DEV.co
Open-Source LLM · stelterlab

Qwen3-Coder-30B-A3B-Instruct-AWQ

Qwen3-Coder-30B-A3B-Instruct is a 30.5B-parameter mixture-of-experts coding model with 3.3B active parameters, quantized to INT4 by stelterlab using llm-compressor. It supports 256K native context (extendable to 1M with Yarn), excels at agentic coding and tool calling, and is available under Apache 2.0. This is a community quantization of Qwen's official base model, marked experimental.

Source: HuggingFace — huggingface.co/stelterlab/Qwen3-Coder-30B-A3B-Instruct-AWQ
30.8B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
42.2k
Downloads (30d)

Key facts

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

FieldValue
Developerstelterlab
Parameters30.8B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads42.2k
Likes6
Last updated2025-08-02
Sourcestelterlab/Qwen3-Coder-30B-A3B-Instruct-AWQ

What Qwen3-Coder-30B-A3B-Instruct-AWQ is

Mixture-of-experts causal language model (48 layers, 32 Q-heads, 4 KV-heads, 128 experts with 8 activated). Native 262K context length. Quantized to INT4 GEMM using llm-compressor; original weights from Qwen AI. Instruct-tuned with support for function calling and tool use. Does not generate thinking blocks. Requires transformers>=4.51.0 for qwen3_moe support. Marked experimental by quantizing party.

Quickstart

Run Qwen3-Coder-30B-A3B-Instruct-AWQ locally

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

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="stelterlab/Qwen3-Coder-30B-A3B-Instruct-AWQ")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

Agentic code generation and tool-use workflows

Model is explicitly optimized for tool-calling and function execution; well-suited for pipelines that require the LLM to invoke external APIs, linters, or testing frameworks autonomously.

Long-context code analysis and repository understanding

256K native context (1M extendable) allows processing of entire codebases, design docs, and multi-file contexts without truncation—valuable for refactoring, audits, and cross-file dependency analysis.

Resource-constrained environments (quantized deployment)

INT4 quantization reduces memory footprint while retaining coding performance; suitable for on-premise, edge, or cost-sensitive cloud deployments where BF16/FP32 models are infeasible.

Running & fine-tuning it

ESTIMATE (verify against your infra): INT4 quantization likely requires 16–20GB VRAM for inference at native 256K context; BF16 full model requires ~60GB VRAM. Inference on single A100 (80GB) or dual A6000 (48GB each) plausible. OOM mitigation recommended (reduce context to 32K if needed, per model card). Requires transformers>=4.51.0 and cuda/compatible accelerator for llm-compressor quantization format.

Not documented in provided card. Feasibility of LoRA/QLoRA on quantized INT4 weights is Unknown—requires testing. Fine-tuning the original unquantized Qwen3-Coder base model is likely more straightforward; quantization may introduce complications. Consult llm-compressor and vLLM project docs for QLoRA compatibility on GEMM INT4 models.

When to avoid it — and what to weigh

  • Production deployments without quantization validation — This is a community quantization marked experimental; Qwen has not published official INT4 benchmarks for this version. Requires in-house evaluation on your specific coding tasks before production use.
  • Need for reasoning or step-by-step thinking — Model explicitly does not generate <think></think> blocks and operates in non-thinking mode. If your workflow relies on chain-of-thought reasoning, consider alternatives or instruct-only models with explicit reasoning support.
  • Single-GPU deployment with <24GB VRAM — 30.5B parameters in BF16/FP32 requires ~60GB; even quantized INT4, estimate ~16–20GB VRAM for full context. Multi-GPU or quantization required for smaller single-GPU setups.
  • Guaranteed SLA and enterprise support — This is a community quantization by stelterlab, not an official release. No guaranteed support, uptime SLA, or vendor backing. Qwen AI support may not cover this quantized variant.

License & commercial use

Apache 2.0 (OSI-approved permissive license). Model itself is unencumbered; no gating. Derivative quantization by stelterlab also released under Apache 2.0.

Apache 2.0 permits commercial use, modification, and distribution with minimal restrictions (attribution required). However, this is a community quantization (experimental) by stelterlab, not an official Qwen release. For production commercial deployments, confirm with Qwen AI that using third-party quantizations does not void any indirect support or indemnification they may offer on the base model. No SLA or commercial support stated for this specific variant.

DEV.co evaluation signals

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

SignalAssessment
MaintenanceActive
DocumentationAdequate
License clarityClear
Deployment complexityModerate
DEV.co fitGood
Assessment confidenceMedium
Security considerations

Standard LLM attack surface (prompt injection, jailbreaking, model extraction via API). No security audit or adversarial robustness claims in card. Community quantization (stelterlab) introduces supply-chain risk: verify weight integrity and review llm-compressor source. Quantization may alter model behavior subtly; test on sensitive tasks before production. No formal vulnerability disclosure process stated.

Alternatives to consider

Qwen3-Coder-30B-A3B-Instruct (official unquantized)

Same base model; avoids experimental quantization; official Qwen support; trade-off is 3x VRAM (~60GB unquantized BF16). Use if serving capacity allows.

DeepSeek-Coder-33B-Instruct or smaller code LLMs

Established code-focused models with published quantization benchmarks; lower experimental risk. Smaller alternatives (7B–13B) fit more constrained hardware.

Claude API or GPT-4 (hosted)

If no on-premise requirement: managed service, no quantization concerns, guaranteed SLA, and proven agentic capabilities. Higher latency and cost per token.

Software development agency

Ship Qwen3-Coder-30B-A3B-Instruct-AWQ with senior software developers

This model requires 16–20GB VRAM (INT4) for optimal inference. Validate quantization quality on your internal coding tasks before production. Contact Devco to review integration with your vector DB, inference framework, and agentic pipeline.

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-30B-A3B-Instruct-AWQ FAQ

Can I use this quantized model commercially?
Apache 2.0 permits commercial use. However, this is a community quantization marked experimental. Test thoroughly before production, and confirm with Qwen AI that using third-party quantizations aligns with any indirect commercial expectations or support model you may have on the base model.
What hardware do I need to run this locally?
Estimate 16–20GB VRAM for INT4 inference at native 256K context (A100, A6000, or equivalent). If OOM occurs, reduce context to 32K per model card guidance. Original BF16 model needs ~60GB.
Is this the official Qwen3-Coder quantization?
No. This is a community quantization by stelterlab using llm-compressor. The base model is official (Qwen AI), but the INT4 quantization is experimental and community-maintained. Qwen AI may not provide direct support for this variant.
Does it support reasoning or chain-of-thought prompting?
No. The model operates in non-thinking mode and does not generate <think></think> blocks. It is optimized for direct agentic coding (tool calling, function invocation), not step-by-step reasoning.

Custom software development services

DEV.co is a software development agency delivering custom software development services to companies building on open source. Our software developers and web developers design, integrate, and ship production systems — spanning web development, APIs, AI, data, and cloud. If Qwen3-Coder-30B-A3B-Instruct-AWQ is part of your open-source llms roadmap, our team can implement, customize, migrate, and maintain it.

Ready to Deploy? Assess Your Infrastructure.

This model requires 16–20GB VRAM (INT4) for optimal inference. Validate quantization quality on your internal coding tasks before production. Contact Devco to review integration with your vector DB, inference framework, and agentic pipeline.