satgeze/ ornith-35b-1m:q2_k

40 Downloads Updated 21 hours ago

Ornith-1.0-35B MoE (APEX-Compact 17GB): 1M context, vision, fits 32GB cards. 50/50 needles to 524K.

vision tools thinking

ollama run satgeze/ornith-35b-1m:q2_k

curl http://localhost:11434/api/chat \
  -d '{
    "model": "satgeze/ornith-35b-1m:q2_k",
    "messages": [{"role": "user", "content": "Hello!"}]
  }'

from ollama import chat

response = chat(
    model='satgeze/ornith-35b-1m:q2_k',
    messages=[{'role': 'user', 'content': 'Hello!'}],
)
print(response.message.content)

import ollama from 'ollama'

const response = await ollama.chat({
  model: 'satgeze/ornith-35b-1m:q2_k',
  messages: [{role: 'user', content: 'Hello!'}],
})
console.log(response.message.content)

Details

Updated 22 hours ago

22 hours ago

79787fc60dd9 · 14GB ·

model

archqwen35moe

·

parameters34.7B

·

quantizationQ2_K

13GB

projector

archclip

·

parameters447M

·

quantizationF16

899MB

params

{ "num_ctx": 262144 }

19B

Readme

Ornith-1.0-35B, 1M context

DeepReinforce’s Ornith-1.0-35B (MoE, 3B active, MIT, Qwen3.5 family) with a 1,048,576-token context baked in, certified needle-perfect through the full range, with vision attached. Native renderer and parser are set, so tool calling works out of the box:

ollama launch claude --model satgeze/ornith-35b-1m

Verified, not claimed

⁵⁰⁄₅₀ needles through 1M (f16 KV, temperature 0, no skipped rungs). The default APEX-Compact build separately scored ⁵⁰⁄₅₀ through 524K on a single RTX 5090.

Pick your tag

Tag	Size	When
`latest` / `apex-compact-mtp`	18GB	Default. SC117 APEX imatrix build, fits 32GB cards, MTP included
`q6_k-mtp`	30GB	The fast one: +43 percent decode under llama.cpp, for 64GB Macs and big GPUs
`q8_0`, `q6_k`, `q5_k_m`, `q4_k_m`, `iq4_xs`, `q3_k_m`, `iq3_xxs`, `q2_k`, `iq2_m`	37 to 12GB	Plain ladder, descending size
`bf16`	69GB	Reference precision

Every tag carries the 1M rope metadata and vision. KV runs about 20KB per token thanks to hybrid attention, so 1M context needs ~20GB of KV on top of weights. Raise with /set parameter num_ctx.

MTP note

The latest and q6_k-mtp tags include the MTP layer (dormant in Ollama until speculative decoding ships there; +14 to +43 percent today under llama.cpp with --spec-type draft-mtp).

Links

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

Credits: DeepReinforce (MIT), Qwen (base + MTP layer), wang-yang and SC117 (MTP GGUF builds), SatGeze (1M extension, vetting, certification).

![Ornith](https://huggingface.co/satgeze/Ornith-1.0-35B-1M-GGUF/resolve/main/banner.jpeg)

# Ornith-1.0-35B, 1M context

DeepReinforce's Ornith-1.0-35B (MoE, 3B active, MIT, Qwen3.5 family) with a 1,048,576-token context baked in, certified needle-perfect through the full range, with vision attached. Native renderer and parser are set, so tool calling works out of the box:

```
ollama launch claude --model satgeze/ornith-35b-1m
```

## Verified, not claimed

![NIAH heatmap](https://huggingface.co/satgeze/Ornith-1.0-35B-1M-GGUF/resolve/main/niah_heatmap.png)

**50/50 needles through 1M** (f16 KV, temperature 0, no skipped rungs). The default APEX-Compact build separately scored 50/50 through 524K on a single RTX 5090.

## Pick your tag

| Tag | Size | When |
|---|---|---|
| `latest` / `apex-compact-mtp` | 18GB | Default. SC117 APEX imatrix build, fits 32GB cards, MTP included |
| `q6_k-mtp` | 30GB | The fast one: +43 percent decode under llama.cpp, for 64GB Macs and big GPUs |
| `q8_0`, `q6_k`, `q5_k_m`, `q4_k_m`, `iq4_xs`, `q3_k_m`, `iq3_xxs`, `q2_k`, `iq2_m` | 37 to 12GB | Plain ladder, descending size |
| `bf16` | 69GB | Reference precision |

Every tag carries the 1M rope metadata and vision. KV runs about 20KB per token thanks to hybrid attention, so 1M context needs ~20GB of KV on top of weights. Raise with `/set parameter num_ctx`.

## MTP note

The `latest` and `q6_k-mtp` tags include the MTP layer (dormant in Ollama until speculative decoding ships there; +14 to +43 percent today under llama.cpp with `--spec-type draft-mtp`).

## Links

Full repo, raw data, charts: [Hugging Face](https://huggingface.co/satgeze/Ornith-1.0-35B-1M-GGUF) | [ModelScope](https://www.modelscope.ai/models/satgeze/Ornith-1.0-35B-1M-GGUF)
Method and harness: [github.com/satindergrewal/aviary-1m](https://github.com/satindergrewal/aviary-1m)

Credits: DeepReinforce (MIT), Qwen (base + MTP layer), wang-yang and SC117 (MTP GGUF builds), SatGeze (1M extension, vetting, certification).

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