DEV.co
Open-Source LLM · Qwen

Qwen3-Coder-480B-A35B-Instruct-FP8

Qwen3-Coder-480B-A35B-Instruct-FP8 is a 480-billion-parameter mixture-of-experts (MoE) code generation model from Alibaba's Qwen team, with 35B parameters active per inference. It targets agentic coding tasks, repository-scale understanding via 256K native context (extendable to 1M), and tool-calling workflows. The FP8 quantization variant reduces memory footprint while maintaining inference compatibility with transformers, vLLM, and sglang. Licensed under Apache 2.0, ungated, and actively maintained as of August 2025.

Source: HuggingFace — huggingface.co/Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8
480.2B
Parameters
apache-2.0
License (OSI-approved)
Unknown
Context (tokens)
529.6k
Downloads (30d)

Key facts

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

FieldValue
DeveloperQwen
Parameters480.2B
Context windowUnknown
Licenseapache-2.0 — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads529.6k
Likes159
Last updated2025-08-21
SourceQwen/Qwen3-Coder-480B-A35B-Instruct-FP8

What Qwen3-Coder-480B-A35B-Instruct-FP8 is

Qwen3-Coder-480B-A35B uses a MoE architecture (160 experts, 8 activated per token) with 62 transformer layers, GQA attention (96 Q-heads, 8 KV-heads), and 262,144-token native context. The FP8-quantized checkpoint employs fine-grained block-wise quantization (block size 128). Best practices include temperature=0.7, top_p=0.8, repetition_penalty=1.05, and max_new_tokens≤65,536. Known limitation: distributed inference in transformers requires CUDA_LAUNCH_BLOCKING=1. No thinking-mode output. Compatible with OpenAI-API-compliant endpoints.

Quickstart

Run Qwen3-Coder-480B-A35B-Instruct-FP8 locally

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

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="Qwen/Qwen3-Coder-480B-A35B-Instruct-FP8")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

Multi-file Repository Analysis & Refactoring

The 256K native context and extended 1M-token support via Yarn enable analysis of large codebases without chunking. Ideal for cross-file dependency tracking, bulk refactoring, and architectural recommendations.

Agent-Driven Development Workflows

Specialized function-call format and tool-calling capabilities designed for agentic frameworks (Qwen Code, CLINE). Deploy as a backend for autonomous coding agents that interpret tool results and iterate.

Enterprise Code-Generation Infrastructure

Apache 2.0 license, ungated model, and FP8 quantization support enable self-hosted, cost-efficient private LLM deployments. Suitable for organizations requiring IP protection and regulatory compliance without external APIs.

Running & fine-tuning it

ESTIMATE: FP8 quantization (~240 GB weights + KV cache) typically requires 2–4× H100 80GB or A100 80GB GPUs for batch=1 at 256K context. Bfloat16 equivalent (480 GB weights) requires 6–8 such GPUs. Context reduction to 32K can reduce VRAM by ~50%. Exact numbers depend on batch size, sequence length, and inference framework optimizations. Verify with your target infrastructure.

Not explicitly stated in model card. LoRA/QLoRA feasibility on FP8 weights is Unknown; recommend checking community issues and Qwen documentation. Fine-tuning infrastructure typically requires same multi-GPU setup as inference. No instruction-tuning data or adapter configs provided.

When to avoid it — and what to weigh

  • Real-time, Ultra-low-latency Endpoints — 480B parameters (even with 35B active) incur significant first-token and per-token latency. Avoid latency-critical SLAs under ~500ms unless aggressively quantized or batched with large hardware.
  • Single-GPU or Consumer Hardware Deployments — Model card explicitly warns of OOM; FP8 variant still requires multi-GPU or high-VRAM systems (A100 80GB, H100, or equivalent). Not suitable for edge devices or laptop inference without severe context truncation.
  • Non-Code-Centric Use Cases — This is a specialized code model; performance on general-purpose NLP, reasoning, or non-English tasks is not documented. Use a general-purpose LLM (Qwen3-72B, Llama 3.1) for diverse domains.
  • Distributed Inference at Scale (transformers library only) — Known bug in transformers' fine-grained FP8 handling for multi-device inference; requires manual environment-variable workaround. Consider vLLM or sglang for production distributed setups.

License & commercial use

Apache License 2.0 (OSI-approved permissive license). Permits commercial use, modification, and distribution with attribution and liability disclaimer.

Apache 2.0 is a standard OSI-permissive license that explicitly permits commercial use. No gating or restrictions stated. You may deploy in production, behind APIs, or in commercial products provided you include a copy of the license and retain copyright notices. Verify with legal counsel for your jurisdiction, but license terms are clear and business-friendly.

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 generates code programmatically—review outputs for unintended side effects (e.g., shell commands, file operations). No explicit security audit or threat model documented. FP8 quantization introduces quantization-specific inference paths; verify with framework maintainers (vLLM/transformers) for any known numerical stability issues. Use in isolated environments if processing untrusted prompts.

Alternatives to consider

Claude 3.5 Sonnet (proprietary API)

Claimed comparable performance on agentic coding but requires API dependency and per-token cost. Use if you prioritize simplicity over IP control and self-hosting.

Qwen3-72B-Coder or smaller Qwen3-Coder variants

Significantly lower VRAM and latency if repository-scale context and 480B performance are not critical. Trade-off between capability and resource efficiency.

Llama 3.1 405B (Meta, gated)

Larger general-purpose model with broader domain coverage, but gated model access and no explicit agentic-coding optimization. Consider for multi-domain setups; avoid for code-only deployments.

Software development agency

Ship Qwen3-Coder-480B-A35B-Instruct-FP8 with senior software developers

Start with a small context (32K) on 2× H100 GPUs to validate performance and integration. Use FP8 quantization to reduce VRAM. Monitor GitHub for updates and check vLLM/sglang compatibility before scaling to production.

Talk to DEV.co

Related open-source tools

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

Qwen3-Coder-480B-A35B-Instruct-FP8 FAQ

Can I use this model commercially in a SaaS product?
Yes. Apache 2.0 explicitly permits commercial use, modification, and distribution. No gating or commercial licensing required. Include a copy of the license and retain copyright attribution in your legal notices.
What GPU hardware do I need to run this model?
Minimum: 2× H100 80GB (or 4× A100 80GB) for FP8 at batch=1 and moderate context (32K–128K). For full 256K context and batch inference, 4–6 H100s or equivalent is typical. Start with smaller quantized models or context windows to validate your setup.
What is the difference between the FP8 and non-FP8 variants?
FP8 reduces weights and activations to 8-bit floating-point, halving memory footprint (~240 GB vs. 480 GB) with minimal quality loss. All inference frameworks (transformers, vLLM, sglang) support FP8. Use FP8 unless you need maximum precision (rare for code generation).
Does this model support fine-tuning or LoRA?
Not documented in the model card. Feasibility Unknown; check Qwen GitHub issues and community reports. If needed, use base model or smaller Qwen3-Coder variants (72B) for LoRA experimentation.

Custom software development services

Need help beyond evaluating Qwen3-Coder-480B-A35B-Instruct-FP8? 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 Qwen3-Coder in Your Infrastructure?

Start with a small context (32K) on 2× H100 GPUs to validate performance and integration. Use FP8 quantization to reduce VRAM. Monitor GitHub for updates and check vLLM/sglang compatibility before scaling to production.