---
type: reference
title: dealign Ornith-1.5-35B-A3B CRACK on Dankserver
description: dealignai/Ornith-1.5-35B-A3B-UNCENSORED (CRACK Q6_K) served by upstream llama.cpp on the RTX 3090. MoE model with the MTP head bundled; selected expert weights offloaded to CPU with an explicit --override-tensor. frank-default-small.
tags:
  - llamacpp
  - cuda
  - dankserver
  - moe
  - ornith
  - mtp
timestamp: 2026-08-24
---

# dealign Ornith-1.5-35B-A3B CRACK on Dankserver

> **Active 2026-08-25.** `frank-default-small` serves this model. It
> replaced the CRACK Q8_0 file on the same upstream llama.cpp binary (b10121).
> The Q6_K file was downloaded from the model repository and validated with
> `llama-gguf` before installation.

`frank-default-small` (port 7770) serves
`Ornith-1.5-35B-A3B-CRACK-Q6_K.gguf` from `dealignai` on Hugging Face. It is
a **MoE** model (256 experts, 8 active) with a hybrid GatedDeltaNet (SSM) +
attention architecture, abliterated (uncensored). The MTP speculative head
is **bundled in the GGUF**, so `draft-mtp` auto-detects it from the target
model.

## Model Files

| File | Size |
|------|------|
| `Ornith-1.5-35B-A3B-CRACK-Q6_K.gguf` | 27.2 GiB |
| `mmproj-Ornith-1.5-35B-A3B-f16.gguf` (vision) | 0.84 GiB |

Store the files under `/home/dank/models/Ornith-1.5-35B-A3B-CRACK/`.

## Layer Audit and Expert Offload

The Q6_K file contains 41 transformer blocks and 753 tensors. Each block
contains 630 MiB of expert tensors. Blocks 3, 7, 11, 15, 19, 23, 27, 31,
35, and 39 contain full attention. Block 40 contains the bundled MTP head.

The non-expert tensors total 2.02 GiB. The layer totals are 660.96 MiB for
ordinary SSM blocks, 655.81 MiB for full-attention blocks, and 662.40 MiB for
block 40. The expert portion is 630 MiB in every block.

The provisional service keeps all non-expert tensors on CUDA. It also keeps
expert tensors from blocks 33 through 40 on CUDA. Those blocks sit next to the
output and include the bundled MTP head. Expert tensors from blocks 0 through
32 move to CPU through this explicit override:

```
OVERRIDE_EXPERT='blk\.(0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32)\.ffn_(up|down|gate|gate_up)_(ch|)exps'
OVERRIDE_BUFT=CPU
```

The previous five-block configuration placed 3.69 GiB of expert weights on
CUDA and 22.13 GiB on CPU. This provisional eight-block configuration adds
approximately 1.89 GiB of expert weights to CUDA while reducing the context
pool by 98304 cells. The final placement will be selected after measuring this
new headroom and discussing layer priorities.

To list the buffer type names on the host, run a bogus override:

```
llama-server --model <ANY> -c 64 -ot foo=BADTYPE
```

On Dankserver the available buffer types are `CPU` and `CUDA0`.

## Speculative Decoding

MoE model with a bundled MTP head -> native `draft-mtp`. The head is
auto-detected from the target (b10121 creates the MTP draft context against
the model file). ngram under MTP hurts on this family; MTP alone is the
proven best. `--spec-draft-n-max 4` is the knee.

```
--spec-type draft-mtp
--spec-draft-n-max 4
```

## Launch Excerpt

```
llama-server \
  -m Ornith-1.5-35B-A3B-CRACK-Q6_K.gguf \
  --mmproj mmproj-Ornith-1.5-35B-A3B-f16.gguf \
  --chat-template-file chat-template.jinja \
  --spec-type draft-mtp \
  --override-tensor "blk\\.(0|1|2|3|4|5|6|7|8|9|10|11|12|13|14|15|16|17|18|19|20|21|22|23|24|25|26|27|28|29|30|31|32)\\.ffn_(up|down|gate|gate_up)_(ch|)exps=CPU" \
  -ngl all -c 491520 -np 3 \
  --cache-type-k f16 --cache-type-v f16 --no-mmap \
  --temp 1.0 --top-p 0.95 --top-k 20
```

## Known Issues

1. The 27.2 GiB Q6_K file and 0.84 GiB vision projector fit the 62 GiB RAM
   box with the selected expert offload.
2. The provisional context pool contains 491520 cells. Three slots receive
   163840 cells each, which is 96K below the model context per stream.
3. `--n-cpu-moe` remains supported. This service currently prefers the
   explicit tensor override because it preserves the final eight expert
   blocks on CUDA pending the next layer-priority decision.
4. The service uses the Qwen-Sharp template from
   `peculiar-ragdoll/Qwen-Sharp-Chat-Templates` through
   `scripts/chat-template.jinja`.

## See Also

- [[deployment]] — deploy and update process
- [[Dankserver System Specs]] — the box
- [[frank-redo-2026-08-17]] — fleet topology history