Apex-35B-V1.0 — CMF

Apex Intelligence Apex-35B-V1.0 — an agentic research, coding and reasoning model fine-tuned from Qwen3.6-35B-A3B (mixture of experts: 35.5B parameters, ~3B active per token, 256 routed experts top-8 plus a shared expert; 30 Gated DeltaNet and 10 gated full-attention layers; 262,144-token context; dynamic thinking; tool calling) — packaged as single CMF files. They run with cortiq, a Rust inference engine with no Python and no ML framework: NVIDIA, AMD and Intel GPUs through Vulkan/DX12, Apple silicon through Metal, and the CPU. The files carry the text model and its MTP layer; the checkpoint's vision tower is not included.

English · Русский · 中文

Quick start

cargo install cortiq-cli
hf download infosave/Apex-35B-V1.0-cmf apex-35b-v1.0-q4tp-v2.cmf --local-dir .
cortiq run apex-35b-v1.0-q4tp-v2.cmf --prompt "Write a Python function that merges two sorted lists." --max-tokens 4096

Use cortiq 0.8.48 or later; prebuilt binaries for Linux, macOS and Windows are on the releases page. The engine picks the GPU and its settings itself. Apex decides on each turn whether to think and for how long, and the thinking counts toward --max-tokens (default 256), so give it room; --no-think asks for a direct answer.

Files

Attention, Gated DeltaNet projections, embeddings and the output head are 8-bit in every file; norms and the router stay 16-bit. q4tp-v2 encodes every routed and shared expert with GPTQ against 262,144 calibration tokens (wikitext-2 train, UltraChat rendered in the model's template, Rust sources) at the size of q4tp; per position it stays closer to the original on chat (mean |ΔNLL| 0.138 against 0.160) and wiki text, and level on code.

Perplexity over the first 4096 tokens of wikitext-2 test and of Python standard-library code, and over 8192 tokens of 12 held-out UltraChat conversations in the chat template; top-1 vs bf16: positions of the chat text where the file's top prediction matches the bf16 checkpoint run in PyTorch over the same tokens. The fine-tune itself scores 26.5 on raw wikitext (Qwen3.6-35B-A3B: 5.6): the number belongs to the model, not to the quantization.

Performance

cortiq bench --core --ignore-eos, one stream, cortiq 0.8.48, NVIDIA A40 (48 GB, Vulkan), the whole model on the card:

Resident weights: 18.7 GB (q4tp, q4tp-v2), 34.3 GB (q8_2f). The q8_2f file reads the prompt one position at a time on Vulkan; for long prompts use a 4-bit file.

Usage

cortiq run apex-35b-v1.0-q4tp-v2.cmf                                  # interactive chat
cortiq run apex-35b-v1.0-q4tp-v2.cmf --prompt "..." --max-tokens 8192 # one answer, thinking as needed
cortiq run apex-35b-v1.0-q4tp-v2.cmf --prompt "..." --no-think        # direct answer
cortiq run apex-35b-v1.0-q4tp-v2.cmf --prompt "..." --greedy          # deterministic output

The files carry Apex's recommended sampling — temperature 1.0, top-p 0.95, top-k 20 — as the run and serve defaults; flags and request fields override it. The embedded chat template is the checkpoint's own (the generation prompt ends with <think>).

OpenAI-compatible server

cortiq serve apex-35b-v1.0-q4tp-v2.cmf --port 8080

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "apex", "max_tokens": 4096,
       "messages": [{"role": "user", "content": "Explain how a B-tree splits a full node."}]}'

Endpoints: /v1/chat/completions (streamed or not), /v1/completions, /v1/models, plus a web dashboard on the same port. --host 127.0.0.1 keeps the server local. A thinking reply reads <think>…</think>answer; "enable_thinking": false gives a direct answer.

Tool calling

Send tools in the OpenAI format; the template lists them the way the model was trained, and the model answers with <tool_call><function=NAME><parameter=KEY>value</parameter></function></tool_call>. The server returns that as message.tool_calls with finish_reason: "tool_calls". In the test conversation (a weather tool, then the answer from its result) the call and the answer were correct with and without thinking.

Hardware

NVIDIA / AMD / Intel (Vulkan, DX12). The 4-bit files run whole on a 24 GB card; a smaller card keeps as many leading layers as fit and the CPU runs the rest. cortiq gpu lists the adapters; CMF_GPU_ADAPTER=1 (index or name substring) picks one. On headless Linux install libvulkan1 libglvnd0 libegl1 libgl1 libglx0 and set XDG_RUNTIME_DIR=/tmp.

Apple silicon. The native Metal backend is the default on macOS.

CPU only. CMF_GPU=0.

Long context

The model supports 262,144 tokens; the KV cache holds 32,768 by default. CMF_MAX_SEQ=262144 raises it. The ten attention layers keep about 40 KB of KV per token (1.3 GB at 32k, 10.5 GB at 262k); the Gated DeltaNet layers keep a fixed state. At a 65,536-token prompt (CMF_MAX_SEQ=131072, A40, q4tp-v2) the prompt reads at 254 tok/s and decoding runs at 28 tok/s.

Verify the download

sha256sum -c apex-35b-v1.0-q4tp-v2.cmf.sha256
cortiq verify apex-35b-v1.0-q4tp-v2.cmf
cortiq info apex-35b-v1.0-q4tp-v2.cmf

Reproduce

cortiq convert --model Apexintelligence-AI/Apex-35B-V1.0 --quant q4tp \
  --output apex-35b-v1.0-q4tp.cmf --also-output q8_2f=apex-35b-v1.0-q8_2f.cmf \
  --sampling temperature=1.0,top_p=0.95,top_k=20,min_p=0,repetition_penalty=1.0
cortiq quantize-moe apex-35b-v1.0-q4tp.cmf --hf Apexintelligence-AI/Apex-35B-V1.0 \
  --calib wiki.train.raw=0.4 --calib chat.txt=0.4 --calib code.txt=0.2 \
  --tokens 262144 --codec q4tp --output apex-35b-v1.0-q4tp-v2.cmf

The conversion streams the checkpoint (peak disk: the outputs); quantize-moe fetches the bf16 experts one layer at a time. The scripts used for these files (reference perplexity, chat parity, benchmarks) are in scripts/apex.

Weights derive from Apex Intelligence's release and remain under its Apache-2.0 terms; see the upstream card.

---

Документация на русском

Apex-35B-V1.0 от Apex Intelligence — модель для агентной работы, исследований, кода и рассуждений, дообученная из Qwen3.6-35B-A3B (смесь экспертов: 35.5 млрд параметров, ~3 млрд активных на токен, 256 экспертов с выбором top-8 и общий эксперт; 30 слоёв Gated DeltaNet и 10 слоёв полного внимания; контекст 262 144 токена; рассуждения по необходимости; вызов инструментов), в виде отдельных файлов CMF для движка cortiq на Rust без Python и ML-фреймворков. В файлах текстовая модель и её MTP-слой; визуальной башни чекпойнта в них нет.

Быстрый старт

cargo install cortiq-cli
hf download infosave/Apex-35B-V1.0-cmf apex-35b-v1.0-q4tp-v2.cmf --local-dir .
cortiq run apex-35b-v1.0-q4tp-v2.cmf --prompt "Напиши функцию на Python, которая сливает два отсортированных списка." --max-tokens 4096

Нужен cortiq 0.8.48 или новее. Модель сама решает, рассуждать ли и сколько; рассуждение входит в --max-tokens (по умолчанию 256), поэтому дайте запас; --no-think просит ответ сразу.

Файлы

Высокий ppl на wikitext — свойство самого дообучения (bf16-оригинал даёт 26.5, Qwen3.6-35B-A3B — 5.6), а не квантования.

Скорость

NVIDIA A40 (48 ГБ, Vulkan), cortiq 0.8.48, модель целиком на карте:

Файл q8_2f читает промпт по одной позиции; для длинных промптов берите 4-битный файл. Сэмплирование по умолчанию — рекомендованное Apex (temperature 1.0, top-p 0.95, top-k 20). cortiq serve даёт OpenAI-совместимый API с вызовом инструментов (message.tool_calls). Контекст до 262 144 токенов: CMF_MAX_SEQ=262144 (по умолчанию 32 768).

中文说明

Apex-35B-V1.0 是 Apex Intelligence 基于 Qwen3.6-35B-A3B 微调的智能体、研究、编程与推理模型 （混合专家：35.5B 参数，每个 token 约 3B 激活，256 个路由专家 top-8 加共享专家； 30 层 Gated DeltaNet 与 10 层门控全注意力；262,144 token 上下文；按需思考； 工具调用）。本仓库提供可由 Rust 推理引擎 cortiq（0.8.48 及以上）加载的 CMF 文件，包含文本模型及其 MTP 层，不含视觉塔。

cargo install cortiq-cli
hf download infosave/Apex-35B-V1.0-cmf apex-35b-v1.0-q4tp-v2.cmf --local-dir .
cortiq run apex-35b-v1.0-q4tp-v2.cmf --prompt "你好" --max-tokens 4096

bf16 原模型在 wikitext 上为 26.5（Qwen3.6-35B-A3B 为 5.6），较高的 wiki ppl 来自 微调本身而非量化。NVIDIA A40（48 GB，Vulkan）上 q4tp-v2 生成 80.3 token/s， 4k 提示词读取 390 token/s，首 token 10.5 s；q8_2f 逐位置读取提示词（84 token/s）， 长提示词请使用 4 位文件。默认采样为 Apex 推荐的 temperature 1.0、top-p 0.95、 top-k 20。cortiq serve 提供 OpenAI 兼容接口与工具调用。