VRAM Requirements for Coding LLMs: What Every Tier Actually Runs

作者 Jakub Rusinowski · 最后更新 2026-08-04 · 硬件数据由我们的 VRAM 引擎

8 GB VRAM runs Qwen3-Coder 8B plus an autocomplete model — a real entry point. 16 GB runs Devstral-2 22B, the first genuinely agent-capable tier. 24 GB runs Qwen 3.6 27B or Qwen 2.5 Coder 32B — the sweet spot. Above that, ~49 GB (2×24 GB or a 96 GB+ Mac) unlocks Qwen3-Coder 80B-A3B. Coding agents need context headroom on top of weights, so budget one notch looser than a chat-only setup.

Coding is the hardest workload to size VRAM for, because three things compete for the same memory: model weights, the KV cache (which grows with context length — and agent tools routinely push 16–32K-token contexts), and any second model you run for autocomplete. Generic "7B needs 5 GB" tables ignore the last two, which is how people end up with an agent that crashes the moment Cline loads a big file.

Every figure below is the Q4_K_M requirement computed by the same engine as our compatibility checker — weights + KV cache + overhead, not a copied spec-sheet number. For any specific GPU + model combination, including quant and context options, the checker gives the exact verdict.

8 GB VRAM — the honest entry point

RTX 3060/4060-class cards and 16 GB Macs. One current-gen small coder, or coder + autocomplete if you keep contexts modest.

12–16 GB VRAM — the first agentic tier

RTX 3060 12GB, RTX 4060 Ti 16GB / 5060 Ti 16GB, RX 7800 XT-class, 24 GB Macs. Dedicated coding agents become reliable here.

24 GB VRAM — the sweet spot

RTX 3090/4090, RX 7900 XTX, 32 GB Macs. The tier this hub recommends aiming for if you’re buying hardware for coding.

48 GB+ — workstation and Mac territory

2×24 GB GPUs, a 48 GB card, or 64–128 GB Apple unified memory (of which ~75% is usable for models).

The two VRAM rules specific to coding

Rule 1 — reserve context headroom. KV cache grows linearly with context length. A model that "fits" at 8K context can overflow at 32K — and agent tools like Cline live at 16–32K. If your card fits a model with less than ~15% VRAM to spare, either lower the quant one step or choose the next model down. (This is exactly what the checker models when you change the context slider.)

Rule 2 — MoE changes the math in your favor. Mixture-of-experts models need all weights resident but only stream active experts per token — so Qwen3-Coder 80B-A3B is *faster* than a dense 27B once loaded, and with llama.cpp expert offload it degrades gracefully onto system RAM instead of refusing to load. If you're between tiers, a small-active MoE (35B-A3B, 80B-A3B) is usually the better stretch than a bigger dense model. More in our MoE models guide.

On Apple Silicon, count ~75% of unified memory as usable for models — a 48 GB Mac is a "36 GB card" for planning purposes.

Upgrading a tier?

If this page moved your target from 8 GB to 16 or 24, these are the cards each tier is priced around — the 4060 Ti 16GB as the value pick, the 3090/4090 for the sweet spot.

没有硬件？先租一块 GPU

Or skip the purchase: rent the exact VRAM tier for an hour and watch actual memory usage with your real workload before spending anything.

• RunPod — RTX 4090 ≈ $0.3–0.7/hr, A100/H100 by the hour; serverless per-second billing available
• Vast.ai — Marketplace of hosted GPUs — often the cheapest per hour (interruptible options)
• Lambda — On-demand A100/H100/B200 instances and clusters, per-hour billing

完整列表见 云 AI 目录.

常见问题

继续阅读

• Everything a 16 GB card runs — full VRAM tier page
• Can-I-Run index — check a specific GPU + model pair
• CUDA out of memory — diagnosis and fixes
• Best Local Coding LLMs in 2026, by VRAM Tier
• Best GPU for Local AI Coding in 2026: Five Picks That Make Sense
• Context Engineering for Local Coding Agents: Budgeting a Window You Pay For
• Why Your Local Coding Agent Feels Slow — and the Six Fixes That Work
• 本地 AI 编程助手 — 完整专题