# MiMo V2.5 with Dflash FP8 KV 1.5M+ & Deep Speed up to 30 tk/s

**URL:** <https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945>\
**Category:** DGX Spark / GB10\
**Created:** [July 7, 2026, 4:14pm UTC](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945 "2026-07-07T16:14:04Z")\
**Posts on this page:** 20\
**Page:** 1

<div class="post-metadata">

**Author:** ![renek](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@renek](https://forums.developer.nvidia.com/u/renek)\
**Post date:** [July 7, 2026, 4:14pm UTC](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945/1 "2026-07-07T16:14:04Z")

</div>

Hi all,

I was working with Claude on MiMo to optimize it more with Dflash for our 2x Spark folks.

For Details please check the Repo but we have probably the best speeds anybody could achieve with such KV Cache on FP8 and no quality loss, im even working to improve further but I did many tests today to confirm and we keep the speed even at 400K context.

Most important is Data on todays update, you see it little bit down the Text.

Have Fun exploring and editing to your likes.

> **[GitHub - HeNryous/mimo-v25-dflash-dgx-spark: vLLM recipe: Xiaomi MiMo-V2.5 + DFlash speculative...](https://github.com/HeNryous/mimo-v25-dflash-dgx-spark)**
>
> vLLM recipe: Xiaomi MiMo-V2.5 + DFlash speculative decoding + fp8 KV cache on 2x DGX Spark (GB10/sm\_121)

---

<div class="post-metadata">

**Author:** ![sjug](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@sjug](https://forums.developer.nvidia.com/u/sjug)\
**Post date:** [July 7, 2026, 5:19pm UTC](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945/2 "2026-07-07T17:19:29Z")

</div>

Thanks for the share, appreciate you not nuking the kv cache.

---

<div class="post-metadata">

**Author:** ![renek](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@renek](https://forums.developer.nvidia.com/u/renek)\
**Post date:** [July 7, 2026, 5:22pm UTC](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945/3 "2026-07-07T17:22:40Z")

</div>

I tried FP4 but didn’t work out for me, also I get up to 70tk/s single stream and just had 700K kv cache usage 20tk/s multi stream, let me know if there are any problems!

---

<div class="post-metadata">

**Author:** ![sjug](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@sjug](https://forums.developer.nvidia.com/u/sjug)\
**Post date:** [July 7, 2026, 6:13pm UTC](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945/4 "2026-07-07T18:13:15Z")

</div>

Is this your custom quant?

---

<div class="post-metadata">

**Author:** ![renek](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@renek](https://forums.developer.nvidia.com/u/renek)\
**Post date:** [July 7, 2026, 6:16pm UTC](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945/5 "2026-07-07T18:16:18Z")

</div>

No lukealonso NVFP4

---

<div class="post-metadata">

**Author:** ![jc2375](https://sea2.discourse-cdn.com/nvidia/user_avatar/forums.developer.nvidia.com/jc2375/32/522115_2.png) [@jc2375](https://forums.developer.nvidia.com/u/jc2375)\
**Post date:** [July 8, 2026, 3:35pm UTC](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945/6 "2026-07-08T15:35:22Z")

</div>

Reading through your repo, this requires a custom vllm build. Is that included in your shell script or a separate step? Thanks for the efforts

---

<div class="post-metadata">

**Author:** ![renek](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@renek](https://forums.developer.nvidia.com/u/renek)\
**Post date:** [July 8, 2026, 4:27pm UTC](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945/7 "2026-07-08T16:27:15Z")

</div>

It’s standard vLLM but specific version, not shipped with my repo, I only provide the mods

---

<div class="post-metadata">

**Author:** ![renek](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@renek](https://forums.developer.nvidia.com/u/renek)\
**Post date:** [July 9, 2026, 10:09am UTC](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945/8 "2026-07-09T10:09:20Z")

</div>

as it was requested i added my image as well to pull and just patch it with my mods

---

<div class="post-metadata">

**Author:** ![renek](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@renek](https://forums.developer.nvidia.com/u/renek)\
**Post date:** [July 9, 2026, 10:53am UTC](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945/9 "2026-07-09T10:53:30Z")

</div>

as well i will add the DCP path, estimated win:  
**~1,3-1,6×** deep (prose 14-18→20-27, struct 28-35→38-52 @ 400K) \*\*debunked

after that i will move on to M3

EDIT: M3 not viable for me because of size and REAP doesnt make too much sense

---

<div class="post-metadata">

**Author:** ![tonyd615](https://sea2.discourse-cdn.com/nvidia/user_avatar/forums.developer.nvidia.com/tonyd615/32/511662_2.png) [@tonyd615](https://forums.developer.nvidia.com/u/tonyd615)\
**Post date:** [July 9, 2026, 6:36pm UTC](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945/10 "2026-07-09T18:36:50Z")

</div>

[GitHub - tonyd2wild/MiMo-V2.5-DFlash-1M-ctx-NVFP4-KV-2x-DGX-Spark: MiMo-V2.5 NVFP4 4-bit weights + NVFP4 4-bit KV cache + DFlash speculative decoding on 2× NVIDIA DGX Spark — 1M context, 3.4M-token KV pool · GitHub](https://github.com/tonyd2wild/MiMo-V2.5-DFlash-1M-ctx-NVFP4-KV-2x-DGX-Spark) 60 tok/s

---

<div class="post-metadata">

**Author:** ![renek](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@renek](https://forums.developer.nvidia.com/u/renek)\
**Post date:** [July 9, 2026, 6:39pm UTC](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945/11 "2026-07-09T18:39:54Z")

</div>

did you check quality degradation with NVFP4 KV Cache ? token wise i was selective over 70 sometimes

---

<div class="post-metadata">

**Author:** ![tonyd615](https://sea2.discourse-cdn.com/nvidia/user_avatar/forums.developer.nvidia.com/tonyd615/32/511662_2.png) [@tonyd615](https://forums.developer.nvidia.com/u/tonyd615)\
**Post date:** [July 9, 2026, 7:11pm UTC](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945/12 "2026-07-09T19:11:49Z")

</div>

For tool calling it did not degrade for the workflows I do

---

<div class="post-metadata">

**Author:** ![renek](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@renek](https://forums.developer.nvidia.com/u/renek)\
**Post date:** [July 9, 2026, 7:22pm UTC](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945/13 "2026-07-09T19:22:32Z")

</div>

but still trade off, if you want more kv cache acceptance and quality go down… :/ thats why i sticked with FP8

---

<div class="post-metadata">

**Author:** ![renek](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@renek](https://forums.developer.nvidia.com/u/renek)\
**Post date:** [July 10, 2026, 9:01am UTC](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945/14 "2026-07-10T09:01:28Z")

</div>

@0rand repo updated to fables optimized state

---

<div class="post-metadata">

**Author:** ![truetotosse](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@truetotosse](https://forums.developer.nvidia.com/u/truetotosse)\
**Post date:** [July 10, 2026, 11:53am UTC](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945/15 "2026-07-10T11:53:43Z")

</div>

@renek @tonyd615 could you please make landing readmes more human readable and move all the history to some other file?

The work you do is greate but it is hard to figure out what is going on.

Like we need some standard how to do good GitHub recipe + with some data what it can achieve.

For example: main readme just pure how to + expected results.

Tool bench, tps c1, c4 etc. at 128, at 100k, at more depths.

---

<div class="post-metadata">

**Author:** ![renek](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@renek](https://forums.developer.nvidia.com/u/renek)\
**Post date:** [July 10, 2026, 11:56am UTC](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945/16 "2026-07-10T11:56:44Z")

</div>

I advised Claude to change it that way, thanks for your input! :)

---

<div class="post-metadata">

**Author:** ![truetotosse](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@truetotosse](https://forums.developer.nvidia.com/u/truetotosse)\
**Post date:** [July 10, 2026, 12:07pm UTC](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945/17 "2026-07-10T12:07:40Z")

</div>

Thank you. Reads very well now.

---

<div class="post-metadata">

**Author:** ![0rand](https://sea2.discourse-cdn.com/nvidia/user_avatar/forums.developer.nvidia.com/0rand/32/512004_2.png) [@0rand](https://forums.developer.nvidia.com/u/0rand)\
**Post date:** [July 10, 2026, 10:25pm UTC](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945/19 "2026-07-10T22:25:23Z")

</div>

Here comes the test. Confirmed - with latest fixed it works to full 1M context. Surely you don’t want to push there - over 40 minutes to prefill but if you need and out of option (need to compress etc) - you can. And decode stays sane.

Amazing job!

```auto
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━┓
┃ Test ┃ c ┃ pp t/s ┃ tg t/s ┃ TTFT (ms) ┃ Total (ms) ┃ Tokens ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━┩
│ pp2048 tg128 @ d0 │ c1 │ 3,501 │ 31.2 │ 833 │ 4,687 │ 2048+128 │
│ pp2048 tg128 @ d0 │ c2 │ 1,536 │ 29.8 │ 1,752 │ 9,081 │ 2048+128 │
│ pp2048 tg128 @ d0 │ c4 │ 2,992 │ 45.0 │ 2,254 │ 11,104 │ 2048+128 │
│ pp2048 tg128 @ d4096 │ c1 │ 3,427 │ 29.2 │ 2,045 │ 6,171 │ 2048+128 │
│ pp2048 tg128 @ d4096 │ c2 │ 3,277 │ 42.2 │ 3,427 │ 8,850 │ 2048+128 │
│ pp2048 tg128 @ d4096 │ c4 │ 2,918 │ 56.3 │ 8,080 │ 15,221 │ 2048+128 │
│ pp2048 tg128 @ d16384 │ c1 │ 3,234 │ 27.0 │ 5,952 │ 10,428 │ 2048+128 │
│ pp2048 tg128 @ d16384 │ c2 │ 3,144 │ 43.2 │ 11,525 │ 16,963 │ 2048+128 │
│ pp2048 tg128 @ d16384 │ c4 │ 3,124 │ 40.2 │ 22,320 │ 30,628 │ 2048+128 │
│ pp2048 tg128 @ d65536 │ c1 │ 2,429 │ 28.5 │ 28,075 │ 32,305 │ 2048+128 │
│ pp2048 tg128 @ d65536 │ c2 │ 2,359 │ 33.1 │ 56,343 │ 62,356 │ 2048+128 │
│ pp2048 tg128 @ d65536 │ c4 │ 2,494 │ 34.2 │ 107,055 │ 116,995 │ 2048+128 │
└─────────────────────────────────────────────┴────────────┴──────────────────────┴──────────────────────┴───────────────────────┴───────────────────────┴───────────────────────┘
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┓
┃ Test ┃ c ┃ pp t/s ┃ tg t/s ┃ TTFT (ms) ┃ Total (ms) ┃ Tokens ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━┩
│ pp2048 tg128 @ d128000 │ c1 │ 1,821 │ 31.7 │ 71,665 │ 75,442 │ 2048+128 │
│ pp2048 tg128 @ d128000 │ c2 │ 1,861 │ 39.9 │ 139,317 │ 144,767 │ 2048+128 │
│ pp2048 tg128 @ d128000 │ c4 │ 1,893 │ 46.4 │ 274,281 │ 282,930 │ 2048+128 │
└───────────────────────────────────────────────┴────────────┴──────────────────────┴─────────────────────┴───────────────────────┴───────────────────────┴──────────────────────┘
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┓
┃ Test ┃ c ┃ pp t/s ┃ tg t/s ┃ TTFT (ms) ┃ Total (ms) ┃ Tokens ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━┩
│ pp2048 tg128 @ d256000 │ c1 │ 1,186 │ 24.2 │ 217,887 │ 222,874 │ 2048+128 │
└───────────────────────────────────────────────┴────────────┴──────────────────────┴─────────────────────┴───────────────────────┴───────────────────────┴──────────────────────┘
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┓
┃ Test ┃ c ┃ pp t/s ┃ tg t/s ┃ TTFT (ms) ┃ Total (ms) ┃ Tokens ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━┩
│ pp2048 tg128 @ d512000 │ c1 │ 704 │ 22.9 │ 730,480 │ 735,774 │ 2048+128 │
└───────────────────────────────────────────────┴────────────┴──────────────────────┴─────────────────────┴───────────────────────┴───────────────────────┴──────────────────────┘
┏━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━━┳━━━━━━━━━━━━━━━━━━━━━━┓
┃ Test ┃ c ┃ pp t/s ┃ tg t/s ┃ TTFT (ms) ┃ Total (ms) ┃ Tokens ┃
┡━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━━╇━━━━━━━━━━━━━━━━━━━━━━┩
│ pp2048 tg128 @ d990000 │ c1 │ 393 │ 17.7 │ 2,524,584 │ 2,531,574 │ 2048+128 │
└───────────────────────────────────────────────┴────────────┴──────────────────────┴─────────────────────┴───────────────────────┴───────────────────────┴──────────────────────┘
```

---

<div class="post-metadata">

**Author:** ![0rand](https://sea2.discourse-cdn.com/nvidia/user_avatar/forums.developer.nvidia.com/0rand/32/512004_2.png) [@0rand](https://forums.developer.nvidia.com/u/0rand)\
**Post date:** [July 10, 2026, 10:35pm UTC](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945/20 "2026-07-10T22:35:10Z")

</div>

At 4 concurrents on content test (hardmode) it pushes up to 100 t/s decode and 120 on 8, not bad at all for multi-threaded work in low context environment.

```auto
╭──────────────────────────────────────────────────────────────────────────── ─── ◆ spec-live ◆ ─── ─────────────────────────────────────────────────────────────────────────────╮
│ ◇ SPECULATIVE DECODING MONITOR ╌╌ ⟨ Draft Model ⟩ ⏱ 04:17 │ 📡 239 │ k=4 │
│ ▸ /root/.cache/huggingface/MiMo-oproj-mxfp8 ← mimo-dflash-test │
│ │
│ ◈ ACCEPTANCE RATE ━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━━╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌╌ 63.2% τ=2.5/4 [k=4] │
│ ✦ Excellent — draft model is well-aligned │
│ │
│ τ Acc Length 2.53 │ Draft Window 4.0 │ Spec Tokens 4 │
│ Accepted t/s 92.2 │ Drafted t/s 119.2 │ Waste Ratio 36.8% │
│ Gen t/s 121.9 │ │
│ ╶──────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╴ │
│ │
│ ╭────────────────────── ⚙ Engine & Session ──────────────────────╮ ╭────────────────────────────────────────── 📊 History (60/60s) ──────────────────────────────────────────╮ │
│ │ KV Cache ▱▱▱▱▱▱▱▱▱▱ 1% │ │ Accept Rate ▃▃▂▂▂▂▁▁ ▁▁▁▂▁▂▂▂▃▄▄▄▄▄▅▅▆▆▆▇▇▇▇█ 63.2% ↕62–63% │ │
│ │ Prefix Cache 12.9% │ │ Gen t/s ▃▄▃▄▄▂▄▄▃▃▄▄▂▂▂ ▂▆▇▅▅▄▇▃▇▄▇▆▇▃▆▄▄▆▄█▅▆▄▇▄▅▇ 121.9 ↕4–123 │ │
│ │ Running 8 │ │ Accepted t/s ▃▂▂▄▂▃▃▂▃▃▂▂▄▄▂ ▂▅▆▆▅▄▇▃▇▃▇▆█▃▆▄▃▇▄▇▅▆▄▇▃▄▇ 92.2 ↕5–95 │ │
│ │ Waiting 0 │ │ Waste ▃▃▂▂▂▂▁▁ ▁▁▁▂▁▂▂▂▃▄▄▄▄▄▅▅▆▆▆▇▇▇▇█ 37% ↕37–38% │ │
│ │ Prompt t/s 2,451 │ ╰─────────────────────────────────────────────────────────────────────────────────────────────────────────╯ │
│ │ │ ╭────────────────────────────────────────── ⌀ Rolling Averages ───────────────────────────────────────────╮ │
│ │ Accepted 12,658 │ │ Avg α 62.3% │ │
│ │ Drafted 20,028 │ │ Avg Gen t/s 75.1 │ │
│ │ Session α 63.2% │ │ Avg Acc t/s 55.1 │ │
│ ╰────────────────────────────────────────────────────────────────╯ ╰─────────────────────────────────────────────────────────────────────────────────────────────────────────╯ │
│ │
│ ╭──────────────────────────────────────────────────────────────────────── ▊ Per-Position Acceptance ─────────────────────────────────────────────────────────────────────────╮ │
│ │ p0 █████████████████████████░░░░░ 85% p1 ████████████████████░░░░░░░░░░ 69% p2 ████████████████░░░░░░░░░░░░░░ 56% p3 █████████████░░░░░░░░░░░░░░░░░ 44% │ │
│ │ 4/4 effective positions (>20%) │ γ=0.80/pos │ │
│ ╰────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────────╯ │
╰───────────────────────────────────────────────────────────── Ctrl+R reset · Ctrl+C exit · Refreshing every 1s ─────────────────────────────────────────────────────────────

```

---

<div class="post-metadata">

**Author:** ![renek](https://developer.download.nvidia.com/images/forums/profile-default-devtalk-84.png) [@renek](https://forums.developer.nvidia.com/u/renek)\
**Post date:** [July 10, 2026, 10:40pm UTC](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945/21 "2026-07-10T22:40:24Z")

</div>

Thank you for testing! :)

[Next page](https://forums.developer.nvidia.com/t/mimo-v2-5-with-dflash-fp8-kv-1-5m-deep-speed-up-to-30-tk-s/375945.md?page=2)