---
type: "post"
id: "diffuspace-publishes-official-clover-35b-a3b-preview-gguf-6f368d"
slug: "diffuspace-publishes-official-clover-35b-a3b-preview-gguf-6f368d"
title: "DiffuSpace publishes official Clover-35B-A3B-Preview GGUF quantizations for llama.cpp self-speculative decoding"
description: "DiffuSpace has published official GGUF quantizations of its Clover-35B-A3B-Preview model for local use with llama.cpp."
retrieval_nugget: "DiffuSpace has published official GGUF quantizations of its Clover-35B-A3B-Preview model for local use with llama.cpp. The release makes smaller local files available, but the documented path depends on a llama.cpp pull request that is still unmerged, so builders should treat it as an experimental deployment route rather than a drop-in runtime update."
published_at: "2026-10-31"
updated_at: "2026-10-31"
record_date: "2026-10-31"
date_kind: "published_at"
topics: []
entities: ["DiffuSpace","Clover-35B-A3B-Preview","llama.cpp","GGUF"]
related_patterns: []
novelty_label: "new_fact"
what_changed: "DiffuSpace has published official Clover GGUF quantizations for llama.cpp self-speculative decoding, beyond the model release and earlier support pull request."
prior_refs: ["memory:d4085037-cd21-8a97-afdb-78fafc9650ea","memory:015eda38-675f-8b40-aa96-585622ee3d76"]
source_url: "https://huggingface.co/Diffuspace/Clover-35B-A3B-Preview-GGUF"
source_urls: ["https://huggingface.co/Diffuspace/Clover-35B-A3B-Preview-GGUF"]
schema_version: "newruntime-agent-readable-v0.2"
stable_id: "post:diffuspace-publishes-official-clover-35b-a3b-preview-gguf-6f368d"
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-publishes-official-clover-35b-a3b-preview-gguf-6f368d/","markdown":"https://newruntime.com/posts/diffuspace-publishes-official-clover-35b-a3b-preview-gguf-6f368d.md","json":"https://newruntime.com/posts/diffuspace-publishes-official-clover-35b-a3b-preview-gguf-6f368d.json"}
---

# DiffuSpace publishes official Clover-35B-A3B-Preview GGUF quantizations for llama.cpp self-speculative decoding

## Retrieval answer

DiffuSpace has published official GGUF quantizations of its Clover-35B-A3B-Preview model for local use with llama.cpp. The release makes smaller local files available, but the documented path depends on a llama.cpp pull request that is still unmerged, so builders should treat it as an experimental deployment route rather than a drop-in runtime update.

## Summary

DiffuSpace has published official GGUF quantizations of its Clover-35B-A3B-Preview model for local use with llama.cpp. The release makes smaller local files available, but the documented path depends on a llama.cpp pull request that is still unmerged, so builders should treat it as an experimental deployment route rather than a drop-in runtime update.

## What happened

DiffuSpace has opened an official Hugging Face repository, `Diffuspace/Clover-35B-A3B-Preview-GGUF`, for GGUF versions of Clover-35B-A3B-Preview. The repository is labelled for text generation and tagged with GGUF, llama.cpp, Clover, diffusion, mixture-of-experts, self-speculative decoding, and imatrix. It carries an Apache-2.0 licence. The repository was created on October 29, according to the supplied release material.

The repository lists three GGUF files: `Clover-35B-A3B-Preview-Q4_0.gguf`, `Clover-35B-A3B-Preview-Q4_K_M.gguf`, and `Clover-35B-A3B-Preview-Q8_0.gguf`. The supplied release material describes the Q4_0 file as 18.24 GiB, Q4_K_M as 21.42 GiB, and Q8_0 as 35.26 GiB. These are quantized versions of the previously released Clover-35B-A3B-Preview weights, intended for local inference rather than a hosted endpoint. The repository also presents a local-app route through llama.cpp rather than through a managed inference provider.

DiffuSpace’s model card provides a llama.cpp route and labels it as a local-app integration. Its instructions begin with installation on macOS or Linux and describe starting a local OpenAI-compatible server. That interface matters operationally: a local process can be presented to software that already expects an OpenAI-compatible server, without the source establishing that every compatible client or configuration has been tested.

The release material adds an important implementation constraint. Users need to build llama.cpp from pull request #31306, which adds the `clover_moe` architecture and was still unmerged when the repository was published. Ordinary released llama.cpp builds may therefore not yet be sufficient. The supplied material says the pull request opened on October 28, one day before the official GGUF repository was created. This is not simply a file-format download with a stable released runtime dependency.

The intended execution mode is described as draft-self decoding. Rather than pairing a target model with a separate draft model, Clover produces draft tokens through block diffusion and verifies them causally. That is the practical reason this repository is more than a conventional format conversion: the local runtime must understand both the model architecture and this decoding path. Builders evaluating it should distinguish the availability of the files from general availability of a supported llama.cpp implementation.

The model card also draws a boundary around performance claims. Reported benchmarks concern the original Clover model, not these particular GGUF quantizations. A builder considering the files therefore has evidence that official quantized artifacts exist and that a documented path is being prepared, but not a published comparison of quality, speed, memory use, or hardware compatibility across Q4_0, Q4_K_M, and Q8_0. Hugging Face shows 16,918 downloads over the previous month for this GGUF repository, compared with 2,021 for the BF16 weights repository, although those counts do not establish how many downloads became successful deployments.

## What the archive adds

This is a new fact on a known story. The earlier Field Note, [DiffuSpace releases Clover-35B-A3B-Preview, an open-weight 35B MoE diffusion language model](https://newruntime.com/posts/diffuspace-releases-clover-35b-a3b-preview-an-open-bd5943/), covered the model release, its 35B total and 3B active parameters, licence, and reported results. This update is specifically about official GGUF delivery and the llama.cpp path for running it locally.

## Trend signal

No supplied pattern is a close fit. The release is primarily a compatibility and packaging update: it makes a model available in a local-runtime format, but it does not by itself establish a durable change in agent architecture, interoperability, security, or economics.

## Why it matters

For teams evaluating open models locally, the release reduces one layer of work: they no longer need to rely on an unofficial conversion to obtain these three GGUF variants. Official publication also identifies the exact artifacts that DiffuSpace is presenting for local use, rather than leaving teams to infer provenance from community repositories.

It also makes the deployment dependency explicit. The useful question is not simply whether Clover can be downloaded, but whether a team can build and operate the required llama.cpp branch, use its self-speculative mode, and validate behavior on its own workload. A team that cannot own that dependency may have a downloadable model but not a practical deployment path.

That distinction matters because quantization changes the operational envelope. Smaller files can make local experiments more practical, while the runtime requirement may add build, upgrade, and reproducibility costs. The official repository is a clearer starting point for testing than community conversions, but it is not yet evidence that a production deployment will match the original model’s reported results or be straightforward to support. The release is most useful to builders able to evaluate a specific runtime build, not to teams seeking a conventional stable package.

## Limits and uncertainty

The supplied material does not provide a quality evaluation for Q4_0, Q4_K_M, or Q8_0 against the original weights, nor does it provide throughput, latency, context-length, memory, or hardware test results for the quantizations. It says the reported benchmarks are for the original model, not the GGUF files. It also does not establish when, or whether, pull request #31306 will merge into a standard llama.cpp release. Download counts are not measures of successful installations, production use, or model preference.

## Skill to build now

Set up a small reproducible local-inference evaluation before choosing a quantization. Record the exact GGUF filename, llama.cpp commit or pull-request build, command-line settings, machine configuration, and a fixed set of representative prompts. Compare output quality and completion behavior before measuring speed. Include at least one request representative of the application that would consume the local OpenAI-compatible endpoint.

This creates a useful baseline if the required llama.cpp support changes, and prevents a benchmark for the original weights from being mistaken for a result from a local quantized deployment. It also gives the team an auditable way to decide whether the smaller Q4 files deliver an acceptable trade-off for its own task.

## Reader takeaway

Official Clover GGUF files are now available in three sizes, giving local-model builders a concrete starting point. But the release should be read as an experimental runtime integration: it requires a llama.cpp build with unmerged Clover support, and it offers no published quantization-specific performance or quality comparison yet.

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)