DEV.co
Open-Source LLM · GadflyII

Qwen3-Coder-Next-NVFP4

Qwen3-Coder-Next-NVFP4 is an NVIDIA FP4-quantized version of Qwen's 80B parameter mixture-of-experts code generation model. It compresses the model from ~149GB to 45GB (70% reduction) while maintaining most capability. The model uses NVFP4 quantization (4-bit weights and activations), supports 256K context length, and is optimized for vLLM serving. It is publicly available under Apache 2.0 license with no access restrictions.

Source: HuggingFace — huggingface.co/GadflyII/Qwen3-Coder-Next-NVFP4
Unknown
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
103.6k
Downloads (30d)

Key facts

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

FieldValue
DeveloperGadflyII
ParametersUnknown
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads103.6k
Likes45
Last updated2026-02-04
SourceGadflyII/Qwen3-Coder-Next-NVFP4

What Qwen3-Coder-Next-NVFP4 is

Base model: Qwen/Qwen3-Coder-Next (80B total, 3B activated per token). Architecture: Hybrid DeltaNet + Attention + MoE with 512 total experts, 10 activated + 1 shared, 48 layers, 262K token context. Quantization: NVFP4 (4-bit weights + 4-bit activations via llmcompressor 0.9.0.1). Key layers remain BF16 (lm_head, MoE gates, DeltaNet linear attention). Calibration: 20 samples from HuggingFaceH4/ultrachat_200k, sequence length 2048. MMLU-Pro performance: 51.27% (BF16: 52.90%, -1.63% delta). Tested to 128K context with FP8 KV cache. Format: compressed-tensors. Requires vLLM 0.16.0+, Transformers 5.0.0+. Developer notes vLLM fork for SM120 Blackwell P2P/TP=2 workaround.

Quickstart

Run Qwen3-Coder-Next-NVFP4 locally

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

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="GadflyII/Qwen3-Coder-Next-NVFP4")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

Large-scale code generation and refactoring

80B parameter model with MoE efficiency enables handling complex codebases and multi-file context. 256K context supports full project analysis. 70% size reduction lowers inference cost vs. full BF16.

Private/on-premise AI coding assistants

45GB footprint and Apache 2.0 license allow self-hosted deployment without licensing friction. Suitable for enterprise DevOps pipelines requiring code review, generation, and documentation at scale.

Research and fine-tuning on specialized code domains

Open-source, quantized base enables rapid iteration on domain-specific adaptations (embedded systems, domain languages, regulatory code). MoE structure allows selective expert fine-tuning.

Running & fine-tuning it

**Minimum for inference:** 2x NVIDIA A100 40GB or equivalent (dual H100, RTX 6000 Ada, or better). ESTIMATE: 45GB model weight + KV cache (context-dependent, up to ~30–50GB for batch size >1 at 64K+ context). FP8 KV cache reduces memory footprint. Multi-GPU recommended (TP=2) for throughput. **No CPU quantization support evident; NVIDIA GPU required.** Single A6000 (48GB) may work for small batches, low context; untested by maintainer beyond 128K.

Unknown from card. Base model architecture (MoE + DeltaNet) supports LoRA, but no guidance on quantized fine-tuning provided. NVFP4 quantization may complicate gradient computation; QLoRA feasibility requires testing. Recommend starting with full-precision LoRA on unquantized base (Qwen/Qwen3-Coder-Next) or quantization-aware fine-tuning using llmcompressor. Community support likely limited; escalate to Qwen team or vLLM project if custom adaptation needed.

When to avoid it — and what to weigh

  • Extremely latency-sensitive services (<10ms SLO) — NVFP4 quantization and MoE routing add overhead. vLLM serving with tensor parallelism (TP=2+) mitigates but introduces coordination latency. Confirm with benchmarks before SLA commitments.
  • Consumer/edge single-GPU deployments — 45GB model size requires modern high-VRAM hardware (2x40GB+ or equivalent). Not practical on consumer GPUs. Recommend smaller quantized models (7B–13B) for edge.
  • Production systems without vLLM/Transformers expertise — Requires vLLM 0.16.0+, Transformers 5.0.0+, and NVIDIA-specific quantization support. Non-trivial deployment; team must be comfortable with containerization, tensor parallelism, and debugging quantization artifacts.
  • Zero-shot accuracy-critical tasks requiring >52% MMLU-Pro — MMLU-Pro drops 1.63pp vs. BF16 (52.90% → 51.27%). If benchmark accuracy is contractual requirement, benchmark application-specific performance first.

License & commercial use

Apache License 2.0 (same as base model Qwen/Qwen3-Coder-Next). No gating, publicly available.

Apache 2.0 is a permissive OSI license explicitly allowing commercial use, modification, and distribution. No proprietary restrictions. However, assume responsibility for: (1) compliance with Qwen base model's terms (verify Qwen/Qwen3-Coder-Next licensing if different); (2) quantization methodology attribution (llmcompressor, vLLM project); (3) infrastructure licensing (NVIDIA software, cloud provider). No warranty or support implied. Recommend legal review for high-stakes commercial deployments.

DEV.co evaluation signals

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

SignalAssessment
MaintenanceModerate
DocumentationAdequate
License clarityClear
Deployment complexityHigh
DEV.co fitGood
Assessment confidenceMedium
Security considerations

Card does not disclose security posture, threat model, or vulnerability management. Standard OSS considerations apply: (1) Verify integrity of model weights and dependencies (vLLM, llmcompressor, Transformers); (2) Quantized weights may be harder to audit than source; (3) MoE routing and gating logic could introduce side-channel exposure under adversarial input—not addressed; (4) Self-hosted deployment requires securing inference endpoints, managing access logs, and monitoring token leakage; (5) No mention of adversarial robustness or jailbreak mitigations. Treat as development/staging artifact until formal security review.

Alternatives to consider

Qwen/Qwen3-Coder-Next (full BF16)

Baseline unquantized model; higher accuracy (52.90% MMLU-Pro), no quantization artifacts, simpler debugging. Trade-off: ~149GB, higher cost, longer latency.

deepseek-coder-33b-instruct (quantized, smaller)

33B parameter alternative; easier single-GPU deployment, lower latency, Apache 2.0 license. Trade-off: ~60% fewer parameters, may underperform on complex reasoning.

Meta Llama-3.1-70B-Instruct (fp8/int8 quantization)

70B general-purpose LLM with strong code ability; widely supported by vLLM/TGI; mature quantization ecosystem. Trade-off: not code-specialized; different license (requires review for commercial use—Llama license has restrictions).

Software development agency

Ship Qwen3-Coder-Next-NVFP4 with senior software developers

Qwen3-Coder-Next-NVFP4 enables enterprise-grade code generation on your infrastructure. Get started with vLLM, evaluate on your codebase, and contact our team for deployment architecture and licensing guidance.

Talk to DEV.co

Related open-source tools

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

Qwen3-Coder-Next-NVFP4 FAQ

Can I use this model commercially in a SaaS product?
Apache 2.0 permits commercial use. However, verify: (1) Qwen base model's license (assume same, but confirm); (2) infrastructure licenses (NVIDIA software may have restrictions); (3) quantization attribution (credit llmcompressor, vLLM contributors). No warranty or support from developer, so assume full liability. Recommend legal review before launch.
What GPU hardware do I need to run this?
Minimum: 2x NVIDIA A100 40GB (or equivalent: H100 80GB, RTX 6000 Ada). Single A6000 48GB may work for low-context, batch-size-1 inference. Estimate 45GB model + 20–50GB KV cache depending on batch and context. FP8 KV cache helps. CPU inference not viable; NVIDIA GPU required.
How much accuracy do I lose with NVFP4 quantization?
MMLU-Pro benchmark shows -1.63 percentage points (52.90% BF16 → 51.27% NVFP4). No other benchmarks provided. Your mileage will vary; test on your application-specific eval set before committing to quantized version.
Do I need to use vLLM, or can I use another serving framework?
vLLM 0.16.0+ is recommended and has native NVFP4 support. Transformers library can load and run the model but lacks serving optimizations. TGI/TensorRT-LLM feasibility is unknown; NVFP4 is relatively new. Plan on vLLM or building custom serving logic if using alternatives.

Software developers & web developers for hire

Need help beyond evaluating Qwen3-Coder-Next-NVFP4? 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.

Deploy a Private Code LLM Today

Qwen3-Coder-Next-NVFP4 enables enterprise-grade code generation on your infrastructure. Get started with vLLM, evaluate on your codebase, and contact our team for deployment architecture and licensing guidance.