GB300 Model Testing Summary
Four models have actually run on this one DGX Station GB300. They are not peers: different engines, different quants, different speculation, different whether the weights fit in HBM. This page is the comparison table I wanted when someone asked βwhatβs best?β
8874c51a, 1M context). DSF Vision-Exp and DeepSeek-V4.1-Flash are proven and currently dark. Full GLM-5.3 is a swap-in, not a daily driver.The table
| Alias | What | Engine | Context | Prefill | Decode C1 | Decode C8 |
|---|---|---|---|---|---|---|
| dsf0731 / DSFVE | DeepSeek-V4-Flash-Vision-Exp (0731-family + vision). Native FP4. | SGLang + DSpark Ξ³5 | 1M | ~32β36k tok/s through 128K; ~18k at 810K | 329 recipe-v2 (C1 historically bounced ~400Β±50) | 1,261 agg (C16 1,786; C64 2,985) |
| glm5.3 | Full GLM-5.3 744B NVFP4. Experimental recipe. | vLLM UVA + slot-cache + MTP K=2 | 256K daily (1M swap-in) | ~3.45k tok/s, 62 s @ 211K | 54.7 (K2+fuse daily) | not the headline β C1-bound by C2C expert-row copies |
| glm5.3f | GLM-5.3-Flash NVFP4 + DFlash2 block 7. Listening now. | SGLang + DFlash2 | 1M | ~22β26k (105K in 4.0 s) | 202.7 answer-only / 265 if thinking counted | 732 agg (91.5 / user) |
| dfs4.1 | DeepSeek-V4.1-Flash native. Engram in Grace, experts on UVA. Not promoted. | vLLM UVA + DSpark k=5 | 1M (KV ~2.2Γ) | ~18k (207K in 11.4 s; 972K in 85 s) | ~90 prose / 145β160 code+JSON | 241 agg (C16 ~312) |
C1 is single-stream output tok/s, warm. C8 is aggregate. Prefill is cold, nonce at the start of the prompt so the prefix cache cannot cheat. Engines and speculation are not interchangeable; do not rank a DSpark row against a DFlash2 row as if they were the same product.
Which one, for what
Interactive speed β DSFVE
Fastest first token, 1M, native vision, agent gates. That is why it was the production recipe on this Station. Recipe v2 is digest-pinned SGLang, static verification, no SPS table.
Daily driver β GLM-5.3-Flash
What is actually up. About 4Γ DSV4.1 on prose, about 4Γ full GLM-5.3, 1M context, tools 10/10 on the round-3 image. Use 202.7, not the old 234/252 β those counted max-effort thinking as output.
Read: round 3 Β· day-two recipe
Intelligence β full GLM-5.3
Capability buy, not a speed buy. Prefill is a different sport (~3.5k vs 18β35k). 256K is fully usable (18/18 needle to 211K). Keep it as a swap, not the default.
Read: 744B testing Β· research pass
DeepSeek-V4.1-Flash is worth having as a recipe. On one Station it is fetch-bound (~90 C1 prose) because a slice of routed experts lives in Grace. It was never promoted over DSF or Flash. Read: Flash 4.1 testing Β· plain-English version.
How to read the numbers
- Flash 202.7 vs 265 vs 234. Round 2 published ~252 C1 while counting max-effort thinking text as output β the chat template ignores
enable_thinkingand only honorsreasoning_effort. Round 3 re-measured answer-only at 202.7 and thinking-counted at 265. The original day-two 234 is the same class of mistake plus an older image. - DSF C1 329 vs ~400. Recipe v2 Iteration 13 is 329.2. Earlier Ξ³5 windows bounced 337β449. Treat per-stream C1 as noisy; read DSF at C8 and above.
- Full GLM-5.3 54.7. September 14 daily: fused hook + MTP K=2 on the 256K profile. Decode is GPU-bound; about half the step is
masked_row_copyof expert rows over C2C. - DSV4.1 ~90 / 18k prefill. v12 off60 + util 0.97, DSpark k=5, 1M. Decode is content-dependent because speculation acceptance is. Prefill is not the limiter; expert fetch is.
Sources
All four rows are dated measurements from posts on the GB300 topic. This page does not re-bench them.
- Recipe v2 from the DSFVE inner loop β C1 329.2, C8 1,260.7, C64 2,985.4; cold prefill 33.9kβ35.7k tok/s (8Kβ128K).
- DeepSeek-V4-Flash-Vision-Exp on One GB300 β prefill ladder through 810K; C1 bounce.
- GLM-5.3-Flash round 3 β 202.7 answer-only, 265 thinking counted, C8 732.
- GLM-5.3-Flash recipe β 105K prefill in 4.0 s (~26.3k tok/s).
- GB300 GLM 5.3 Testing β 54.69 C1 daily; ~3.45k prefill; 256K / 1M profiles.
- GB300 DeepSeek Flash 4.1 Testing β ~90 C1 prose, 18k prefill, C8 241 / C16 312.