How much VRAM do you need for a 200B–300B model?
Short answer: about 164.9 GB at Q4_K_M, the usual quant for local use. Below: every quant, longer contexts, the cheapest hardware that fits, and every 200B–300B open model we track.
- Q4_K_M
- 164.9 GB 142.5 GB weights
- Q8_0
- 288.3 GB 249.8 GB weights
- FP16
- 541.7 GB 470.2 GB weights
No single consumer GPU we track fits it at Q4. Cheapest Mac or mini PC: Mac Studio M5 Ultra 256GB ($9,499).
Cheapest hardware for each quant
| Setting | Memory needed | Cheapest GPU | Cheapest Mac / mini PC |
|---|---|---|---|
| Q4_K_M, ~4k context | 164.9 GB | None we track | Mac Studio M5 Ultra 256GB 192 GB usable · $9,499 |
| Q8_0, ~4k context | 288.3 GB | None we track | Mac Studio M5 Ultra 512GB 384 GB usable · $22,299 |
| FP16, ~4k context | 541.7 GB | None we track | None we track |
| Q4_K_M, 32k context | 170.4 GB | None we track | Mac Studio M5 Ultra 256GB 192 GB usable · $9,499 |
No single consumer GPU we track holds a 200B–300B model at Q4. Your options: two or more GPUs (see the ready-made builds or the planner), a large unified-memory machine (see the best Mac for local LLMs), or a smaller quant with lower quality.
Memory by context length
Each extra token of context adds to the KV cache. Numbers for Qwen3-235B-A22B, from its published config; a Q8 KV cache halves the cache part.
| Context | Q4_K_M, FP16 cache | Q4_K_M, Q8 cache | Q8_0, FP16 cache |
|---|---|---|---|
| 4k tokens | 164.9 GB | 164.9 GB | 288.3 GB |
| 8k tokens | 165.7 GB | 164.9 GB | 289.0 GB |
| 16k tokens | 167.3 GB | 165.7 GB | 290.6 GB |
| 32k tokens | 170.4 GB | 167.3 GB | 293.8 GB |
| 64k tokens | 176.7 GB | 170.4 GB | 300.1 GB |
| 128k tokens | 189.3 GB | 176.7 GB | 312.7 GB |
Hardware that runs it
Bar: Q4 memory needed vs. usable memory. Speed is estimated decode at Q4 (how we estimate); MoE shows an upper bound only.FitsTight
| Hardware | Q4 need vs. usable | Q4 / Q8 / FP16 | Est. tok/s |
|---|---|---|---|
| Mac Studio M5 Ultra 256GBunified memory · $9,499 | 164.9 / 192.0 GB | Q4FitsQ8Won’t fitFP16Won’t fit | up to 72 |
| Mac Studio M5 Ultra 512GBunified memory · $22,299 | 164.9 / 384.0 GB | Q4FitsQ8FitsFP16Won’t fit | up to 72 |
200B–300B models we track
| Model | Q4_K_M | Q8_0 | FP16 |
|---|---|---|---|
| Step-3.7-Flash201.4B params · stepfun-ai · MoE | 141.4 GB | 247.0 GB | 464.1 GB |
| command-a-plus-05-2026-bf16218.8B params · CohereLabs · MoE | 153.5 GB | 268.3 GB | 504.1 GB |
| MiniMax-M2228.7B params · MiniMaxAI · MoE | 160.4 GB | 280.4 GB | 527.0 GB |
| MiniMax-M2.5228.7B params · MiniMaxAI · MoE | 160.4 GB | 280.4 GB | 527.0 GB |
| MiniMax-M2.7228.7B params · MiniMaxAI · MoE | 160.4 GB | 280.4 GB | 527.0 GB |
| MiniMax-M2.1228.7B params · MiniMaxAI · MoE | 160.4 GB | 280.4 GB | 527.0 GB |
| Qwen3-235B-A22B235.1B params · Qwen · MoE | 164.9 GB | 288.3 GB | 541.7 GB |
| Qwen3-235B-A22B-Instruct-2507235.1B params · Qwen · MoE | 164.9 GB | 288.3 GB | 541.7 GB |
| Qwen3-235B-A22B-Thinking-2507235.1B params · Qwen · MoE | 164.9 GB | 288.3 GB | 541.7 GB |
| Qwen3-VL-235B-A22B-Instruct235.7B params · Qwen · MoE | 165.3 GB | 289.0 GB | 543.0 GB |
| DeepSeek-V2.5235.7B params · deepseek-ai · MoE | 165.4 GB | 289.0 GB | 543.2 GB |
| DeepSeek-Coder-V2-Instruct235.7B params · deepseek-ai · MoE | 165.4 GB | 289.0 GB | 543.2 GB |
| DeepSeek-V2.5-1210235.7B params · deepseek-ai · MoE | 165.4 GB | 289.0 GB | 543.2 GB |
| DeepSeek-V2-Chat-0628235.7B params · deepseek-ai · MoE | 165.4 GB | 289.0 GB | 543.2 GB |
| DeepSeek-Coder-V2-Instruct-0724235.7B params · deepseek-ai · MoE | 165.4 GB | 289.0 GB | 543.2 GB |
| K-EXAONE-236B-A23B237.1B params · LGAI-EXAONE · MoE | 166.3 GB | 290.7 GB | 546.3 GB |
| Intern-S1240.7B params · internlm · MoE | 168.8 GB | 295.1 GB | 554.6 GB |
| Solar-Open2-250B250.3B params · upstage · MoE | 175.5 GB | 306.8 GB | 576.7 GB |
| Llama-3_1-Nemotron-Ultra-253B-v1253.4B params · nvidia | 177.7 GB | 310.6 GB | 583.8 GB |
| Inkling-Small266.0B params · thinkingmachines · MoE | 186.4 GB | 326.0 GB | 612.7 GB |
| ERNIE-5.5-Flash-Base286.2B params · baidu · MoE | 200.6 GB | 350.8 GB | 659.4 GB |
| MemT-1.0-Flash287.9B params · MemTensor · MoE | 201.8 GB | 352.8 GB | 663.3 GB |
| DeepSeek-V4-Flash290.9B params · deepseek-ai · MoE | 203.8 GB | 356.5 GB | 670.2 GB |
| Ling-plus292.5B params · inclusionAI · MoE | 205.0 GB | 358.4 GB | 673.8 GB |
| Hy3298.8B params · tencent · MoE | 209.3 GB | 366.1 GB | 688.2 GB |
| Hy3-preview298.8B params · tencent · MoE | 209.3 GB | 366.1 GB | 688.2 GB |
Quick answers
- How much VRAM do I need for a 200B–300B model?
- About 164.9 GB at Q4_K_M, 288.3 GB at Q8_0 and 541.7 GB at FP16, at a ~4k context. That is for Qwen3-235B-A22B (235.1B parameters): weights plus about 15% for the KV cache and 1 GB for the runtime.
- How much VRAM does a 200B–300B model need for 32k context?
- About 170.4 GB at Q4_K_M with an FP16 KV cache, or 167.3 GB with a Q8 cache, based on Qwen3-235B-A22B's published config. Models with different attention layouts need more or less.
- What is the cheapest GPU for a 200B–300B model?
- None of the single consumer GPUs we track hold it at Q4. Use several GPUs, a large unified-memory machine, or CPU offload (slow).
- Can a Mac run a 200B–300B model?
- Yes. The cheapest Mac or unified-memory mini PC we track that runs it at Q4 is the Mac Studio M5 Ultra 256GB (256 GB, about 192.0 GB usable, $9,499).
- Does a MoE model of the same size need less VRAM?
- No. A mixture-of-experts model still has to hold all of its weights in memory, so it needs about the same VRAM as a dense model with the same total parameters. It runs faster, because only a few experts are read per token.
Other sizes: 1B–2B · 2B–3B · 4B–6B · 7B–8B · 12B–14B · 16B–21B · 24B–27B · 30B–32B · 40B–50B · 70B · 120B · 300B–480B · 500B–800B · 1T+ · All sizes · Ollama VRAM by tag · GGUF sizes by quant