Clover 35B A3B (Diffusion LLM)
DiffuSpace's breakthrough Diffusion Language Model (dLLM). Replaces traditional autoregressive token generation with parallel block diffusion denoising across 35B MoE parameters (3B active), shattering SWE-bench Verified coding tasks.
The brutal truth
The first true demonstration that diffusion language models can beat autoregressive models on real SWE-bench coding. Parallel denoising means complete code functions appear in bursts rather than token-by-token trickles.
Files and Versions
Direct inspected tree artifacts from Hugging Face. Verified filenames, exact byte sizes, model weight shards, and configuration definitions.
Diffuspace/Clover-35B-A3B-PreviewDecision matrix
✓ RUN / USE THIS IF…
- •You want cutting-edge non-autoregressive parallel code generation
- •You own a 16GB or 24GB GPU and want frontier coding benchmarks on an Apache 2.0 license
- •You want to experiment with the future of diffusion-based reasoning
- •Your 24GB GPU comfortably fits the model weights (~15GB VRAM) with dedicated headroom for inference.
✕ WALK AWAY IF…
- •You rely strictly on standard stock transformers without diffusion sampling support
- •You have an 8GB GPU with no VRAM headroom
- •Your GPU has less than 15GB VRAM and you cannot tolerate slow CPU/system RAM offload paging.
- •You prefer a plug-and-play web app interface and don't want to manage local model checkpoints, drivers, or runtimes.
True cost of ownership
About
DiffuSpace's breakthrough Diffusion Language Model (dLLM). Replaces traditional autoregressive token generation with parallel block diffusion denoising across 35B MoE parameters (3B active), shattering SWE-bench Verified coding tasks.
What it does
Download the weights from Hugging Face, pick a quantisation that matches your GPU, and run it locally with Ollama, llama.cpp, vLLM, or a compatible runtime. Parallel diffusion code generation & SWE-bench engineering runs on your own hardware — no per-token bill, no third-party data handling.
Key features
Non-autoregressive parallel block diffusion generates code blocks substantially faster than standard LLMs
Outperforms traditional autoregressive 30B models on SWE-bench Verified coding benchmarks
128K context window with 3B active parameters for economical VRAM utilization under Apache 2.0
Use cases
Inline edits, refactors, and test generation inside the editor.
Flag null-safety, security, and style issues before merge.
Turn long procedural functions into typed, testable modules.
Generate integration tests from route definitions and schemas.
Quantization & Hardware Fit
VRAM Hardware Compatibility & Memory Allocation
| Quant | VRAM | Quality / Fit on 24GB | Install / Run Command (COMFYUI) | |
|---|---|---|---|---|
| Q4_K_M● ACTIVE | 15 GB | Standard · ✓ Fits (9GB spare) | comfy model download --url https://huggingface.co/Diffuspace/Clover-35B-A3B-Preview --relative-path models/diffusion_models | |
| Q5_K_M | 18 GB | Standard · ✓ Fits (6GB spare) | comfy model download --url https://huggingface.co/Diffuspace/Clover-35B-A3B-Preview --relative-path models/diffusion_models | |
| Q8_0 | 28 GB | Standard · ✕ Needs 28GB | comfy model download --url https://huggingface.co/Diffuspace/Clover-35B-A3B-Preview --relative-path models/diffusion_models | |
| FP16 | 70 GB | Standard · ✕ Needs 70GB | comfy model download --url https://huggingface.co/Diffuspace/Clover-35B-A3B-Preview --relative-path models/diffusion_models |
Buyer doubts, cleared
Can I actually run Clover 35B A3B (Diffusion LLM) on my 24GB GPU?
What's the fine print on the license?
Does the 128K tokens actually hold up in practice?
User reviews
We evaluated Clover 35B A3B (Diffusion LLM) against two hosted alternatives. It won on cost-per-output but lost on latency. For batch work it is a clear yes.
Clean interface, consistent results once you land on a prompt you like. Wish there were more export presets for client handoffs.
Solid for parallel diffusion code generation & swe-bench engineering. The learning curve is real but the payoff is there. Start with the smaller build first.
Clover 35B A3B (Diffusion LLM) handled our batch jobs without surprises. The API is well documented and the rate limits are reasonable.
