satgeze/ ornith-9b-1m:mtp-q4_k_m

32 Downloads Updated 22 hours ago

Ornith-1.0-9B agentic coder: 1M context, vision, MTP baked in. Needle-verified.

vision tools thinking

ollama run satgeze/ornith-9b-1m:mtp-q4_k_m

curl http://localhost:11434/api/chat \
  -d '{
    "model": "satgeze/ornith-9b-1m:mtp-q4_k_m",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='satgeze/ornith-9b-1m:mtp-q4_k_m',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'satgeze/ornith-9b-1m:mtp-q4_k_m',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 22 hours ago

22 hours ago

a874f747d1df · 6.7GB ·

model

archqwen35

·

parameters9.2B

·

quantizationQ4_K_M

5.8GB

projector

archclip

·

parameters456M

·

quantizationF16

918MB

params

{ "num_ctx": 262144 }

19B

Readme

Ornith-1.0-9B, 1M context

DeepReinforce’s Ornith-1.0-9B (MIT, Qwen3.5 family) with a 1,048,576-token context baked in via YaRN metadata, vision tower attached, and the official MTP layer restored. Native renderer and parser are set, so tool calling works out of the box, including as a Claude Code backend:

ollama launch claude --model satgeze/ornith-9b-1m

Verified, not claimed

10 needles per rung at depths 5 to 95 percent, temperature 0, full ladder with no skipped rungs:

Config	64K to 524K	1M
q8_0 tag + f16 KV	¹⁰⁄₁₀ every rung	¹⁰⁄₁₀
latest (Q4) + q8 KV	¹⁰⁄₁₀ every rung	⁷⁄₁₀ (quantization tax, documented)

Pick your tag

Tag	Size	When
`latest` / `mtp-q4_k_m`	6.7GB	Default. MTP layer baked in
`q8_0` / `mtp-q8_0`	11GB	Max quality: the config that scored ¹⁰⁄₁₀ at 1M
`q6_k`, `q5_k_m`, `iq4_xs`, `q3_k_m`, `iq3_xxs`, `q2_k`, `iq2_m`	7.4 to 3.6GB	Smaller VRAM, descending quality
`bf16`	18GB	Reference precision

Every tag carries the 1M rope metadata and the vision tower. Raise context with /set parameter num_ctx 1048576 (needs ~32GB free for KV at the full million).

MTP note

Speculative-decoding tensors are included in the latest and q8_0 tags and load harmlessly, but Ollama has no speculative decoding engine yet, so the speed boost is dormant here. It activates automatically when Ollama ships it. For the +25 to 38 percent decode speedup today, run the same file with llama.cpp:

llama-server -m ornith-1.0-9b-1M-MTP-Q4_K_M.gguf -c 1048576 -np 1 --jinja \
  --spec-type draft-mtp --spec-draft-n-max 3 --mmproj mmproj-ornith-9b-f16.gguf

Links

Full repo, raw per-needle data, charts: Hugging Face | ModelScope Method and harness: github.com/satindergrewal/aviary-1m

Credits: DeepReinforce (Ornith-1.0, MIT), Qwen (base + MTP layer), protoLabsAI (MTP GGUF restoration), SatGeze (1M extension, vetting, certification).

![Ornith](https://huggingface.co/satgeze/Ornith-1.0-9B-1M-GGUF/resolve/main/banner.jpeg)

# Ornith-1.0-9B, 1M context

DeepReinforce's Ornith-1.0-9B (MIT, Qwen3.5 family) with a 1,048,576-token context baked in via YaRN metadata, vision tower attached, and the official MTP layer restored. Native renderer and parser are set, so tool calling works out of the box, including as a Claude Code backend:

``` 
ollama launch claude --model satgeze/ornith-9b-1m
```

## Verified, not claimed

![NIAH heatmap](https://huggingface.co/satgeze/Ornith-1.0-9B-1M-GGUF/resolve/main/niah_heatmap_9b.png)

10 needles per rung at depths 5 to 95 percent, temperature 0, full ladder with no skipped rungs:

| Config | 64K to 524K | 1M |
|---|---|---|
| q8_0 tag + f16 KV | 10/10 every rung | **10/10** |
| latest (Q4) + q8 KV | 10/10 every rung | 7/10 (quantization tax, documented) |

## Pick your tag

| Tag | Size | When |
|---|---|---|
| `latest` / `mtp-q4_k_m` | 6.7GB | Default. MTP layer baked in |
| `q8_0` / `mtp-q8_0` | 11GB | Max quality: the config that scored 10/10 at 1M |
| `q6_k`, `q5_k_m`, `iq4_xs`, `q3_k_m`, `iq3_xxs`, `q2_k`, `iq2_m` | 7.4 to 3.6GB | Smaller VRAM, descending quality |
| `bf16` | 18GB | Reference precision |

Every tag carries the 1M rope metadata and the vision tower. Raise context with `/set parameter num_ctx 1048576` (needs ~32GB free for KV at the full million).

## MTP note

Speculative-decoding tensors are included in the `latest` and `q8_0` tags and load harmlessly, but Ollama has no speculative decoding engine yet, so the speed boost is dormant here. It activates automatically when Ollama ships it. For the +25 to 38 percent decode speedup today, run the same file with llama.cpp:

```
llama-server -m ornith-1.0-9b-1M-MTP-Q4_K_M.gguf -c 1048576 -np 1 --jinja \
  --spec-type draft-mtp --spec-draft-n-max 3 --mmproj mmproj-ornith-9b-f16.gguf
```

## Links

Full repo, raw per-needle data, charts: [Hugging Face](https://huggingface.co/satgeze/Ornith-1.0-9B-1M-GGUF) | [ModelScope](https://www.modelscope.ai/models/satgeze/Ornith-1.0-9B-1M-GGUF)
Method and harness: [github.com/satindergrewal/aviary-1m](https://github.com/satindergrewal/aviary-1m)

Credits: DeepReinforce (Ornith-1.0, MIT), Qwen (base + MTP layer), protoLabsAI (MTP GGUF restoration), SatGeze (1M extension, vetting, certification).

Paste, drop or click to upload images (.png, .jpeg, .jpg, .svg, .gif)