Can I Run Clover 35B-A3B Locally? (2026): 24GB, No Diffusion
DiffuSpace put Clover-35B-A3B-Preview on Hugging Face on 24 October 2026. It is a 35B mixture-of-experts model with about 3B active parameters, trained to write 32 tokens at a time by diffusion. The weights are 10 to 35 GiB as GGUF, so a 24GB card holds a 4-bit file. The catch: the only llama.cpp support is an open pull request, and since 1 November it runs Clover token by token, without the diffusion speed-up.
Bottom Line
- Yes, on a 24GB card. The official Q4_K_M file is 21.42 GiB, and the KV cache is small. A 16GB card runs a 3-bit file (Q3_K_S, 14.14 GiB).
- You need an unmerged llama.cpp build. Support lives in PR #31306, open as of 1 November 2026. Ollama and LM Studio fail to load the model.
- No diffusion speed-up locally yet. On 1 November the PR became “AR mode only”: one token per step, like any other model. The fast block-diffusion path is not in any official local runtime.
- Context is almost free. Only 10 of 40 layers keep a KV cache. Full 131,072-token context costs about 2.5 GiB at f16.
- Action: if you want Clover’s agent scores today, build the PR and run Q4_K_M on 24GB. If you want the speed DiffuSpace advertises, wait for the follow-up PR.
Ready to buy? See the tested hardware list with current prices.
What Clover 35B-A3B Is
| Clover-35B-A3B-Preview | |
|---|---|
| Maker | DiffuSpace (Shenzhen) |
| On Hugging Face | 24 October 2026 |
| License | Apache-2.0 |
| Base model | Qwen3.6-35B-A3B |
| Parameters | 34.66B total, about 3B active |
| Experts | 256 routed, 8 active, plus 1 shared |
| Layers | 40: 30 Gated DeltaNet (linear), 10 full attention |
| Full-attention KV | 2 KV heads, 256 dimensions each |
| Context | 131,072 tokens |
| Diffusion block | 32 tokens, confidence threshold 0.95 |
| Input | Text only |
How the diffusion works. Clover writes in blocks of 32 tokens. Each block starts fully masked. Every pass fills all positions above 95% confidence, or the single most confident one if none qualify. A finished block becomes the prefix for the next block. Fewer passes than tokens is where the speed comes from.
Why it trends. The model reached Chinese business press on 29 October (21st Century Business Herald, via sfccn.com), with a funding round and a 680 tok/s claim. More than 20 community GGUF and derivative repos appeared on Hugging Face between 28 and 30 October. The official GGUF repo had 20,211 downloads when we read it.
Every GGUF Size
Sizes from the Hugging Face API, read 1 November 2026. 1 GiB = 1,073,741,824 bytes.
| Repo | Quant | Size (GiB) |
|---|---|---|
| mudler (APEX) | I-Micro | 9.86 |
| AtomicChat | AD-IQ3_XXS-IQ2_S | 12.73 |
| maternion | Q3_K_S | 14.14 |
| maternion | Q3_K_M | 15.62 |
| maternion | IQ4_XS | 17.64 |
| Diffuspace (official) | Q4_0 | 18.24 |
| maternion | Q4_K_M | 19.72 |
| Diffuspace (official) | Q4_K_M | 21.42 |
| maternion | Q6_K | 26.56 |
| maternion | Q8_0 | 34.37 |
| Diffuspace (official) | Q8_0 | 35.26 |
bartowski, unsloth, lmstudio-community and mlx-community had no Clover repo on 1 November. The official, maternion and AtomicChat files use the tensor names of the current PR head, so they load in it.
Why the Weights Are the Limit, Not the Cache
Clover inherits the Qwen3.6 hybrid layout. Three of every four layers are Gated DeltaNet layers with a fixed-size state. Only the 10 full-attention layers grow a KV cache.
Our arithmetic from config.json: 10 layers × 2 (K and V) × 2 KV heads × 256 dimensions × 2 bytes = 20 KiB per token at f16.
| Context | KV cache (f16) |
|---|---|
| 8,192 | 0.16 GiB |
| 32,768 | 0.63 GiB |
| 131,072 | 2.5 GiB |
The Gated DeltaNet state adds about 60 MiB, whatever the context. Compare a dense model: Nanbeige4.3 33B needs 208 KiB per token, ten times more. So for Clover, pick the biggest quant that fits and keep the full context.
What Fits Your Machine
| Machine | File to run | Context | Notes |
|---|---|---|---|
| 12GB card | AD-IQ3_XXS-IQ2_S (12.73) or APEX I-Micro (9.86) | Short | Tight. Or offload experts to RAM. |
| 16GB card | Q3_K_S (14.14) | 32K+ | Fully on GPU. |
| 24GB card | Q4_K_M (21.42) | 32K-64K | The comfortable pick. |
| 32GB card | Q6_K (26.56) | Full | Near-lossless. |
| Mac, 32GB | IQ4_XS (17.64) | 32K | Measured 22.5 tok/s on an M1 Pro. |
| Mac or Spark, 64GB+ | Q8_0 (35.26) | Full | Best quality. |
For a 24GB card, a used RTX 3090 averaged $1,563 on resale trackers as of October 2026. For 16GB, a new RTX 5060 Ti 16GB holds Q3_K_S on the GPU; check current listings for its price.
A card smaller than the file still works. Clover is a mixture-of-experts model, so you can keep attention on the GPU and push expert tensors to system RAM. See llama.cpp MoE offload flags explained for --n-cpu-moe. We have not tested that flag on the Clover PR.
How Fast It Runs
Every number here is community-reported. None of them is ours. The numbers do not transfer between machines, builds or settings.
| Hardware | Quant and mode | Generation | Source |
|---|---|---|---|
| Apple M1 Pro 32GB | IQ4_XS, llama-bench tg128 | 22.5 tok/s (pp512 289 tok/s) | austindixson GGUF card |
| NVIDIA GB10 (DGX Spark) | Q4_K_M, llama-server, normal decoding | 13.47 tok/s | PR #31306 comment, 31 Oct |
| NVIDIA GB10 (DGX Spark) | Q4_K_M, old draft-self mode | 8.10 tok/s | PR #31306 comment, 31 Oct |
| RTX 3090 | Q4_K_M, unofficial diffusion port, 480-token reply | 101.6 tok/s | Benzidrine/clover-llamacpp |
| NVIDIA GB300 | BF16, DiffuSpace engine | 680 tok/s | DiffuSpace blog (vendor) |
Read the table with three caveats:
- The GB10 test found self-speculation 40% slower than normal decoding, at an 82.56% draft acceptance rate. That is one reason the PR dropped it.
- The 3090 number is not a chat speed. The unofficial port does not stop at
<|im_end|>, so it always fills the block budget. Its speed also rises with output length: 64.4 tok/s at 96 tokens, 101.6 at 480. - No RTX 4090 or 5090 number exists yet for the PR’s normal decoding. We do not estimate one.
The vendor 680 tok/s figure is batch 1 on a datacenter GPU with DiffuSpace’s own engine. No local runtime reproduces it today.
Run It
Build the PR branch. This example is for an NVIDIA card; on a Mac, drop the CUDA flag (Metal is on by default).
git clone https://github.com/ggml-org/llama.cpp && cd llama.cpp
git fetch origin pull/31306/head:clover && git checkout clover
cmake -B build -DGGML_CUDA=ON && cmake --build build -j --config Release
curl -LO https://huggingface.co/Diffuspace/Clover-35B-A3B-Preview/resolve/main/chat_template.jinja
./build/bin/llama-server -m Clover-35B-A3B-Preview-Q4_K_M.gguf \
-ngl 99 -c 32768 -fa on --jinja --chat-template-file chat_template.jinja- Pass the chat template. Some community GGUFs have no embedded template. Tool calling needs it.
- Ignore the
draft-selfcommand on the official GGUF card. It worked only on the old PR head. - Watch for repetition. The official card warns that very long generations can repeat. Community cards add
--presence-penalty 0.9 --repeat-penalty 1.05. - Expect long thinking. The llama.cpp path has no length or think-end controls. In the GB10 test, speculative runs hit the token cap while still in thinking.
Should You Switch From Qwen3.6-35B-A3B?
DiffuSpace’s benchmark table, vendor-reported:
| Benchmark | Clover | Qwen3.6-35B-A3B |
|---|---|---|
| SWE-bench Verified | 73.6 | 65.8 |
| SWE-bench Pro | 46.5 | 42.7 |
| Terminal-Bench 2.1 | 51.7 | 43.1 |
| BFCL-v4 | 67.9 | 63.4 |
| τ³ | 65.3 | 43.4 |
| IFBench | 72.3 | 61.3 |
| GPQA-Diamond | 83.8 | 83.3 |
| HLE | 12.8 | 21.2 |
| AIME26 | 83.3 | 90.4 |
Switch for agent work if you can build llama.cpp. The agent and coding rows gain 4 to 22 points on the vendor’s table. The memory cost is the same as the base model, because the layout is the same.
Stay on Qwen3.6-35B-A3B if you need Ollama or LM Studio, hard math, or a model that is out of preview. Clover loses 7 points on AIME26 and 8 on HLE. And today you pay the base model’s per-token cost without the speed-up that is Clover’s main idea.
Sources
- Model card and config.json: Diffuspace/Clover-35B-A3B-Preview (read 2026-11-01)
- Official GGUF sizes: Diffuspace/Clover-35B-A3B-Preview-GGUF; community GGUFs: maternion, AtomicChat, austindixson Mac IQ4_XS
- llama.cpp PR #31306: body (“AR mode only”), GB10 test comment by tsurumitsu (2026-10-31)
- Unofficial diffusion port: Benzidrine/clover-llamacpp
- Vendor speed claims and benchmark table: DiffuSpace blog, GitHub repo
- Press: sfccn.com, 2026-10-29
- RTX 3090 price: resaleprices.com, read 2026-10-30
See Also
- Can I Run SenseNova Loop 9B Locally? — another new architecture that needs a patched llama.cpp, 5.6 GB at Q4_K_M
- Open text diffusion models you can run locally — every dLLM that runs on your own hardware
- Best local LLM for 24GB VRAM — what else the Q4_K_M tier holds
- Best local LLM for 16GB VRAM — the 3-bit tier
- llama.cpp MoE offload flags explained — run a 35B MoE on a smaller card
- Can I run Nanbeige4.3 33B locally? — a dense 33B with a 10x larger cache
- Qwen3.6 vs Gemma 4 for agents — the base model in context
- Best models for NVIDIA DGX Spark — the GB10 box from the speed table
Before you order parts, check the tested hardware list for current prices by tier.