Apex-35B-V1.0 APEX-I MiniPlus V2.1
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.
Family cross-links
Architecture and precision
- Base model:
Apexintelligence-AI/Apex-35B-V1.0(Apache-2.0). BF16 single GGUF reference from prithivMLmods/Apex-35B-V1.0-MXFP4-MoE-GGUF. - Architecture: 40 layers, 30 linear-attention/SSM + 10 full-attention; 256 routed experts + 1 shared, 8 active; hidden 2048; expert FF 512; 262,144-token context; 248,320 vocab; 753 tensors including the MTP block 40.
- iMatrix: public
Apexintelligence-AI_Apex-35B-V1.0-imatrix.ggufby bartowski, 384,343,168 bytes, SHA256b1f64857f0fa390657d8e9155e0bb0ac39ba7503b1f9dbd3c5705237016440e3, verified before use. - Recurrent shielding:
ssm_alpha,ssm_a,ssm_dt.bias,ssm_conv1dand RMSNorms remainF32. - Layer 0: no special boundary; the same SSM types as the rest of the layers.
- MTP: block 40, 20 tensors; F32 norms,
Q4_Kq/k/v andnextn.eh_proj,Q6_Koutput,Q4_0routed experts,Q5_Kshared experts. No DSpark draft is included for this family. - Vision projector: this repository ships the text GGUF; the upstream
mmproj-Apexintelligence-AI_Apex-35B-V1.0-bf16.ggufcompanion is not included.
Surgical Tensor Quantization Map (Audited from GGUF)
The exact tensor breakdown below has been verified directly from the compiled binary weights:
| Layer Group | Sub-Component / Tensor | Qty | Precision | Engineering Rationale |
|---|---|---|---|---|
| Global Output Head | output.weight |
1 | Q6_K |
Preserves near-FP16 token classification; stable syntax and bracket behavior. |
| Global Embeddings | token_embd.weight |
1 | Q4_K |
High-fidelity vocabulary representation. |
| All Normalizations | output_norm, attn_*_norm, post_attention_norm, ssm_norm + MTP norms |
138 | F32 |
Uncompressed numerical stability across all 40 layers plus block 40. |
| Expert Routers | blk.*.ffn_gate_inp, ffn_gate_inp_shexp (40 layers + MTP) |
82 | F32 |
Uncompressed routing over 256 experts; zero router drift. |
| Attention Gates | blk.*.attn_gate.weight (30 Hybrid Layers) |
30 | Q8_0 |
High-precision gating across the hybrid linear-attention layers. |
| Shared Foundation Experts | blk.*.ffn_{gate,down,up}_shexp (All 40 Layers) |
120 | Q5_K |
Active on every token; protected in high precision. |
| Periodic Full Attention | blk.{3,7,11,...}.attn_q/k/v (10 Anchor Layers) |
30 | Q4_K |
Full quadratic-attention anchors for deep retrieval. |
| Periodic Full Attention | blk.{3,7,11,...}.attn_output (10 Anchor Layers) |
10 | Q6_K |
Armored attention output projection over long context. |
| Recurrent SSM Scales | blk.*.ssm_alpha, ssm_a, ssm_conv1d, ssm_dt |
120 | F32 |
F32 shielding against recurrent-state drift. |
| Linear Attention & SSM | blk.*.attn_qkv, ssm_beta, ssm_out |
90 | Q3_K |
Linear-attention projections kept lean for CPU execution. |
| Edge MoE Experts | Layers 0-9 & 30-39 (ffn_*_exps) |
60 | Q3_K |
Routed experts calibrated with the importance matrix. |
| Core MoE Experts | Layers 10-29 (ffn_*_exps) |
60 | IQ3_XXS |
Deep-band routed experts calibrated with the importance matrix. |
| MTP Block 40 | blk.40.{attn_output,attn_q/k/v,nextn.eh_proj,ffn_*_exps,ffn_*_shexp} |
11 | Q6_K output; Q4_K q/k/v + eh_proj; Q4_0 routed experts; Q5_K shared |
Companion MTP head; 20 tensors total including its 7 F32 norms and 2 routers counted above. |
The Qty column sums to 753. The complete audited list is tensor-map.tsv.