Quick Navigation Index
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.
- Model Overview & Technical Specifications
- Empirical Importance Matrix Calibration
- Bundled Multimodal Vision Projector
- Surgical Tensor Quantization Map
- Recommended Server & Inference Configuration
- Integrated MTP Speculative Decoding
- Community Compute Fund & Student Support
1. Model Overview & Technical Specifications
This repository provides Sage-A1.5 APEX-I-NanoPlus V2.1, a surgical ultra-compact quantization of InternScience/Sage-A1.5.
The architecture comprises 40 hybrid layers (linear attention / DeltaNet recurrence interleaved with periodic quadratic attention anchors every 4 layers), 256 routed micro-experts per layer with top-k=8 activation, and an integrated speculative Multi-Token Prediction (MTP) block in blk.40, totaling exactly 753 tensors.
NanoPlus V2.1 is engineered to fit within compact VRAM envelopes (such as 16GB VRAM GPUs or 24GB consumer cards with massive context windows) or stream via fast system RAM, while strictly protecting all normalizations, routers, attention gates, and anchor output heads.
Key Characteristics
- Base Model:
InternScience/Sage-A1.5 - Architecture:
qwen35moe(Hybrid Recurrent + Periodic Full Attention MoE + Integrated MTP) - Total Tensors: 753 (733 main model layers + 20 integrated MTP tensors)
- Quantization Profile: APEX-I-NanoPlus V2.1
- Calibration Corpus: 1,001,472 tokens (VAL-1M multi-domain corpus)
- MoE Expert Coverage: 100.0% (30,720 / 30,720 expert projection matrices observed, 0 unobserved)
2. Empirical Importance Matrix Calibration
Quantization uses the official empirical Importance Matrix Sage-A1.5-VAL-1M-I-Matrix.GGUF from IsValorum/Sage-A1.5-VAL-1M-I-Matrix-GGUF.
Calibration Parameters
- Hardware: Dedicated NVIDIA A100 SXM4 80GB GPU
- Token Volume: 1,001,472 tokens across 489 sequential chunks (chunk size 2048)
- Composition: Code (35.0%), Reasoning & Mathematics (25.0%), Multi-Domain Prose & Natural Languages (25.0%), Tool Use & Structured JSON (15.0%)
- Natural Language Diversity: English (134,960 tok), Spanish (97,146 tok), Mandarin Chinese (96,322 tok), French (77,578 tok), German (73,636 tok), Russian (71,940 tok), Japanese (70,894 tok), Portuguese (69,576 tok), Arabic (69,566 tok), Italian (69,456 tok), Korean (69,444 tok)
- Programming Languages: Python, JavaScript, TypeScript, C++, Rust, Go, Java, Bash, SQL, JSON Schema
- MoE Coverage Metric: 100.0% expert activation across all 40 layers and 256 micro-experts
3. Bundled Multimodal Vision Projector
Sage-A1.5 features multimodal visual reasoning capabilities. For visual inference, the high-precision multimodal projector Sage-A1.5-mmproj.gguf is bundled directly in this repository.
When launching llama-server or CLI tools, include --mmproj Sage-A1.5-mmproj.gguf to enable visual inputs.
4. Surgical Tensor Quantization Map
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; eliminates syntax errors and bracket drops. |
| Global Embeddings | token_embd.weight |
1 | Q4_K |
High-fidelity vocabulary embedding representation. |
| All Normalizations | output_norm, attn_*_norm, post_attention_norm, ssm_norm |
131 | F32 |
100% uncompressed numerical stability across all 40 layers. |
| Expert Routers | blk.*.ffn_gate_inp, ffn_gate_inp_shexp |
80 | F32 |
100% uncompressed routing fidelity across 256 micro-experts; zero router drift. |
| Attention Gates | blk.*.attn_gate.weight (30 Hybrid Layers) |
30 | Q8_0 |
High-precision attention gating across hybrid DeltaNet recurrence layers; eliminates crosstalk. |
| Shared Foundation Experts | blk.*.ffn_{gate,down,up}_shexp (All 40 Layers) |
120 | Q4_K |
Foundation knowledge backbone active on 100% of tokens; protected in linear Q4_K for fast streaming. |
| Periodic Full Attention | blk.{3,7,11,...}.attn_q/k/v (10 Anchor Layers) |
30 | Q4_K |
Full quadratic attention anchor checkpoints for deep needle-in-a-haystack retrieval. |
| Periodic Full Attention | blk.{3,7,11,...}.attn_output (10 Anchor Layers) |
10 | Q6_K |
High-precision attention output projection over deep context. |
| Recurrent SSM Scales | blk.*.ssm_alpha, ssm_a, ssm_conv1d, ssm_dt |
120 | F32 |
Guarded in uncompressed FP32 to prevent DeltaNet recurrent state drift. |
| Linear Attention & SSM | blk.*.attn_qkv, ssm_beta, ssm_out |
90 | Q4_K |
Linear AVX2 execution; zero SIMD CPU stalls during system RAM streaming. |
| Border MoE Down-Proj | Layers 0–1 & 38–39 (ffn_down_exps) |
4 | Q3_K |
Linear SIMD execution optimized for token entry and exit stability. |
| Border MoE Gate/Up | Layers 0–1 & 38–39 (ffn_gate/up_exps) |
8 | IQ3_XXS |
High-density boundary protection guided by imatrix. |
| Core MoE Down-Proj | Layers 2–37 (ffn_down_exps) |
36 | IQ3_XXS |
Fortified 3.06 bpw residual stream; preserves core mathematical, coding, and reasoning capacity. |
| Core MoE Gating | Layers 2–37 (ffn_gate_exps) |
36 | IQ2_S |
High-precision 2.50 bpw SwiGLU gating; eliminates activation noise. |
| Core MoE Up-Proj | Layers 2–15 (ffn_up_exps) |
14 | IQ2_S |
Enhanced 2.50 bpw precision for sensitive early-intermediate feature extraction. |
| Core MoE Up-Proj | Layers 16–37 (ffn_up_exps) |
22 | IQ2_XXS |
High 2.06 bpw compression in deep MoE layers to reach compact memory footprint. |
| Integrated MTP Block | blk.40.attn_output (Speculative Output) |
1 | Q6_K |
High-precision output for speculative draft token validation. |
| Integrated MTP Attention | blk.40.attn_q/k/v, nextn.eh_proj |
4 | Q4_K |
Guarded hidden projection and attention matrices. |
| Integrated MTP Shared | blk.40.ffn_{gate,down,up}_shexp |
3 | Q4_K |
Protected shared expert for speculative draft generation. |
| Integrated MTP Experts | blk.40.ffn_{gate,down,up}_exps |
3 | Q4_0 |
Linear quantization for fast speculative forward passes. |
| Integrated MTP Norms | blk.40.*_norm, nextn.*_norm, blk.40.ffn_gate_inp* |
9 | F32 |
100% uncompressed numerical stability in speculative routing and RMSNorms. |
The complete 100% audited tensor list is attached in tensor-map.tsv in this repository (753 / 753 tensors).
5. Recommended Server & Inference Configuration
Launch llama-server with the following recommended parameters:
llama-server \
-m Sage-A1.5.APEX-I-NanoPlus-V2.1.gguf \
--mmproj Sage-A1.5-mmproj.gguf \
-c 32768 \
-ngl 99 \
--threads 16 \
--port 8080
Recommended Generation Parameters
- Temperature: 0.6 - 0.7
- Top-P: 0.95
- Min-P: 0.05
- Repeat Penalty: 1.05
6. Integrated MTP Speculative Decoding
Sage-A1.5 incorporates a native Multi-Token Prediction (MTP) block stored directly in blk.40. This enables speculative self-decoding without requiring an external draft model:
llama-server \
-m Sage-A1.5.APEX-I-NanoPlus-V2.1.gguf \
--draft-max 2 \
-c 32768 \
-ngl 99
7. Community Compute Fund & Student Support
All IsValorum quantizations will always remain completely free and open to the public without paywalls.
However, cloud GPU compute is expensive and funded entirely out-of-pocket. If you find these handcrafted builds to be of high quality and they benefit your day-to-day workflow, consider supporting future releases on Ko-fi. Every contribution directly funds the cloud compute hours required for new model runs and helps support my university education:
(When supporting on Ko-fi, feel free to leave a note with your Hugging Face handle and the specific model you would like prioritized).