Clover-35B-A3B-Preview: Specs, License & How to Run

Clover-35B-A3B-Preview is a text-only diffusion language MoE from DiffuSpace. Official hosts we opened on 2026-10-31: the ungated card Diffuspace/Clover-35B-A3B-Preview (created 2026-10-24T03:46:30Z; last modified 2026-10-28T03:07:54Z; 176 likes / 1,505 downloads), the Apache 2.0 LICENSE, config.json, the inference repo DiffuSpace/Clover-35B-A3B-Preview (117 stars / 5 forks; created 2026-10-27T11:30:09Z; last push 2026-10-28T10:34:42Z; no LICENSE blob), the launch post Clover-35B, and the upstream card Qwen/Qwen3.6-35B-A3B. API gated was false. Hugging Face pipeline_tag is text-generation; model_type is clover_moe.

This page hangs off ChatGPT alternatives and closed LLM APIs. That mapping is editorial. We have no sourced quality comparison against ChatGPT. Head-to-head facts versus the same-size Apache-2.0 coding MoE already on this site: Clover-35B-A3B-Preview vs Flint. There is no Qwen3.6-35B-A3B page here.

The card calls Clover the first DiffuSpace release, combining a 35B-parameter mixture-of-experts backbone with approximately 3B active parameters per token, block diffusion, and a 128K context window. Those sentences stay vendor. Chat-template comments say the model accepts text content only.

What we verified (2026-10-31)

Field	Value	Source
Host	huggingface.co/Diffuspace/Clover-35B-A3B-Preview (not gated)	HF API
Created / last modified	2026-10-24T03:46:30Z / 2026-10-28T03:07:54Z	same
Likes / downloads	176 / 1,505	same
License tag + file	apache-2.0 / Apache License 2.0	HF cardData + LICENSE
Pipeline tag	text-generation	HF API
Architecture	clover_moe / CloverMoeModelLM	card + config
Base model	Qwen/Qwen3.6-35B-A3B	card YAML
Params (card)	35B total / ~3B active	card
safetensors.parameters	34,661,350,848	HF API
usedStorage (BF16)	69,342,784,805 bytes (~64.6 GiB)	same
Layers / experts	40 layers; 256 routed, top-8; 1 shared	card + config
Context	128K (card); 131,072 (config)	card + config
Block diffusion defaults	block_size 32; LowConfidence threshold 0.95	card + config
GitHub inference repo	117 stars / 5 forks; no LICENSE file	GitHub API

HF tags include base_model:Qwen/Qwen3.6-35B-A3B and base_model:finetune:Qwen/Qwen3.6-35B-A3B. Config also records moe_block_coreset_enabled: true and moe_block_coreset_top_p: 0.9 (card: 90% of aggregated routing mass, not 90% of all experts).

How to run (from the official card, not tested)

The card asks for Transformers 5.19.0, PyTorch 2.10 or later, trust_remote_code=True, and a GPU that can hold BF16 weights (approximately 69 GB) plus inference state. Copied install and generate names only:

pip install "transformers==5.19.0" accelerate safetensors

import torch
from transformers import AutoModelForCausalLM, AutoTokenizer

model_id = "Diffuspace/Clover-35B-A3B-Preview"
tokenizer = AutoTokenizer.from_pretrained(model_id, trust_remote_code=True)
model = AutoModelForCausalLM.from_pretrained(
    model_id, trust_remote_code=True, dtype=torch.bfloat16, device_map="cuda:0"
).eval()

Card generation defaults copied as labels: block_size=32, threshold=0.95, do_sample=False. The card says requests that exceed the 128K limit (including space to resolve the final full block) are rejected. Official SGLang in DiffuSpace/Clover-35B-A3B-Preview is the recommended serve/eval path; the card warns that the Hugging Face PyTorch path can produce different text than SGLang kernels. forkalt has not run either.

License

LICENSE is Apache License 2.0. Card tag apache-2.0. Appendix copyright on that file: Copyright 2026 Alibaba Cloud. The card’s “Upstream attribution” section says Clover is derived from Qwen3.6-35B-A3B and that the LICENSE preserves the upstream grant. The GitHub inference repo had no LICENSE file (404) on 2026-10-31. Not legal advice.

What we did not invent

• Numeric agentic-bench scores stay on the launch blog. They are not copied here.
• VRAM, tokens/s, and “state-of-the-art” stay vendor.
• We do not rank Clover against Flint, ChatGPT, or Qwen3.6-35B-A3B.

Related on this site: Flint — Clover-35B-A3B-Preview vs Flint — Ornith-1.5-122B — Sage-A1.5 — Nemotron 3.5 Bolt (local-agent VLM, different job) — LFM3. Serving: vLLM. Catalog: home.