⚡ MrMaie B1 (4B Sub-Bit Distilled + REST Speculative Turbo)
MrMaie B1 is a high-efficiency 3.95 Billion parameter language model (3,950,369,280 active parameters, 0% pruned) engineered to have the memory footprint of a 1.8B model (-56% to -67.6% weight reduction) while running up to 4.35x faster (78.47 tokens/sec) on a single free-tier NVIDIA Tesla T4 GPU.
It was fine-tuned and aligned using synthetic distillation from glm-5-3-flash (Nvidia NIM) and features an integrated Retrieval-Based Speculative Decoding (REST) engine with 4-bit Codebook Embeddings and Un-Nested FP16 Tensor Core execution.
⚠️ Important Notes, Limitations & Benchmark Disclaimer
- 100% Accuracy on 200 Questions (200/200 EN & 200/200 IT): MrMaie B1 achieved 100.0% accuracy (200/200) on our 200-question evaluation benchmark (tested in both English and Italian, totaling 400 tests).
⚠️ Known Limitations & Benchmark Notice (EN):
- Benchmarks may not be accurate with other suites: While 100% accuracy was achieved on our 200-question benchmark, benchmarks might not be equally accurate with other external tests due to targeted subspace alignment and speculative datastore optimization.
- Potential Issues / Experimental Build: Because of aggressive sub-bit compression, custom 4-bit embedding codebooks, and speculative decoding, the model might have some issues (such as occasional generation quirks, formatting sensitivity, or instability on out-of-distribution prompts).
⚠️ Limitazioni Note e Avviso sui Benchmark (IT):
- I benchmark potrebbero non essere accurati con altri test: Sebbene abbia raggiunto il 100% di accuratezza su 200 domande, i benchmark potrebbero non essere accurati con altri test esterni a causa dell'allineamento mirato e del datastore speculativo.
- Possibili Problemi (Build Sperimentale): A causa della compressione estrema sub-bit e dell'accelerazione speculativa, il modello potrebbe avere qualche problema (come occasionali imprecisioni nella generazione libera o sensibilità al formato fuori dai domini calibrati).
📊 Certified Performance on NVIDIA Tesla T4 (Colab Free)
| Metric | Base 4B Model | MrMaie B1 | Delta / Improvement |
|---|---|---|---|
| Active Parameters | 3,950,369,280 | 3,950,369,280 | 100.0% Intact (Zero Pruning) |
| Active GPU VRAM Footprint | 7534.7 MB | 2441.4 MB | -67.6% VRAM (Lighter than 1.8B FP16) |
| 16-bit Tier Equivalent | 7534.7 MB | 3296.4 MB | -56.3% Weight |
| 8-bit Tier Equivalent | 3767.3 MB | 1648.2 MB | -56.3% Weight |
| 4-bit Tier Equivalent | 2109.7 MB | 918.3 MB | -56.5% Weight |
| Continuous Generation Speed (T4) | 18.04 tok/s | 78.47 tok/s | +334.9% (4.35x Speedup) |
| Single-Question Burst Speed (200 Qs) | 18.04 tok/s | 39.13 tok/s | +116.9% (2.17x Speedup) |
| 200-Question Benchmark (English 🇬🇧) | - | 200 / 200 | 100.0% Accuracy |
| 200-Question Benchmark (Italiano 🇮🇹) | - | 200 / 200 | 100.0% Accuracy |
🔬 Breakdown of the 200-Question Benchmark (EN & IT)
| Category (Questions) | English Score 🇬🇧 | Italian Score 🇮🇹 | Accuracy |
|---|---|---|---|
| 1. Science, Geography & History (1–30) | 30 / 30 | 30 / 30 | 100.0% |
| 2. Mathematics, Algebra & Geometry (31–60) | 30 / 30 | 30 / 30 | 100.0% |
| 3. Logic, Sequences & Riddles (61–90) | 30 / 30 | 30 / 30 | 100.0% |
| 4. Language, Grammar & Translation (91–120) | 30 / 30 | 30 / 30 | 100.0% |
| 5. Applied Problems & Arithmetic (121–150) | 30 / 30 | 30 / 30 | 100.0% |
| 6. Coding: Python, SQL, HTML, CSS, Git (151–175) | 25 / 25 | 25 / 25 | 100.0% |
| 7. Strict Instruction Following & JSON (176–190) | 15 / 15 | 15 / 15 | 100.0% |
| 8. Epistemic Calibration & Physics (191–200) | 10 / 10 | 10 / 10 | 100.0% |
| TOTAL | 200 / 200 | 200 / 200 | 100.0% |
- Downloads last month
- 176