Clover-35B-A3B-Preview APEX-I-MiniPlus-V2.1 GGUF

APEX Quantization Series by IsValorum

[!IMPORTANT] Independently computed using dedicated cloud compute clusters, funded 100% out-of-pocket.

If you consider these releases to be of high quality and they help you in your day-to-day workflow, consider supporting on Ko-fi. Every contribution directly sustains future cloud compute hours and helps support my university studies.

[!WARNING] This file does not load in vanilla llama.cpp. Clover uses the clover_moe architecture, which upstream does not implement. Use the precompiled builds in IsValorum/llama.cpp-clover-moe, which ships Linux x86_64 CPU and CUDA 12 builds plus a Windows x86_64 universal package (NVIDIA CUDA 12.8, Vulkan for AMD and Intel GPUs, and CPU), or compile the runtime from the patched branch (PR #31306, commit 4838ffb0c8ade92f34d9157deea38d92758ab824). The CUDA builds need an NVIDIA GPU; the CPU builds run on any x86-64 machine.

MiniPlus raises the shared experts to Q5_K and keeps the edge experts at 3-bit Q3_K, the higher-fidelity tier for this family.

Family cross-links

MicroPlus is not published for this family. There is no abliterated edition of Clover: the release is a single post-trained preview.

Related artifacts: BF16 source · Q8_0 · 1.25M I-Matrix · patched llama.cpp · opposite tier

Architecture and precision

• Base model: Diffuspace/Clover-35B-A3B-Preview, a clover_moe derivative of Qwen3.6-35B-A3B with an added block-diffusion path
• Architecture: 40 layers, 30 gated DeltaNet (linear SSM) layers plus 10 full-attention layers, 256 routed experts with top-8 routing, 131,072 token context
• Parameters: 34,661,350,848
• iMatrix: VAL-1.25M, computed over 1.25M tokens, SHA256 05d9669137d8ca843f0a7d3a8f00975e84d48644832d2734c564a1778b2f2d5d
• Recipe: recipe_moe_miniplus_v2_1_clover.txt (853 entries), one line per tensor, no fallback
• Recurrent shielding: decay factors, 1D convolutions, and RMSNorms in F32
• Block-diffusion armoring: ssm_bd_logit_scale.weight, ssm_bd_history.weight, ssm_bd_history.bias, and ssm_conv1d.weight_right in F32 across all 30 GDN layers
• Output head in Q6_K, token embedding in Q4_K, attention gates in Q8_0, router logits (ffn_gate_inp, ffn_gate_inp_shexp) in F32

Empirical evaluation metrics

No empirical evaluation is published for this release. This block is reserved for the pipeline that fills it automatically.

File and metrics

tensor-map.tsv in this repository lists all 853 tensors with their final type, shape, and parameter count. quant-report.json holds the machine-readable version of the table above.

Audited tensor map sample

The complete audited list is tensor-map.tsv.

Runtime requirement

Vanilla llama.cpp cannot run this model. The clover_moe architecture is not implemented upstream, and stock builds abort at load time with an unknown-architecture error. Everything you need is published at IsValorum/llama.cpp-clover-moe, built from upstream PR #31306, commit 4838ffb0c8ade92f34d9157deea38d92758ab824:

The runtime archives ship llama-cli, llama-server, llama-diffusion-cli, llama-quantize, llama-imatrix, and llama-perplexity. llama-diffusion-cli is the entry point for the block-diffusion path. They are dynamically linked, so point the loader at the bundled libraries before running:

tar xzf llama-clover-4838ffb-linux-x86_64-cpu.tar.gz
export LD_LIBRARY_PATH=$PWD/bin:$LD_LIBRARY_PATH
./bin/llama-cli -m Clover-35B-A3B-Preview.APEX-I-MiniPlus-V2.1.gguf -ngl 99 -c 8192

On Windows, extract the universal ZIP and run from its root:

set PATH=%~dp0bin;%PATH%
llama-completion.exe -m Clover-35B-A3B-Preview.APEX-I-MiniPlus-V2.1.gguf -p "hi" -n 1 -c 128 -b 16 -ub 16 -ngl 8 -no-cnv

Drop -ngl 99 on the CPU build. The Linux x86_64 binaries require GLIBC_2.38 or newer, so run them on Ubuntu 24.04 or equivalent; the Windows package has no glibc requirement and picks its backend at load time. macOS is not published yet: build from the same branch and commit, as for any other BLAS or CUDA configuration.

[!NOTE] What the prebuilt runtimes cover. Linux x86_64: any CPU, and NVIDIA CUDA 12 for sm_80/86/89/90. Windows x86_64, in a single universal package: NVIDIA CUDA 12.8 from sm_50 through sm_120, AMD and Intel GPUs through Vulkan 1.4.363, and any CPU with per-CPU backends. Not covered: macOS, ARM64, ROCm, and Blackwell on Linux; those need a build from the same branch and commit.

Reproducibility

llama-quantize --imatrix Clover-VAL-1.25M-I-Matrix.gguf \
  --tensor-type-file recipe_moe_miniplus_v2_1_clover.txt \
  Clover-35B-A3B-Preview-BF16.gguf \
  Clover-35B-A3B-Preview.APEX-I-MiniPlus-V2.1.gguf Q4_0 64

The 853-line recipe overrides every tensor, including the 120 block-diffusion tensors that the canonical APEX-I MoE V2.1 recipes do not contain. Without those lines llama-quantize would silently fall back to the command-line type for them.

Weight parity with the parent MoE

The block-diffusion path adds 120 tensors and 740,160 parameters, 0.002 percent of the model. Held in F32 they cost 2,960,640 bytes over a quantization of the parent architecture without them, 0.024 percent of the file. Every one of the 733 shared tensors receives the same type it would receive in the parent MoE, so bits per weight over the shared set is identical and the footprint difference is the diffusion block alone.

Limitations

• Preview release. No benchmark of any kind is published here, and none should be assumed.
• No abliterated edition exists, and none should be assumed.