DEV.co
Open-Source LLM · cyankiwi

Qwen3-Coder-Next-AWQ-8bit

Qwen3-Coder-Next is an open-weight 80B-parameter language model optimized for code generation and agentic tasks. It uses a mixture-of-experts (MoE) architecture with only 3B active parameters, enabling efficient inference while maintaining performance comparable to much larger models. The model supports a 256k context window, tool calling, and integrates with popular coding environments and IDEs. It is available as an AWQ 8-bit quantized variant and licensed under Apache 2.0.

Source: HuggingFace — huggingface.co/cyankiwi/Qwen3-Coder-Next-AWQ-8bit
24.1B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
108.3k
Downloads (30d)

Key facts

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

FieldValue
Developercyankiwi
Parameters24.1B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads108.3k
Likes6
Last updated2026-03-27
Sourcecyankiwi/Qwen3-Coder-Next-AWQ-8bit

What Qwen3-Coder-Next-AWQ-8bit is

Qwen3-Coder-Next is a causal language model with 80B total parameters (3B activated). Architecture: 48 layers with hybrid layout combining Gated DeltaNet, Gated Attention, and MoE (512 experts, 10 activated per token, 1 shared). Context length: 262,144 tokens natively. This HuggingFace variant (cyankiwi/Qwen3-Coder-Next-AWQ-8bit) is an 8-bit quantized version using AWQ compression, available in safetensors format. Supports transformers library, compatible with SGLang ≥v0.5.8 and vLLM ≥0.15.0 for serving. Recommended sampling: temperature=1.0, top_p=0.95, top_k=40.

Quickstart

Run Qwen3-Coder-Next-AWQ-8bit 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-Coder-Next-AWQ-8bit")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

Local Coding Agent Development

Deploy as a self-hosted agent in development environments. The 3B active parameters and quantization enable efficient local inference on modest hardware, with tool-calling capabilities suited for autonomous code editing, refactoring, and bug fixing workflows.

Long-Context Code Repository Analysis

Leverage 256k context to process entire codebases, documentation, and test suites in a single inference pass. Useful for codebase understanding, cross-file refactoring, and comprehensive code review tasks without context truncation.

Cost-Effective API Endpoint Deployment

Run as an OpenAI-compatible API (via SGLang or vLLM) on shared infrastructure. The MoE efficiency and quantization minimize VRAM and latency, reducing per-token serving costs compared to dense larger models.

Running & fine-tuning it

ESTIMATE (requires verification): 8-bit quantized 80B model ≈ 160–200 GB in float32 equivalent; at 8-bit ≈ 80–100 GB. Full 256k context on 2–4 GPUs with 80GB VRAM each recommended (e.g., A100, H100). For single-GPU inference, reduce context to 32k–64k and expect OOM risks. Serving with SGLang/vLLM on 4 GPUs (tensor parallel size 2) documented in model card.

Not stated in model card. Qwen3-Coder-Next is presented as a pre-trained and post-trained base model. No explicit mention of LoRA/QLoRA support, instruction tuning, or fine-tuning recipes. Approach as a production base model; fine-tuning feasibility Unknown and requires vendor documentation or community testing.

When to avoid it — and what to weigh

  • Requiring Reasoning-Mode Thinking Blocks — This model explicitly does not support thinking mode or <think></think> block generation. If your workflow depends on chain-of-thought intermediate reasoning output, this is not suitable.
  • Extreme Multi-Language or Domain Generalization Needed — The model is optimized specifically for coding. Non-code tasks (creative writing, translation, domain-specific expertise outside software engineering) may underperform compared to general-purpose models.
  • Real-Time Single-GPU Consumer Hardware — While quantized, the 80B total parameter model and 256k context default require substantial VRAM. OOM errors are documented; running at full context on single consumer GPUs is not recommended without context reduction.
  • Strict Latency SLAs Without Tuning — MoE and large context lengths introduce inference complexity. Without careful deployment tuning (tensor parallelism, context length reduction), latency may not meet sub-100ms SLA targets.

License & commercial use

Licensed under Apache License 2.0 (apache-2.0), an OSI-approved permissive open-source license. Full text available at https://opensource.org/licenses/Apache-2.0. No gatekeeping (gated=false).

Apache 2.0 is a permissive OSI license that permits commercial use, modification, and distribution with minimal restrictions (requires copyright notice and license copy). No additional commercial restrictions are stated. However, verify compliance with any downstream dependencies (transformers, SGLang, vLLM) and your internal legal/compliance policies. No explicit warranty or support clause in the license; use is 'as-is'.

DEV.co evaluation signals

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

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

Model card does not discuss security posture, data provenance, safety filtering, or known vulnerabilities. As an open-weight code model: (1) Review training data sources and potential code-based exploits or backdoors if deployment is critical. (2) Quantization (8-bit) does not inherently mitigate model-based risks. (3) Tool-calling capability enables code execution; secure sandboxing of invoked tools is the deployer's responsibility. (4) No mention of adversarial robustness or alignment testing. Conduct your own security review before production use in high-stakes environments.

Alternatives to consider

DeepSeek-Coder-V2 (or similar dense code models)

Dense alternatives may offer better single-GPU inference and simpler deployment if MoE scheduling overhead or multi-GPU requirements are blockers. Trade-off: fewer parameters activated per token, less efficient, higher VRAM at parity.

Code Llama (or Llama 3.1-based code finetuned models)

Established, widely-supported code models with stronger ecosystem support and more documentation. Trade-off: smaller context (typically 4–8k), potentially less recent training, may have different licensing (Llama 2 Community License is not Apache 2.0).

Claude 3.5 Sonnet or GPT-4 Turbo (proprietary API)

Closed-source alternatives eliminate deployment complexity and security concerns around model integrity. Trade-off: no local deployment, per-token costs, data privacy/lock-in, no ability to fine-tune or audit model behavior.

Software development agency

Ship Qwen3-Coder-Next-AWQ-8bit with senior software developers

Qwen3-Coder-Next offers efficient, open-source code generation with agentic capabilities and long context for repository analysis. Start with SGLang or vLLM on GPU infrastructure, or test locally with Ollama/LMStudio. Verify hardware requirements and fine-tuning support for your use case before production.

Talk to DEV.co

Related open-source tools

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

Related on DEV.co

Explore the category and the services that help you build with it.

Qwen3-Coder-Next-AWQ-8bit FAQ

Can I use Qwen3-Coder-Next commercially?
Yes. The Apache 2.0 license permits commercial use, including in products and services, provided you include a copy of the license and copyright notice. No explicit commercial restrictions are stated. Verify compatibility with your internal policy and any downstream dependencies (transformers, SGLang, vLLM).
What hardware do I need to run this model locally?
Estimate: 80–100 GB VRAM for the 8-bit quantized model. For full 256k context, use 2–4 GPUs with 80GB VRAM each (A100, H100). For single-GPU inference, reduce context to 32k–64k to avoid OOM. The model card documents OOM issues and recommends context reduction if server startup fails.
Does this model support thinking/reasoning blocks like o1 or Claude 3.5?
No. The model card explicitly states it does not generate <think></think> blocks and operates only in non-thinking mode. If your workflow requires intermediate reasoning output, this model is not suitable.
How do I deploy this as an API?
Use SGLang (≥v0.5.8) or vLLM (≥0.15.0). Both provide OpenAI-compatible endpoints. Example: `python -m sglang.launch_server --model Qwen/Qwen3-Coder-Next --port 30000 --tp-size 2` for tensor-parallel serving on 4 GPUs. Reduce context length if startup fails.

Software development & web development with DEV.co

Need help beyond evaluating Qwen3-Coder-Next-AWQ-8bit? DEV.co is a software development agency offering software development services and web development for teams of every size. Our software developers and web developers build custom software, web applications, APIs, and open-source llms integrations — and maintain them long-term.

Ready to Deploy a Local Code Agent?

Qwen3-Coder-Next offers efficient, open-source code generation with agentic capabilities and long context for repository analysis. Start with SGLang or vLLM on GPU infrastructure, or test locally with Ollama/LMStudio. Verify hardware requirements and fine-tuning support for your use case before production.