LG AI Research · Chat and reasoning · 749B parameters · 37B active · apache-2.0
K-EXAONE 2.0 750B A37B
Needs 899 GB at FP8 / INT8. The smallest real configuration today is 4× MI325X at Cyfuture AI, $12.29 per hour, about $8,972 a month at 730 hours.
Parameters749B37B active per token
Context256k tokensmaximum published
Checkpoint1,499 GBBF16, as published
Licenseapache-2.0check before commercial use
ReleasedJul 31, 2026documented date
Smallest configuration · FP8 / INT8
4× MI325X1024 GB VRAM · 88% used
- Per hour
- $12.29
- Per month
- $8,972
- Cheapest at
Cyfuture AI
Memory · 3 precisions · 5 context lengthsHow much VRAM it needs
Weights plus runtime overhead, then the KV cache that grows with the context you serve.
| Precision | Weights | Weights + headroom | 4k context | 32k context | 128k context | 256k context |
|---|---|---|---|---|---|---|
| BF16published checkpoint | 1,499 GB | 1,799 GB | 1,800 GB | 1,809 GB | 1,841 GB | 1,883 GB |
| FP8 / INT8estimated | 749 GB | 899 GB | 901 GB | 910 GB | 941 GB | 983 GB |
| INT4 / MXFP4estimated | 449 GB | 540 GB | 541 GB | 550 GB | 582 GB | 624 GB |
| KV cacheone sequence, FP16 cache, included above | — | — | 2 GB | 11 GB | 42 GB | 84 GB |
Cross-checked on Hugging Face Oct 28, 2026: 749.4B parameters in the safetensors headers, 1498.7 GB of weight files, attention layout from config.json (78 layers, 8 KV heads × 128, 58 sliding-window layers). Method: weights as published at the native precision, verified on Hugging Face 2026-10-28; other precisions estimated at 1 (8-bit) and 0.6 (4-bit) bytes per parameter; × 1.2 runtime headroom; KV cache = 4096 bytes per token per layer × 20 full-attention layers + 58 sliding-window layers capped at None tokens (FP16 cache, one sequence).config.json ↗
Smallest configuration per GPU · 110 GPUs · cheapest eligible on-demand offer per GPU count, every provider on fileWhat it costs to run, GPU by GPU
The fewest GPUs of one type whose total memory fits the model at short context, priced from the cheapest eligible on-demand offer for exactly that count (45-day window, newest reading per provider), with the next providers beside it. Monthly = 730 hours.
| GPU | Smallest config | VRAM used | $ / hour | $ / month | Cheapest at |
|---|---|---|---|---|---|
| MI325XBest fit | 4× 256 GB1024 GB total | 88% | $12.29 | $8,972 | Cyfuture AInot scored · read Nov 2 · receipt |
| MI300X | 8× 192 GB1536 GB total | 59% | $13.68 | $9,986 |  TensorWaveTrust 64 · read Nov 2 · receipt · also Cyfuture AI $22.38, AtmosCompute $26.77 |
| H200 | 8× 141 GB1128 GB total | 80% | $21.04 | $15,359 |  Airon AInot scored · read Nov 2 · receipt · also Geodd $21.76, Lium $23.92 |
| B300 | 4× 288 GB1152 GB total | 78% | $23.29 | $17,004 |  UpCloudTrust 75 · read Nov 2 · receipt · also FPT AI Factory $27.96, Verda $45.40 |
| MI355X | 8× 288 GB2304 GB total | 39% | $23.60 | $17,228 | TensorWaveTrust 64 · read Nov 2 · receipt |
| H200 NVL | 8× 141 GB1128 GB total | 80% | $26.09 | $19,042 |  IONOSnot scored · read Nov 2 · receipt · also Geodd $27.12, Vast.ai $39.47 |
| B200 | 8× 180 GB1440 GB total | 62% | $32.91 | $24,026 | UpCloudTrust 75 · read Nov 2 · receipt · also Airon AI $33.04, Deep Infra $39.12 |
| B300 CC | 4× 288 GB1152 GB total | 78% | $46.31 | $33,806 |  VerdaTrust 81 · read Nov 2 · receipt |
| GB300 | 4× 288 GB1152 GB total | 78% | $52.21 | $38,113 | VerdaTrust 81 · read Nov 2 · receipt |
| B200 CC | 8× 180 GB1440 GB total | 62% | $72.37 | $52,830 | VerdaTrust 81 · read Nov 2 · receipt |
Throughput estimates · Oct 4, 2026Tokens per second, and what a million costs
Chat and reasoning. Speed estimated from memory bandwidth and model size, never measured. The cost per 1M output tokens is an estimate at the prices read on Oct 4, 2026; it is not recomputed here.
| GPU | Cards | Precision | tok/s per user | per 1M output tokens · estimate |
|---|---|---|---|---|
| MI300X | 8 | FP8 | 26.8 | $4.57 |
| H200 | 8 | FP8 | 34.9 | $5.56 |
| B200 | 8 | FP8 | 46.8 | $6.08 |
cards, precision and speed estimated from memory bandwidth and model size; cost recomputed by the site from verified prices · cheapest per output at the prices of Oct 4, 2026: MI300XWhy estimates, and their limits →
K-EXAONE 2.0 750B A37B · related