PrimaLabs

PrimaLabs is adaptive inference infrastructure for open models. Faster tokens, higher throughput, lower cost per token as the result

[PrimaLabs @PrimalabsAI]

Fastest MiMo-V2.6-Pro endpoint on Artificial Analysis.

407 tok/s at 10K input. 9x Xiaomi's own API.
5.6s to first token vs 47.2s.
$0.87 per 1M output tokens.

Try it today.

[PrimaLabs @PrimalabsAI]

@Spangler3000 @chrisfzz2 @jundotkim the 262k to 500k drop is pretty reasonable. curious whether those numbers hold with tool-call traces in the context too.

[PrimaLabs @PrimalabsAI]

@SSHCodes @Da7_Tech 2,002 tok/s prefill on one card is the interesting number here. how does decode hold up once the prompt is short?

[PrimaLabs @PrimalabsAI]

@Moha__niang MOPD is the patched checkpoint for the tool-call repetition bug, not a new video model. the HF tag is just noisy.

[PrimaLabs @PrimalabsAI]

@rikkarth @Darkfibr3 @LilithDatura we're hosting it. MiMo-V2.6-Pro hits 407 tok/s on Artificial Analysis (Sep 24 snapshot), about $0.87 per 1M output tokens.
https://t.co/ty6LGd3Rlv
https://t.co/hs41vV0L9l

[PrimaLabs @PrimalabsAI]

@agonzalp basically this. once the weights are open, the product is routing, latency, and how you actually ship on top of it

[PrimaLabs @PrimalabsAI]

@aquilesfd yeah speed is mostly the endpoint, not the weights. MiMo-V2.6-Pro hits 407 tok/s on Artificial Analysis vs 45 on Xiaomi's API. same model, very different feel.
https://t.co/ty6LGd4pb3
https://t.co/hs41vV1iYT

[PrimaLabs @PrimalabsAI]

@stretchcloud this is the part that actually matters. shipping the learning loop, not just the checkpoint, is what lets other teams stop replaying the same experiments

[PrimaLabs @PrimalabsAI]

@AndrewDelPrete yeah that pairing is hard to beat. mimo for the heavy agent loops, flash when you just need something cheap and fast enough

[PrimaLabs @PrimalabsAI]

@gauravdhiman_ai @OpenRouter Yes- we DM’d you a voucher for free credits :)

[PrimaLabs @PrimalabsAI]

@AdolfoUsier @LeoBuilds_ @NovaXCode Coding speed matters. PrimaLabs hits 407 tok/s on MiMo-V2.6-Pro vs 45 on Xiaomi's API (Artificial Analysis, Sep 24).
https://t.co/ty6LGd3Rlv
https://t.co/hs41vV0L9l

[PrimaLabs @PrimalabsAI]

@gmi_cloud Good to see more endpoints. PrimaLabs is currently #1 MiMo-V2.6-Pro on Artificial Analysis at 407 output tok/s.
https://t.co/ty6LGd4pb3
https://t.co/hs41vV1iYT

[PrimaLabs @PrimalabsAI]

@BetterCallMedhi Yeah, open weights change the economics. PrimaLabs serves MiMo-V2.6-Pro at $0.87 per 1M output tokens, #1 speed on Artificial Analysis.
https://t.co/ty6LGd4pb3
https://t.co/hs41vV1iYT

[PrimaLabs @PrimalabsAI]

@ValsAI Solid recap. We're the fastest MiMo-V2.6-Pro endpoint on Artificial Analysis at 407 tok/s, about 9x Xiaomi's API.
https://t.co/ty6LGd3Rlv
https://t.co/hs41vV0L9l

[PrimaLabs @PrimalabsAI]

@FireworksAI_HQ Nice. Speed holds up when context gets long too. PrimaLabs is #1 MiMo-V2.6-Pro on Artificial Analysis: 327 tok/s at 100K input.
https://t.co/ty6LGd4pb3
https://t.co/hs41vV1iYT

[PrimaLabs @PrimalabsAI]

@michaelking_z this is the production test. PrimaLabs lists MiMo-V2.6-Pro at $0.87 per 1M output tokens, alongside 407 tok/s at 10K input on AA. https://t.co/ty6LGd4pb3 https://t.co/hs41vV1iYT

[PrimaLabs @PrimalabsAI]

@MichaelGannotti this is a useful signal. MiMo-V2.6-Pro also leads the AA provider table for output speed: 407 tok/s at 10K input. https://t.co/ty6LGd4pb3 https://t.co/hs41vV1iYT

[PrimaLabs @PrimalabsAI]

@XiaomiMiMo the gallery is a strong demo. for production, the provider table matters too. AA shows PrimaLabs at 407 tok/s at 10K input. https://t.co/ty6LGd3Rlv https://t.co/hs41vV0L9l

[PrimaLabs @PrimalabsAI]

@mcndjxlefnd @tphuang @opencode the benchmark point is fair. on AA, PrimaLabs measured 407 tok/s at 10K input vs 45 on Xiaomi's API. worth checking the provider table. https://t.co/ty6LGd3Rlv https://t.co/hs41vV0L9l

[PrimaLabs @PrimalabsAI]

@DarylUrbanski @MiaAI_lab @XiaomiMiMo good to hear. on AA, PrimaLabs is the fastest MiMo-V2.6-Pro endpoint at every tested context. 407 tok/s at 10K input vs 45 on Xiaomi's API. https://t.co/ty6LGd3Rlv https://t.co/hs41vV0L9l

[PrimaLabs @PrimalabsAI]

@mariathasanish @akjags @thekaransinghal @OpenAI @deepseek @XiaomiMiMo Better tasks and environments are the durable advantage. The serving layer still determines agent feel: 5.6s TTFT on our endpoint.

https://t.co/ty6LGd3Rlv
https://t.co/hs41vV0L9l