Axion-1.0 35B - NanoPlus 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.
[!NOTE] Vision: this repo ships the text GGUF. For multimodal use, grab
Axion-1.0-35B.mmproj-bf16.gguf(or its quantized versions) from the upstream GGUF repo.MTP/DSpark: the main GGUF has no MTP tensors and no DSpark draft is included for this family.
Family cross-links
| Tier | File | Base | Status |
|---|---|---|---|
| NanoPlus | Axion-1.0-35B.APEX-I-NanoPlus-V2.1.gguf |
Axion-1.0-35B | Published |
| MiniPlus | see repo | Axion-1.0-35B | Published |
Architecture and precision
- Base model:
Axion1-0/Axion-1.0-35B(Apache-2.0). BF16 GGUF reference from prithivMLmods/Axion-1.0-35B-GGUF. - Architecture: 40 layers, 30 linear-attention/SSM + 10 full-attention; 256 routed experts + 1 shared, 8 active; hidden 2048; 262,144-token context; 248,320 vocab.
- iMatrix: public
Axion-1.0-35B.imatrix.ggufby mradermacher, size and sha256 verified on download before use. - Recurrent shielding:
ssm_alpha,ssm_a,ssm_dt.bias,ssm_conv1dand RMSNorms remainF32. - Layer 0:
attn_gateinQ8_0; the linear-attention projections (attn_qkv,ssm_beta,ssm_out) inQ3_K.
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 |
131 | F32 |
Uncompressed numerical stability across all layers. |
| Expert Routers | blk.*.ffn_gate_inp, ffn_gate_inp_shexp |
80 | 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 | Q4_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 needle-in-a-haystack 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 | Q4_K |
Linear-attention projections kept lean for CPU execution. |
| Edge MoE Experts | Layers 0-9 & 30-39 (ffn_*_exps) |
4 | Q3_K |
Routed experts calibrated with the importance matrix. |
| Edge MoE Experts | Layers 0-9 & 30-39 (ffn_*_exps) |
24 | IQ2_S |
Routed experts calibrated with the importance matrix. |
| Edge MoE Experts | Layers 0-9 & 30-39 (ffn_*_exps) |
8 | IQ2_XXS |
Routed experts calibrated with the importance matrix. |
| Edge MoE Experts | Layers 0-9 & 30-39 (ffn_*_exps) |
24 | IQ3_XXS |
Routed experts calibrated with the importance matrix. |
| Core MoE Experts | Layers 10-29 (ffn_*_exps) |
26 | IQ2_S |
Deep-band routed experts calibrated with the importance matrix. |
| Core MoE Experts | Layers 10-29 (ffn_*_exps) |
14 | IQ2_XXS |
Deep-band routed experts calibrated with the importance matrix. |
| Core MoE Experts | Layers 10-29 (ffn_*_exps) |
20 | IQ3_XXS |
Deep-band routed experts calibrated with the importance matrix. |
The machine-readable version is tensor-map.tsv.
Files
| File | Size | Notes |
|---|---|---|
Axion-1.0-35B.APEX-I-NanoPlus-V2.1.gguf |
12,554,402,496 bytes | Main NanoPlus V2.1 GGUF |
| tensor-map.tsv | 33,214 bytes | Per-tensor types |
| chat_template.jinja | from source | Axion chat template |
Usage
llama-cli -m Axion-1.0-35B.APEX-I-NanoPlus-V2.1.gguf -p "Hello" -ngl 99 -c 32768
llama-server -m Axion-1.0-35B.APEX-I-NanoPlus-V2.1.gguf --jinja -ngl 99 -c 32768
Limitations
- Base benchmarks are self-reported by the Axion team; no independent evaluation is claimed for this tier.
- Quantization quality is theoretical until you test it on your own workload.
License
Apache-2.0, inherited from the base model. See the source model card.