Clover-35B-A3B-Preview — CMF

DiffuSpace Clover-35B-A3B-Preview — a diffusion language model on the Qwen3.6-35B-A3B backbone (mixture of experts: 35B parameters, ~3B active per token, 256 experts top-8 plus a shared expert; 30 Gated DeltaNet and 10 gated full-attention layers; 128k context). It is built to write in 32-token blocks by repeated denoising passes, and its trunk also reads as an ordinary causal model — packaged as single CMF files for cortiq, a Rust inference engine with no Python and no ML framework. cortiq decodes causally by default, the fastest on every machine we measured; block denoising and self-speculation are one setting away.

English · Русский · 中文

Quick start

cargo install cortiq-cli
hf download infosave/Clover-35B-A3B-Preview-cmf clover-35b-a3b-preview-q4tp-v2.cmf --local-dir .
cortiq run clover-35b-a3b-preview-q4tp-v2.cmf --prompt "Write a Python function that merges two sorted lists." --no-think --max-tokens 512

Use cortiq 0.8.52 or later; prebuilt binaries for Linux, macOS and Windows are on the releases page. Without --no-think the model reasons in a <think> block first, so give it --max-tokens in the thousands.

Files

Attention, Gated DeltaNet projections, embeddings and the head are 8-bit in every file. q4tp-v2 calibrates the experts with GPTQ on 262k tokens (wiki, chat and code text) at the same size as q4tp. wiki ppl: perplexity of the model's causal path over 12 × 512 tokens of wikitext-2 test. first-pass top-1: rows of the first denoising pass whose top token matches the official PyTorch implementation (bf16), three chat prompts × 32 rows (two for q8_2f); the misses are near-ties.

Performance

Tokens per second after the prompt, q4tp-v2:

The whole model and its head take 18.7 GB on the card. q8_2f does not fit 32 GB whole and runs at 46–49 tok/s there.

Decoding modes

CMF_CLOVER_MODE picks how a request uses the model. On the RTX 5090:

In denoise on an M4: 3.2 and 6.5–9 tok/s. CMF_CLOVER_THRESHOLD lowers the denoising threshold, CMF_CLOVER_DRAFT sets the rows of a self-speculative round.

Usage

cortiq run clover-35b-a3b-preview-q4tp-v2.cmf                                  # interactive chat
cortiq run clover-35b-a3b-preview-q4tp-v2.cmf --prompt "..." --max-tokens 4096 # with reasoning
cortiq run clover-35b-a3b-preview-q4tp-v2.cmf --prompt "..." --no-think        # direct answer

Every mode applies the release's response penalties (presence 0.9, repetition 1.05, a gradual push toward the stop token after 8192 answer tokens and toward </think> after 4096). --temperature samples instead of greedy.

OpenAI-compatible server

cortiq serve clover-35b-a3b-preview-q4tp-v2.cmf --port 8080

curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{"model": "clover", "enable_thinking": false,
       "messages": [{"role": "user", "content": "Explain how a hash map works."}]}'

/v1/chat/completions (streamed or not), /v1/completions, /v1/models and a web dashboard on the same port.

Hardware

NVIDIA / AMD / Intel (Vulkan, DX12). A card that holds the file runs the whole model; a smaller one runs the first layers on the card and the rest on the CPU. cortiq gpu lists the adapters; on headless Linux install libvulkan1 libglvnd0 libegl1 libgl1 libglx0 and set XDG_RUNTIME_DIR=/tmp.

Apple silicon. The prompt runs on Metal and the decode on the CPU, which is the faster of the two for this file on a 24 GB Mac: Metal cannot keep the 19 GB of weights resident there. The weights stay in the file's mapping, so a 24 GB Mac runs the q4tp files.

CPU only. CMF_GPU=0.

Verify the download

sha256sum -c clover-35b-a3b-preview-q4tp-v2.cmf.sha256
cortiq verify clover-35b-a3b-preview-q4tp-v2.cmf
cortiq info clover-35b-a3b-preview-q4tp-v2.cmf

Reproduce

cortiq convert --model Diffuspace/Clover-35B-A3B-Preview --quant q4tp \
  --sampling temperature=0,top_p=1,top_k=0,min_p=0,repetition_penalty=1 \
  --output clover-35b-a3b-preview-q4tp.cmf
cortiq quantize-moe clover-35b-a3b-preview-q4tp.cmf --hf Diffuspace/Clover-35B-A3B-Preview \
  --calib wiki.train.txt=0.4 --calib chat.txt=0.4 --calib code.txt=0.2 \
  --tokens 262144 --codec q4tp --output clover-35b-a3b-preview-q4tp-v2.cmf

Scope

Causal decoding, LowConfidence block diffusion and self-speculation. Weights derive from DiffuSpace's release (built on Qwen3.6-35B-A3B) and remain under its Apache-2.0 terms; see the upstream card and its attributions.

---

Документация на русском

Clover-35B-A3B-Preview от DiffuSpace — диффузионная языковая модель на основе Qwen3.6-35B-A3B (35 млрд параметров, ~3 млрд активных на токен). Она умеет писать блоками по 32 токена, раскрывая их проходами денойзинга, а её основа работает и как обычная каузальная модель. Здесь она в виде отдельных файлов CMF для движка cortiq на Rust; по умолчанию cortiq декодирует каузально: это быстрее всего на всех машинах, где мы мерили.

cargo install cortiq-cli
hf download infosave/Clover-35B-A3B-Preview-cmf clover-35b-a3b-preview-q4tp-v2.cmf --local-dir .
cortiq run clover-35b-a3b-preview-q4tp-v2.cmf --prompt "Объясни, как работает хеш-таблица." --no-think --max-tokens 512

Нужен cortiq 0.8.52 или новее. Файл по умолчанию — q4tp-v2 (эксперты 4 бита с GPTQ, 19.36 ГБ, wiki ppl 16.80); q4tp — простая конвертация (18.22), q8_2f — почти без потерь (35 ГБ, 15.86).

Токены в секунду после промпта, q4tp-v2: RTX 5090 (Vulkan) — 120 на прозе и 121 на коде, промпт на 2048 токенов читается со скоростью 1143 ток/с; при 16 ГБ видеопамяти — 54 и 53; Mac mini M4 (24 ГБ, промпт на Metal ~170 ток/с, декод на CPU) — 13–18 и 14–18; только CPU, 15 ядер Xeon — 13.8 и 10.9.

CMF_CLOVER_MODE выбирает режим: ar (по умолчанию) — каузальный декод со штрафами из релиза; draft — самоспекуляция, проход денойзинга предлагает 16 токенов, каузальный путь их проверяет (RTX 5090: 42 и 104 ток/с); denoise — блочный денойзинг LowConfidence из релиза (RTX 5090: 18–26 и 63–95; M4: 3.2 и 6.5–9). cortiq serve даёт OpenAI-совместимый API с потоковой выдачей.

中文说明

Clover-35B-A3B-Preview 是 DiffuSpace 基于 Qwen3.6-35B-A3B 的扩散语言模型（35B 参数，每个 token 约 3B 激活）。它可以用 32 个 token 的块经多次去噪生成，其主干也可作 普通因果模型运行。本仓库提供可由 Rust 推理引擎 cortiq（0.8.52 及以上）加载的 CMF 文件；默认文件为 q4tp-v2（专家 4 位 GPTQ，19.36 GB）。cortiq 默认因果解码，这在 我们测过的所有机器上都最快。

每秒 token 数（提示之后，q4tp-v2）：RTX 5090（Vulkan）文本 120、代码 121，2048 个 token 的提示 1143 token/s；16 GB 显存 54 和 53；Mac mini M4（24 GB，提示走 Metal 约 170 token/s，解码走 CPU）13–18 和 14–18；仅 CPU（15 核 Xeon）13.8 和 10.9。

CMF_CLOVER_MODE 选择模式：ar（默认）为因果解码；draft 为自推测，一次去噪提出 16 个 token，由因果路径验证（RTX 5090：42 和 104）；denoise 为发布版的 LowConfidence 块去噪（RTX 5090：18–26 和 63–95；M4：3.2 和 6.5–9）。

hf download infosave/Clover-35B-A3B-Preview-cmf clover-35b-a3b-preview-q4tp-v2.cmf --local-dir .
cortiq run clover-35b-a3b-preview-q4tp-v2.cmf --prompt "你好" --no-think