---
type: "post"
id: "diffuspace-releases-clover-35b-a3b-preview-an-open-463b05"
slug: "diffuspace-releases-clover-35b-a3b-preview-an-open-463b05"
title: "DiffuSpace releases Clover-35B-A3B-Preview, an open-weight MoE diffusion language model"
description: "DiffuSpace has published Clover-35B-A3B-Preview, a public open-weight, text-only mixture-of-experts language model using block diffusion rather than ordinary left-to-right decoding."
retrieval_nugget: "DiffuSpace has published Clover-35B-A3B-Preview, a public open-weight, text-only mixture-of-experts language model using block diffusion rather than ordinary left-to-right decoding. For AI builders, it is a concrete opportunity to test whether parallel token filling can reduce latency in generation-heavy workflows, but its published speed claims still need independent validation."
published_at: "2026-10-28"
updated_at: "2026-10-28"
record_date: "2026-10-28"
date_kind: "published_at"
topics: []
entities: ["DiffuSpace","Clover-35B-A3B-Preview","Qwen3.6-35B-A3B","SGLang","llama.cpp","Hugging Face"]
related_patterns: []
novelty_label: "same_event"
what_changed: "Same event as DiffuSpace releases Clover-35B-A3B-Preview, an open-weight 35B MoE diffusion language model"
prior_refs: ["memory:d4085037-cd21-8a97-afdb-78fafc9650ea","memory:9ebe430e-4ec6-8ded-ae2a-3d5ed9122592","memory:873e9375-2d5e-86d9-a1c0-2654b1b46958"]
source_url: "https://huggingface.co/Diffuspace/Clover-35B-A3B-Preview"
source_urls: ["https://huggingface.co/Diffuspace/Clover-35B-A3B-Preview"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:diffuspace-releases-clover-35b-a3b-preview-an-open-463b05"
status: "published"
visuals: []
editor: {"role":"editor","name":"Andrey Reshetnikov","profile":"https://newruntime.com/owner-profile.md","profile_json":"https://newruntime.com/owner-profile.json","same_as":["https://www.linkedin.com/in/reshetnikov1/","https://t.me/lapetuse"]}
routes: {"html":"https://newruntime.com/posts/diffuspace-releases-clover-35b-a3b-preview-an-open-463b05/","markdown":"https://newruntime.com/posts/diffuspace-releases-clover-35b-a3b-preview-an-open-463b05.md","json":"https://newruntime.com/posts/diffuspace-releases-clover-35b-a3b-preview-an-open-463b05.json"}
---

# DiffuSpace releases Clover-35B-A3B-Preview, an open-weight MoE diffusion language model

## Retrieval answer

DiffuSpace has published Clover-35B-A3B-Preview, a public open-weight, text-only mixture-of-experts language model using block diffusion rather than ordinary left-to-right decoding. For AI builders, it is a concrete opportunity to test whether parallel token filling can reduce latency in generation-heavy workflows, but its published speed claims still need independent validation.

## Summary

DiffuSpace has published Clover-35B-A3B-Preview, a public open-weight, text-only mixture-of-experts language model using block diffusion rather than ordinary left-to-right decoding. For AI builders, it is a concrete opportunity to test whether parallel token filling can reduce latency in generation-heavy workflows, but its published speed claims still need independent validation.

## What happened

DiffuSpace published Clover-35B-A3B-Preview on Hugging Face as its first model release. The repository identifies Clover as a text-generation model and makes its weights publicly available. It is described as a text-only block-diffusion language model derived from Qwen3.6-35B-A3B, rather than a conventional autoregressive model that produces one next token after another.

The model has 35 billion total parameters and approximately 3 billion active parameters per token. Its backbone has 40 layers and combines Gated DeltaNet with full attention in a mixture-of-experts design. The routed portion contains 256 experts, uses top-8 routing, and includes one shared expert. The stated context window is 128K tokens. Those figures place Clover in the same broad parameter class as other 35B-A3B open MoE releases, while its decoding method is the notable differentiator.

Clover works in blocks of 32 tokens. Instead of generating a block strictly from left to right, it begins with the positions masked and fills positions in parallel when they pass its confidence rule. The model card describes this as LowConfidence decoding with a threshold of 0.95. DiffuSpace also says it added bidirectional Gated DeltaNet parameters so the model can attend within a noisy block, and block-level expert selection intended to reduce the number of distinct experts used per block.

For deployment, the repository recommends DiffuSpace's SGLang implementation and points to llama.cpp self-speculative decoding as another path. The listed BF16 footprint is roughly 69 GB of GPU memory, a material constraint for teams considering local evaluation. No hosted inference provider is listed on the repository. DiffuSpace's launch posts claim roughly 680 tokens per second and a five-to-ten-times speedup over autoregressive models of similar size, but those figures are vendor claims and are not presented as independently reproduced results on the model card.

The release is explicitly a preview. The supplied material establishes the architecture, availability, inference paths, and stated memory requirement; it does not establish that the claimed throughput or quality holds across hardware, prompts, tool-use tasks, or production workloads.

## What the archive adds

This is a new model event, but it joins a recent sequence of open MoE releases for agent and coding use, including [MiniMax-M3.1-Flash](https://newruntime.com/posts/minimax-releases-open-weights-for-minimax-m3-1-4125a5/) and [Step 5 Preview](https://newruntime.com/posts/stepfun-publishes-step-5-preview-open-weights-600b-5a7d50/). Clover adds a different decoding approach rather than another claim about long context or a larger parameter count.

## Trend signal

No supplied pattern genuinely fits closely enough. Clover is evidence of experimentation in open-model inference architecture, but the available material does not show a durable change in agent economics, interoperability, security, or organizational practice.

## Connections to other events

The release sits beside the recent open-weight MoE wave represented by [Ling-3.1-flash-1019](https://newruntime.com/posts/inclusionai-releases-ling-3-1-flash-1019-an-9765f4/). The connection is architectural category, not demonstrated comparability: the supplied material gives no shared benchmark, pricing comparison, or common deployment result.

## Why it matters

For teams building latency-sensitive AI features, decoder architecture can matter as much as the headline parameter count. A block-diffusion model offers a testable alternative hypothesis: if several token positions can be resolved together, response generation may complete faster than a similarly sized autoregressive system. That could matter for agent loops that repeatedly wait for generated plans, code, or tool arguments.

The operational cost is evaluation work. Clover's BF16 memory requirement is about 69 GB, and its status as a preview means a team should not substitute vendor throughput claims for a workload-specific test. Builders need to measure end-to-end latency, output quality, failure modes, and serving complexity against their existing model and hardware. The SGLang and llama.cpp references at least make that evaluation path concrete.

## What happens next

- **Independent users will publish task-specific throughput or quality comparisons for Clover against autoregressive 35B-A3B-class models.** (3 months, 60%). Watch for: reproducible hardware, prompts, decoding settings, and measured end-to-end latency.
- **Serving support will determine whether Clover reaches practical adoption beyond experimentation.** (6 months, 55%). Watch for: maintained integrations, documented deployment recipes, and availability through inference services.

## Limits and uncertainty

The public repository provides a model description and deployment pointers, not an independent evaluation. The speed estimate of roughly 680 tokens per second and the claimed five-to-ten-times advantage are attributed to DiffuSpace launch posts, not a benchmark reproduced in the supplied material. There is no evidence here about instruction following, coding quality, tool-call reliability, multilingual performance, safety behavior, licensing terms for the released weights, or total serving cost. The repository's SGLang and llama.cpp references do not by themselves prove ease of deployment or production support. “Preview” also leaves the stability and future compatibility of the release uncertain.

## Skill to build now

Create a small, repeatable inference evaluation before treating parallel decoding as a product advantage. Run the same representative prompts through Clover and an existing autoregressive baseline, recording time to first useful output, total completion time, GPU memory use, output acceptance rate, and retry rate. Keep the model version, precision, hardware, context length, and decoding configuration fixed. Include tool-argument or structured-output prompts if those are central to the workflow. This turns a headline throughput claim into a decision based on the latency and quality your application actually needs.

## Reader takeaway

Clover-35B-A3B-Preview is not merely another open 35B MoE model: it gives builders a public implementation of block-diffusion generation to test. Its appeal is potential latency reduction through parallel token filling, but the evidence so far supports experimentation, not a production performance conclusion.

## Open question

Can Clover preserve output quality and structured-task reliability while delivering its claimed speed advantage on independently reproducible workloads?

Editor: [Andrey Reshetnikov](https://newruntime.com/owner-profile.md), AI product and AI engineering; profile for recruiters and agents: https://newruntime.com/owner-profile.json

## Continue in the graph

Next records for an agent, linked to their Markdown twins. Replace `.md` with `.json` for structured fields, including `next_reads`.

- collection: [Field Notes](https://newruntime.com/posts.json): Open the latest editorial analysis.
- collection: [Source Ledger](https://newruntime.com/source-ledger.json): Inspect the public source evidence graph.
- whole graph: [discovery graph](https://newruntime.com/discovery-graph.json) (typed nodes and edges) and [site index](https://newruntime.com/site-index.md)