01
FASTEST
GLM 5.3
#1BLACKBOX AI461.0t/s
#2Inco (FAST)388.2t/s
#3Nebius (FP4)373.2t/s
#4Makora295.3t/s
#5Together AI240.6t/s
+19% AHEAD OF Inco (FAST)VIEW MODEL →
OUTPUT SPEED · TOKENS PER SECONDRANKED FASTEST FIRST
02
FASTEST
Kimi K3MAX
#1BLACKBOX AI292.6t/s
#2Inco (FAST)255.0t/s
#3Databricks143.0t/s
#4Baseten94.0t/s
#5Together AI42.0t/s
+15% AHEAD OF Inco (FAST)VIEW MODEL →
First on both boards
Two models, two boards, one provider at the top of both. The gap over the runner-up is the part that shows up in your own streaming latency.
461.0 t/s
GLM 5.3 · RANKED #1
+19%
AHEAD OF THE #2 PROVIDER
292.6 t/s
KIMI K3 MAX · RANKED #1
+15%
AHEAD OF THE #2 PROVIDER
What the
numbers mean
Speed claims are only useful with their units, their scope, and their shelf life attached.
- OUTPUT SPEED
- Tokens per second on streaming completions, measured per provider endpoint. Higher is better. Every value on a board uses the same unit and the same scale.
- ENDPOINT TIERS
- FAST, ULTRA, and MAX are the providers' own names for their speed tiers. A row is one endpoint, not a company average, so a provider can appear more than once.
- SAME WEIGHTS
- Every provider on a board serves the same open weights. The spread is the serving stack: kernels, quantization, hardware, and scheduling.
- A SNAPSHOT
- Provider speeds move as endpoints are re-tuned and capacity changes. Read a board as the state of the run, not a standing guarantee.
Measure it on your own prompts.
Both models run on the same key and the same endpoint. Point your SDK at it, stream a completion, and count the tokens yourself.