infosave/Clover-35B-A3B-Preview-cmf

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 writes the answer in 32-token blocks: each block starts as mask tokens and is resolved by repeated denoising passes over the whole block, several tokens at once — packaged as single CMF files for cortiq, a Rust inference engine with no Python and no ML framework.

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 0.8.41 or later (block diffusion, Metal); 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

file	experts	size	wiki ppl	first-pass top-1	recommended for
clover-35b-a3b-preview-q4tp-v2.cmf	4-bit, GPTQ	19.36 GB	16.80	93/96	default: GPUs with 24 GB+
clover-35b-a3b-preview-q4tp.cmf	4-bit	19.38 GB	18.22	93/96	the plain conversion
clover-35b-a3b-preview-q8_2f.cmf	8-bit	35.00 GB	15.86	62/64	near-lossless; GPUs with 48 GB+, big-memory CPUs

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

A pass is one denoising forward over the 32-row block. LowConfidence fixes every row above the threshold after each pass (the release default 0.95), so predictable text resolves several tokens a pass and free prose about one. Tokens per second after the prompt, q4tp-v2, cortiq 0.8.41:

hardware	backend	pass	prose	code
RTX 5090 (32 GB)	Vulkan	34 ms	26 tok/s	80–95 tok/s
Mac mini M4 (24 GB)	Metal	0.3–0.5 s	3.2 tok/s	6.5–9 tok/s

On the RTX 5090 the whole model and its head run on the card (18.8 GB). A lower threshold trades care for speed: CMF_CLOVER_THRESHOLD=0.8 measured prose 25–30 and code 124–150 tok/s on the 5090, 3.7 and 9.7 tok/s on the M4. cortiq run prints the passes, tokens per pass and the rate after every answer.

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

Generation follows the release: greedy LowConfidence over 32-token blocks, the block's experts restricted to the set that carries 90% of its routing, and 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). Knobs: CMF_CLOVER_THRESHOLD (0–1), CMF_CLOVER_BLOCK (8, 16 or 32). --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. Streaming delivers each block as it resolves.

Hardware

NVIDIA / AMD / Intel (Vulkan, DX12). A card that holds the file runs every pass as one GPU submission; a smaller one runs the first layers on the card and the rest on the CPU (a 16 GB budget held 30 of 40 layers). cortiq gpu lists the adapters; on headless Linux install libvulkan1 libglvnd0 libegl1 libgl1 libglx0 and set XDG_RUNTIME_DIR=/tmp.

Apple silicon. The passes, block commits and the prompt run on Metal; the weights stay in the file's mapping (the experts a block uses are bound as they are routed), so a 24 GB Mac runs the q4tp files. The prompt reads at ~57 tok/s on an M4.

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

LowConfidence block diffusion, the release's default decoding. The self-speculative mode of DiffuSpace's SGLang and llama.cpp backends is not part of this runtime. 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). Ответ пишется блоками по 32 токена: блок стартует из масок и раскрывается повторными проходами денойзинга, по нескольку токенов за проход. Здесь она в виде отдельных файлов CMF для движка cortiq на Rust.

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 0.8.41 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). На RTX 5090 проход по блоку занимает 34 мс, модель целиком на карте (18.8 ГБ): проза ~26 ток/с, код 80–95 ток/с при пороге 0.95; CMF_CLOVER_THRESHOLD=0.8 ускоряет до 25–30 и 124–150 ток/с. На Mac mini M4 (24 ГБ) проходы, коммиты блоков и промпт идут на Metal: проза 3.2 ток/с, код 6.5–9 ток/с. 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）。模型以 32 个 token 为一块生成：每块从掩码开始， 经多次去噪一次确定多个 token。本仓库提供可由 Rust 推理引擎 cortiq（0.8.41 及以上） 加载的 CMF 文件；默认文件为 q4tp-v2（专家 4 位 GPTQ，19.36 GB）。RTX 5090 上 每次去噪 34 ms，整模型驻留显存（18.8 GB）：阈值 0.95 时文本约 26 token/s， 代码 80–95 token/s。Mac mini M4（24 GB）走 Metal：文本 3.2 token/s，代码 6.5–9 token/s。

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