking-AWQ-4bit

Qwen3-Next-80B-A3B-Thinking is a 80-billion-parameter open-source reasoning model from Alibaba's Qwen team. It uses sparse mixture-of-experts and hybrid attention to activate only 3B parameters per token, offering efficient long-context reasoning (up to 262K native, extensible to 1M tokens). The quantized 4-bit AWQ version from cyankiwi is gated-free and Apache 2.0 licensed. Benchmarks show reasoning performance approaching or exceeding Gemini 2.5 Flash Thinking on complex tasks. It is designed for reasoning-heavy workloads and requires significant VRAM and inference framework support.

Source: HuggingFace — huggingface.co/cyankiwi/Qwen3-Next-80B-A3B-Thinking-AWQ-4bit
83.8B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
105.3k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
Developercyankiwi
Parameters83.8B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads105.3k
Likes23
Last updated2026-05-06
Sourcecyankiwi/Qwen3-Next-80B-A3B-Thinking-AWQ-4bit

What Qwen3-Next-80B-A3B-Thinking-AWQ-4bit is

The model combines Gated DeltaNet and Gated Attention in a hybrid layout (48 layers, 12 blocks of 3×DeltaNet+MoE + 1×Attention+MoE) with 512 experts (10 activated + 1 shared per token). Multi-token prediction during pretraining accelerates inference. This is the 4-bit AWQ quantized version, reducing memory footprint. Context length is 262,144 tokens natively with RoPE-based extensibility. Supports only thinking mode by default; outputs include internal reasoning steps. Requires transformers main branch for `qwen3_next` architecture support.

Quickstart

Run Qwen3-Next-80B-A3B-Thinking-AWQ-4bit locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="cyankiwi/Qwen3-Next-80B-A3B-Thinking-AWQ-4bit")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Complex Reasoning and Math Problems

Exceeds Qwen3-30B-Thinking and matches/exceeds Gemini-2.5-Flash-Thinking on AIME25, HMMT25, and SuperGPQA. Suitable for competitive math, theorem proving, and multi-step logical inference where thinking-mode depth is beneficial.

Long-Context Document Analysis

Native 262K token context with efficient MoE activation (3B/80B) enables cost-effective processing of full books, codebases, or legal contracts. Inference throughput 10× higher than Qwen3-32B for 32K+ context.

Self-Hosted Reasoning Pipelines

Apache 2.0 license and ungated access allow deployment in private inference clusters. 4-bit quantization reduces on-device VRAM; hybrid architecture suits resource-constrained multi-GPU setups with vLLM or SGLang.

Running & fine-tuning it

Estimate (4-bit AWQ): ~48 GB VRAM for single-GPU inference (at context length <32K). Multi-GPU (4× H100/A100): recommended for full context length (262K) with tensor parallelism. Activation: ~3B parameters per token. Typical inference: 30–50 tokens/sec per GPU depending on context length and framework optimization. Full precision (80B) would require ~160+ GB VRAM.

Unknown. Model card does not document LoRA, QLoRA, or full fine-tuning stability on this architecture. Hybrid attention + high-sparsity MoE may require custom gradient handling. GSPO (mentioned for post-training) is proprietary Qwen technique. Recommend testing on small dataset before production fine-tuning; consider reaching out to Qwen team for guidance.

When to avoid it — and what to weigh

  • Single-GPU or Consumer-Grade Hardware — Even 4-bit quantized, 80B requires ~48 GB VRAM (estimate). Not suitable for laptops or single RTX 4090. Requires enterprise GPU clusters or cloud inference.
  • Sub-Millisecond Latency Requirements — Thinking mode generates extended reasoning tokens; responses include internal state. Not optimized for ultra-low-latency chat or streaming frontends. Best for batch/offline workloads.
  • Lack of Specialized Inference Framework — Hybrid attention + MoE require vLLM ≥0.5.2 or SGLang ≥0.5.2 for efficient throughput. Hugging Face transformers alone will not leverage multi-token prediction or optimized kernels.
  • Non-English or Niche Domain Fine-Tuning Without Support — Model card does not detail LoRA/QLoRA stability for this architecture. Fine-tuning hybrid attention + MoE carries unknown stability risks; requires empirical validation.

License & commercial use

Apache License 2.0 (SPDX: apache-2.0). Permissive OSI-approved open-source license. Allows modification, commercial use, and redistribution with license and copyright attribution.

Apache 2.0 explicitly permits commercial use, including deployment in products and services, provided the license and attribution are retained. No gating or restrictions noted. However, verify compliance with any downstream inference provider's terms (e.g., cloud platform ToS). Quantization by cyankiwi is also Apache 2.0 licensed, so commercial deployment of this quantized variant is permitted.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationStrong
License clarityClear
Deployment complexityHigh
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Standard LLM considerations apply. Model has not undergone third-party adversarial testing documented here. Thinking mode may increase verbosity and reasoning transparency, which could expose internal logic in sensitive scenarios. Recommend: (1) validate outputs for proprietary reasoning patterns before deployment; (2) implement prompt injection filters; (3) monitor for reasoning-based jailbreaks (extended thinking could expose novel attack surfaces); (4) audit quantization method (AWQ) for floating-point precision degradation in security-critical tasks. No exploit or vulnerability details available.

Alternatives to consider

Qwen3-235B-A22B-Thinking-2507 (base model, unquantized)

Higher capacity (235B total, 22B active); superior benchmarks on most reasoning tasks (AIME25: 92.3%, MMLU-Pro: 84.4%). Trade-off: 2.8× more VRAM, lower inference throughput, higher latency. Best if unlimited resources.

Gemini 2.0 Flash Thinking API (proprietary, cloud-hosted)

Comparable reasoning performance on many benchmarks; managed inference, no deployment overhead. Trade-off: API-only, potential latency, cost per token, closed-source. Best if vendor lock-in acceptable.

DeepSeek-R1-Distill-Qwen-32B (open-source, smaller reasoning model)

Similar reasoning focus, 32B (32× smaller), better for resource-constrained setups. Trade-off: lower absolute performance (no AIME25/HMMT25 data), shorter native context (~32K). Best if inference speed critical.

Software development agency

Ship Qwen3-Next-80B-A3B-Thinking-AWQ-4bit with senior software developers

Evaluate Qwen3-Next-80B-A3B-Thinking in your infrastructure. Use SGLang or vLLM to optimize inference throughput. Verify VRAM availability and benchmark on your reasoning workloads. Apache 2.0 license enables immediate commercial deployment.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Qwen3-Next-80B-A3B-Thinking-AWQ-4bit FAQ

Can I use this model commercially in a production SaaS application?
Yes, Apache 2.0 permits commercial use. You must retain the license and attribution (cyankiwi as quantizer, Qwen as base model author). Verify your inference provider's terms (cloud platform, API gateway). No additional licensing fees or commercial restrictions apply.
How much VRAM do I need to run this 4-bit version?
Estimate ~48 GB for single-GPU inference at shorter context lengths (<32K tokens). For full 262K context with multiple concurrent requests, use tensor parallelism across 4× GPUs (H100/A100 class). Exact requirements depend on batch size, context length, and framework overhead. Test with your workload.
Why does the model only output thinking mode? How do I get regular chat responses?
Qwen3-Next-80B-A3B-Thinking is fine-tuned for reasoning-only tasks; the chat template automatically inserts `<think>` tag. The model outputs internal reasoning followed by a final answer. For general-purpose chat, use the base model (Qwen/Qwen3-Next-80B-A3B-Base) or a different model variant.
What inference framework should I choose: SGLang vs. vLLM?
Both support Qwen3-Next (≥0.5.2 for SGLang, ≥0.5.2 for vLLM). SGLang is recommended by the model card for reasoning parsing and MTP support. vLLM is more mature for general LLM serving. Start with SGLang if reasoning transparency is required; vLLM if you prioritize stability and broader integrations.

Custom software development services

Adopting Qwen3-Next-80B-A3B-Thinking-AWQ-4bit is usually one piece of a larger software development effort. As a software development agency, DEV.co provides software development services and web development expertise — pairing senior software developers and web developers with your team to design, build, and operate open-source llms software in production.

Ready to Deploy Advanced Reasoning?

Evaluate Qwen3-Next-80B-A3B-Thinking in your infrastructure. Use SGLang or vLLM to optimize inference throughput. Verify VRAM availability and benchmark on your reasoning workloads. Apache 2.0 license enables immediate commercial deployment.