infosave/Ornith-1.5-122B-cmf

Ornith-1.5-122B — CMF

This repository contains standalone CMF packages for the text tower of Ornith-1.5-122B, executed by the Rust cortiq runtime. The artifacts were converted from ornith-ai/Ornith-1.5-122B revision 7681ef608624e7532d8164128bcffdbdbaf1b1af.

Text only. The upstream 122B checkpoint includes a visual tower; CMF intentionally packages only model.language_model. These files do not take image or video input and do not claim Ornith's vision-language behaviour.

Runtime requirement: use cortiq 0.8.30 or newer. Earlier versions must not be assumed compatible with this Qwen3.5/GDN MoE text profile.

Quick start · 122B artifacts · Measured Vulkan run · Limits · Русская документация

Quick start

The 122B q2tp-v2 file is the smallest published 122B profile. Download one model and its matching external checksum, then validate both transport and CMF integrity:

REPO=infosave/Ornith-1.5-122B-cmf
MODEL=ornith-1.5-122b-q2tp-v2.cmf

hf download "$REPO" "$MODEL" "$MODEL.sha256" --local-dir .
sha256sum -c "$MODEL.sha256"                 # macOS: shasum -a 256 -c "$MODEL.sha256"
cortiq verify "$MODEL"

cortiq run "$MODEL" \
  --prompt "Explain the difference between a mutex and a semaphore." \
  --no-think --max-tokens 128

cortiq verify validates the CMF envelope, directory, and per-tensor hashes. It is not a quality, memory-fit, or performance test. For a deterministic smoke command use --greedy; it cannot be combined with sampling penalties.

Ornith-1.5-122B artifacts

file	exact bytes	SHA-256	intended trade-off
ornith-1.5-122b-q2tp-v2.cmf	48,097,144,622	ebe80b9a6c112a7dd82fe31dca8058c7cc8e7fe03616f49e1431c1b8060ad01a	lowest storage footprint; fastest of these three on the measured 24 GB Vulkan host
ornith-1.5-122b-q4tp.cmf	67,904,220,974	fb4706070790e4b3fe1eda81ec7c3fb997a8653334092f0b4e741b6f4d26c13a	higher-precision direct q4tp profile
ornith-1.5-122b-q8_2f.cmf	125,055,889,198	44f92bfdbeb0ca94f1097cb85fc1eb6bce1ba55f405b2bd558cab4f778a1c429	high-precision CMF comparison profile; not FP16/upstream ground truth

All three are mixed-precision CMF files; their names are not claims that every stored value has a uniform bit width:

• q4tp: routed and shared MoE matrices use Q4TiledP; the backbone and output matrices follow the hybrid Q8_2f policy where applicable, while norms and controls retain their required floating-point representation.
• q2tp-v2: routed/shared gate_proj and up_proj use Q2TiledP; the routed/shared down_proj remains Q4TiledP; backbone/head matrices follow the Q8_2f policy, with required F16/F32 controls retained.
• q8_2f: model matrices use Q8_2f subject to the same required floating-point controls.

Why there is no q4tp-v2 file

q2tp-v2 names the current q2tp encoder revision. In contrast, a q4tp-v2 label is reserved here for a data-aware, calibrated/refit q4 workflow. The published 122B q4 artifact is a direct, non-data-calibrated conversion, so it is deliberately named q4tp, not q4tp-v2. This avoids misrepresenting a direct MoE conversion as a calibrated quality result.

Architecture retained by the text conversion

property	122B CMF text target
model family	Qwen3.5 MoE / Gated DeltaNet hybrid
layers	48 total: 36 Gated DeltaNet (linear-attention) + 12 full-attention
hidden / attention	hidden 3072; 32 query heads, 2 KV heads, head dimension 256
MoE	256 routed experts, top-8 selected per token, plus one always-on shared expert
MTP	one physical MTP block retained
CMF structure	38,384 tensors; embedded tokenizer; CMF v2
modality	text tower only; visual tensors are intentionally omitted

The architecture table describes what was preserved in the text conversion. It is not an upstream benchmark claim and does not validate vision, image/video preprocessing, tool execution, or an external hosted API.

Measured Vulkan run

This is a narrow, reproducible execution measurement, not a general throughput promise or quality benchmark.

• Host: Ubuntu 24.04, NVIDIA RTX PRO 4000 Blackwell (24,467 MiB reported), driver 595.91.07; WGPU/Vulkan.
• Runtime: cortiq 0.8.30 release build.
• Command shape: cortiq bench MODEL --tokens N --ignore-eos --core --json, CMF_GPU=wgpu, WGPU_BACKEND=vulkan, no MTP; one stream.
• The runtime used a safe device prefix and a host tail, rather than claiming the full 48-layer model was GPU-resident.

profile	generated tokens	steady decode tok/s	TTFT	device prefix	resident weight bytes
q2tp-v2	64	5.72	5.76 s	13 / 48 layers	13,075,174,400
q4tp	64	4.59	5.34 s	9 / 48 layers	12,925,678,592
q8_2f	32	1.04	20.23 s	5 / 48 layers	13,353,047,040

For every measured token the graph recorded a GPU prefix and no graph miss; it did not record a full-model graph. The q2/q4/q8 files also completed a bounded GPU-configured generation smoke. The q2tp-v2 profile is therefore the practical choice on this particular 24 GB Vulkan machine. Results depend on driver, free VRAM, host RAM/storage, prompt length, selected experts, and runtime version.

The same host was set up with the repository's Vulkan prerequisites:

apt-get update
apt-get install -y libglvnd0 libgl1 libglx0 libgles2 libegl1 libvulkan1 vulkan-tools
export XDG_RUNTIME_DIR=/tmp
vulkaninfo --summary

MTP result

The MTP tensor block is physically present and CMF integrity verification passed. On a CPU q4tp diagnostic, forcing CMF_MTP=1 drafted 7 and accepted 5 tokens in a 16-token core bench; the short greedy answer's emitted text matched the MTP-off run. That is a functional check, not a quality evaluation.

On the measured Vulkan prefix configuration, forced MTP was not beneficial: its 64-token q2/q4 runs drafted 15 tokens and accepted 0, with 5.36 / 4.12 steady tok/s respectively. Leave MTP disabled for this exact Vulkan profile unless it is re-measured on the target workload.

Limits and integrity

• cortiq verify succeeded for all three 122B files (38,384 tensors each). Each profile also loaded and generated a bounded CPU smoke response; that only demonstrates basic execution, not broad answer quality.
• No held-out perplexity, task-suite score, or comparison with the original BF16 checkpoint is published here. q8_2f is a CMF comparison profile, not an FP16 reference.
• The 122B file sizes are about 44.79 GiB (q2tp-v2), 63.24 GiB (q4tp), and 116.47 GiB (q8_2f) before operating-system and runtime headroom. They must not be presented as validated fits for a 24 GB unified-memory Mac.
• No Apple-silicon/Metal performance, capacity, or end-to-end quality result is claimed for these 122B files. Check cortiq gpu, available memory, and the complete target workload rather than extrapolating from Vulkan.
• Large-model runtime performance is route- and context-dependent. Do not raise undocumented GPU-memory limits to force a larger resident graph.

License and attribution

The Ornith 1.5 source checkpoints listed in this card are released under the MIT license. This repository provides quantized, text-only derivatives. CMF and cortiq are available from the CMF project.

---

Русская документация

Здесь находятся самостоятельные CMF-пакеты текстовой башни Ornith-1.5-122B для Rust-runtime cortiq. Файлы сконвертированы из ornith-ai/Ornith-1.5-122B, revision 7681ef608624e7532d8164128bcffdbdbaf1b1af.

Это только текстовая модель: visual tower исходного checkpoint намеренно не включена. Image/video input и исходное vision-language поведение не заявляются.

Требуется cortiq 0.8.30+.

Файлы 122B

файл	точный размер	SHA-256	назначение
ornith-1.5-122b-q2tp-v2.cmf	48,097,144,622 B	ebe80b9a6c112a7dd82fe31dca8058c7cc8e7fe03616f49e1431c1b8060ad01a	наименьший файл; самый быстрый из трёх на измеренном Vulkan-host с 24 GB
ornith-1.5-122b-q4tp.cmf	67,904,220,974 B	fb4706070790e4b3fe1eda81ec7c3fb997a8653334092f0b4e741b6f4d26c13a	прямой профиль q4tp с более высокой точностью матриц
ornith-1.5-122b-q8_2f.cmf	125,055,889,198 B	44f92bfdbeb0ca94f1097cb85fc1eb6bce1ba55f405b2bd558cab4f778a1c429	высокоточный CMF-профиль для сравнения, но не FP16/upstream ground truth

Это mixed-precision файлы, а не равномерно 2-/4-/8-битные веса:

• q4tp: routed и shared MoE-матрицы — Q4TiledP; backbone/head по гибридной политике используют Q8_2f, а нормы и управляющие тензоры остаются в нужном floating типе.
• q2tp-v2: gate_proj/up_proj routed и shared experts — Q2TiledP, down_proj — Q4TiledP, backbone/head — Q8_2f с сохранением F16/F32 controls.
• q8_2f: матрицы — Q8_2f, кроме обязательных floating controls.

Почему нет q4tp-v2

Суффикс q2tp-v2 означает текущую ревизию q2tp encoder. q4tp-v2 здесь зарезервирован для data-aware калибровки/refit. Опубликованный 122B q4tp сконвертирован напрямую и такой MoE-refit пока не проходил, поэтому файл честно называется q4tp, а не q4tp-v2.

Архитектура

Для 122B сохранены: 48 слоёв (36 Gated DeltaNet + 12 full-attention), hidden 3072, 32 Q-head / 2 KV-head, 256 routed experts с top-8 и один shared expert, а также один физический MTP block. В CMF 38,384 тензора и embedded tokenizer. Это описание text conversion, а не заявление о vision, качестве или benchmark upstream-модели.

Измеренный Vulkan

Ubuntu 24.04, NVIDIA RTX PRO 4000 Blackwell (24,467 MiB), driver 595.91.07, WGPU/Vulkan, cortiq 0.8.30, один stream, bench --core, MTP выключен:

профиль	токенов	steady decode	TTFT	GPU prefix
q2tp-v2	64	5.72 tok/s	5.76 s	13 / 48 слоёв
q4tp	64	4.59 tok/s	5.34 s	9 / 48 слоёв
q8_2f	32	1.04 tok/s	20.23 s	5 / 48 слоёв

Это безопасный device prefix с host tail; ни один результат не означает, что все 48 слоёв целиком resident в GPU. Для всех измеренных токенов был GPU prefix без graph miss. Скорость зависит от драйвера, VRAM, RAM/накопителя, длины контекста и выбранных экспертов. На этом host практичнее q2tp-v2.

MTP-физически присутствует и прошёл integrity-проверку. CPU-диагностика q4tp с CMF_MTP=1 получила 7 drafted / 5 accepted в 16-token bench, но на данном Vulkan prefix форсированный MTP дал 15 drafted / 0 accepted и снизил скорость. Для этой конфигурации оставляйте MTP выключенным, пока не появится измерение на вашей задаче.

Проверка и ограничения

REPO=infosave/Ornith-1.5-122B-cmf
MODEL=ornith-1.5-122b-q2tp-v2.cmf
hf download "$REPO" "$MODEL" "$MODEL.sha256" --local-dir .
sha256sum -c "$MODEL.sha256"                 # macOS: shasum -a 256 -c "$MODEL.sha256"
cortiq verify "$MODEL"

verify успешно прошёл для всех трёх 122B-файлов; ограниченный CPU smoke также загрузил и сгенерировал ответ. Это не является тестом качества. Пока не опубликованы PPL, task-suite или сравнение с исходным BF16 checkpoint. Размеры файлов — примерно 44.79 / 63.24 / 116.47 GiB ещё до headroom ОС и runtime, поэтому для 24 GB Mac нет заявления о Metal-fit, скорости или end-to-end качестве. Все перечисленные исходные Ornith checkpoints имеют MIT-лицензию.