Orivael-SRD-Lab/gemma4_e2b_srd4_q4km

Gemma 4 E2B-it — SRD-4 Q4_K_M GGUF

Orivael SRD-4 quantization of google/gemma-4-E2B-it, packed into a HMAC-signed .axm container and extracted to GGUF Q4_K_M for llama.cpp / LM Studio / PocketPal.

Patent pending — Orivael Inc.
The SRD container format and signing protocol are the subject of pending patent claims (ORVL-023, ORVL-024).

---

What is SRD?

Stochastic Residual Dithering (SRD) is a post-training quantization technique developed by Orivael Inc. It applies calibrated noise dithering to residual quantization error, recovering perplexity lost in aggressive bit-width reduction — with no re-training or fine-tuning required.

Every weight shard is HMAC-SHA256 signed inside an .axm container, providing cryptographic provenance: the fingerprint is a public commitment to exactly these weights.

---

Architecture — Mixture of Experts

Gemma 4 E2B is a Mixture-of-Experts (MoE) model:

Property	Value
Total parameters	~5B
Active parameters per token	~2B (25%)
Experts	8 total · 2 active per token
Hidden size	2560
Layers	34
Vocabulary	262,144 tokens
Max context	32,768 tokens
Attention	GQA

SRD uses a MoE-aware chunk split (20 / 20 / 37 / 23 % of depth) rather than the dense-model 40–77 % range, since expert routing layers are distributed across the full depth.

---

Compression

Metric	Value
BF16 baseline	~10.0 GB
SRD-4 container (.axm)	~4.6 GB
Q4_K_M GGUF (this file)	3.11 GB
Compression vs BF16	69 %
Bits per weight (avg)	~4.85 bpw
HMAC governance	Yes — fingerprinted .axm
Re-training required	No

---

Files

File	Size	Description
gemma4_e2b_srd4_q4km.gguf	3.11 GB	GGUF Q4_K_M — llama.cpp / LM Studio / PocketPal
gemma4_e2b_met_sidecar.json	~5 KB	MoE-aware MET slot map + RAM hydration policy

---

Usage

llama.cpp

./llama-cli \
  -m gemma4_e2b_srd4_q4km.gguf \
  -p "<start_of_turn>user\nHello, how are you?<end_of_turn>\n<start_of_turn>model\n" \
  -n 200 --temp 0.7 -ngl 99

LM Studio

Download and open gemma4_e2b_srd4_q4km.gguf directly in LM Studio. Use the Gemma Instruct chat template.

Python (llama-cpp-python)

from llama_cpp import Llama

llm = Llama(
    model_path="gemma4_e2b_srd4_q4km.gguf",
    n_gpu_layers=-1,
    n_ctx=4096,
)
output = llm.create_chat_completion(messages=[
    {"role": "user", "content": "Hello, how are you?"}
])
print(output["choices"][0]["message"]["content"])

---

Hardware Requirements

Device	VRAM / RAM	Notes
RTX 3080 / 4070 (10 GB)	10 GB VRAM	Full GPU offload (-ngl 99)
RTX 3070 (8 GB)	8 GB VRAM	Full GPU offload
Apple M2/M3 (16 GB)	16 GB unified	Full Metal offload
Jetson Orin Nano 8 GB	8 GB	-ngl 99, fits comfortably
CPU only (16 GB RAM)	16 GB	Slow but works

---

Axiom MET Sidecar

The included gemma4_e2b_met_sidecar.json provides Memory-Efficient Trajectory (MET) hydration budgets for on-device inference engines that support Axiom's intent-aware layer loading:

{
  "moe": {
    "num_experts": 8,
    "experts_per_token": 2,
    "active_frac": 0.25,
    "chunk_note": "20/20/37/23 split — architecture-agnostic for MoE"
  },
  "hydration_policy": {
    "INFORM": ["early"],
    "HARM":   ["early", "factual", "reasoning", "governance"]
  }
}

---

Benchmark

PPL benchmark in progress.
WikiText-2 perplexity (stride 512, context 2048, 100 chunks) will be posted here once validated with a current llama.cpp build. An earlier run produced an invalid result due to a stale llama.cpp version at conversion time — that number is not published here.

Metric	SRD Q4_K_M	Community Q4_K_M
bpw	~4.85	~4.85
Size	3.11 GB	pending
WikiText-2 PPL	pending re-run	pending
HMAC governance	Yes	No
MoE-aware sidecar	Yes	No

---

License

Weights are derived from google/gemma-4-E2B-it and are subject to the Gemma Terms of Use.
The SRD quantization method and .axm container format are proprietary to Orivael Inc. (patent pending ORVL-023, ORVL-024).

---

Citation

@misc{orivael2026srd,
  title  = {Stochastic Residual Dithering (SRD): Post-Training Quantization
             with Cryptographic Provenance},
  author = {Orivael Inc.},
  year   = {2026},
  note   = {Patent pending ORVL-023, ORVL-024}
}