LG AI Research · Chat and reasoning · 237B parameters · 23B active · other
K-EXAONE 236B A23B
Needs 285 GB at FP8 / INT8. The smallest real configuration today is 4× A100 80GB at Vast.ai, $2.40 per hour, about $1,754 a month at 730 hours.
Parameters237B23B active per token
Context256k tokensmaximum published
Checkpoint474 GBBF16, as published
Licenseothercheck before commercial use
Released—documented date
Smallest configuration · FP8 / INT8
4× A100 80GB320 GB VRAM · 89% used
- Per hour
- $2.40
- Per month
- $1,754
- Cheapest at
Vast.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 | 474 GB | 570 GB | 570 GB | 571 GB | 576 GB | 582 GB |
| FP8 / INT8estimated | 237 GB | 285 GB | 285 GB | 287 GB | 291 GB | 298 GB |
| INT4 / MXFP4estimated | 142 GB | 171 GB | 171 GB | 173 GB | 178 GB | 184 GB |
| KV cacheone sequence, FP16 cache, included above | — | — | 1 GB | 2 GB | 7 GB | 13 GB |
Cross-checked on Hugging Face Oct 28, 2026: 237.1B parameters in the safetensors headers, 474.2 GB of weight files, attention layout from config.json (48 layers, 8 KV heads × 128, 36 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 × 12 full-attention layers + 36 sliding-window layers capped at 128 tokens (FP16 cache, one sequence).config.json ↗
Smallest configuration per GPU · 109 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 |
|---|---|---|---|---|---|
| A100 80GBBest fit | 4× 80 GB320 GB total | 89% | $2.40 | $1,754 | Vast.aiTrust 78 · read Nov 2 · receipt · also Edgevana $4.60, Gcore $5.63 |
| A100 40GB | 8× 40 GB320 GB total | 89% | $3.20 | $2,338 | Vast.aiTrust 78 · read Nov 2 · receipt · also Denvr Cloud $9.20, Sesterce Cloud $12.32 |
| RTX A6000 | 8× 48 GB384 GB total | 74% | $3.36 | $2,453 |  LiumTrust 40 · read Nov 2 · receipt · also LeaderGPU $3.58, Verda $6.62 |
| L40S | 8× 48 GB384 GB total | 74% | $3.74 | $2,727 | Vast.aiTrust 78 · read Nov 2 · receipt · also Lium $6.00, GPU.ai $6.40 |
| RTX PRO 6000 | 4× 96 GB384 GB total | 74% | $3.84 | $2,805 | Vast.aiTrust 78 · read Nov 2 · receipt · also UpCloud $3.89, Hetzner $6.44 |
| MI350X | 1× 288 GB288 GB total | 99% tight | $3.99 | $2,913 |  OpenRelaynot scored · read Nov 2 · receipt |
| Gaudi 2 | 4× 96 GB384 GB total | 74% | $4.18 | $3,052 |  Cyfuture AInot scored · read Nov 2 · receipt |
| A40 | 8× 48 GB384 GB total | 74% | $4.61 | $3,364 |  LeaderGPUTrust 59 · read Nov 2 · receipt · also Exoscale $8.36 |
| A100X | 4× 80 GB320 GB total | 89% | $5.12 | $3,739 | Vast.aiTrust 78 · read Nov 2 · receipt |
| RTX 6000 Ada | 8× 48 GB384 GB total | 74% | $5.20 | $3,796 |  GPU.ainot scored · read Nov 2 · receipt · also io.net $5.46, QuantaCloud $6.21 |
| MI300X | 2× 192 GB384 GB total | 74% | $5.71 | $4,167 | Cyfuture AInot scored · read Nov 2 · receipt · also Hot Aisle $5.98, AtmosCompute $6.69 |
| MI355X | 1× 288 GB288 GB total | 99% tight | $5.99 | $4,373 |  Daytonanot scored · read Nov 2 · receipt · also OpenRelay $6.30 |
| MI325X | 2× 256 GB512 GB total | 56% | $6.21 | $4,536 | Cyfuture AInot scored · read Nov 2 · receipt |
| H100 | 4× 80 GB320 GB total | 89% | $6.84 | $4,991 |  GcoreTrust 80 · read Nov 2 · receipt · also GPU.ai $7.92, Edgevana $7.98 |
| B200 | 2× 180 GB360 GB total | 79% | $8.22 | $6,002 |  UpCloudTrust 75 · read Nov 2 · receipt · also Deep Infra $9.78, Elice AI Cloud $13.64 |
| L40 | 8× 48 GB384 GB total | 74% | $8.40 | $6,132 |  OblivusTrust 82 · read Nov 2 · receipt · also CoreWeave $10.00, AtmosCompute $12.21 |
| H200 NVL | 4× 141 GB564 GB total | 50% | $10.67 | $7,788 | Vast.aiTrust 78 · read Nov 2 · receipt · also IONOS $13.04, Geodd $13.56 |
| B300 | 2× 270 GB540 GB total | 53% | $11.65 | $8,506 | UpCloudTrust 75 · read Nov 2 · receipt · also FPT AI Factory $13.98, Vast.ai $17.00 |
| H200 | 4× 141 GB564 GB total | 50% | $14.76 | $10,775 |  Theta EdgeCloudnot scored · read Nov 2 · receipt · also Canopy Wave $16.00, Hugging Face $20.00 |
| H100 NVL | 4× 94 GB376 GB total | 76% | $17.86 | $13,038 |  AtmosComputenot scored · read Nov 2 · receipt · also EmpirioLabs AI $22.40, Open Telekom Cloud $25.90 |
| B200 CC | 2× 180 GB360 GB total | 79% | $18.09 | $13,206 |  VerdaTrust 81 · read Nov 2 · receipt |
| GB300 | 2× 279 GB558 GB total | 51% | $26.11 | $19,060 | VerdaTrust 81 · read Nov 2 · receipt |
| B300 CC | 4× 270 GB1080 GB total | 26% | $46.31 | $33,806 | VerdaTrust 81 · read Nov 2 · receipt |
| A800 80GB PCIe | 8× 80 GB640 GB total | 44% | $50.80 | $37,088 |  Volcano Enginenot scored · 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 | 2 | FP8 | 54.0 | $0.909 |
| B200 | 2 | FP8 | 94.8 | $1.20 |
| H200 | 2 | FP8 | 67.6 | $1.25 |
| H100 SXM | 4 | FP8 | 58.5 | $1.49 |
| RTX PRO 6000 | 4 | FP8 | 34.9 | $2.02 |
| H100 PCIe | 4 | FP8 | 34.0 | $2.16 |
| L40S | 8 | FP8 | 24.1 | $2.67 |
| A100 80GB | 8 | BF16 | 37.0 | $2.86 |
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 236B A23B · related