namanadep/Codestral-Mamba-7B-vs-Qwen-7B

Codestral Mamba 7B vs. Qwen 2.5 7B: Technical Benchmark & Architectural Comparison

This repository contains the full research, automated test suite, raw JSON benchmark logs, technical reports, and PowerPoint presentations evaluating Codestral Mamba 7B (mamba-codestral:7b) against Qwen 2.5 7B (qwen2.5:7b-instruct).

Created & published by Naman Adep using NAMAN_HF_TOKEN.

---

📊 Summary Performance Comparison

Metric	Codestral Mamba 7B	Qwen 2.5 7B Instruct	Delta / Key Takeaway
Architecture	Selective State Space Model (SSM S6)	Multi-Head Self-Attention Transformer	Mamba eliminates $O(N^2)$ quadratic memory scaling.
Average Latency	4.28s	7.40s	42.2% faster completion time for Codestral Mamba.
Hardware Throughput	194.8 t/s	193.3 t/s	Identical speed on NVIDIA H200 NVL GPUs.
Average Output Length	1,930 chars	4,347 chars	Qwen 2.5 generates 2.25x more content (docstrings & edge cases).
KV-Cache Scaling	Constant $O(1)$ Memory State	$O(N)$ Memory Growth	Mamba uses fixed VRAM up to 256k tokens.

---

📂 Uploaded Files & Repository Inventory

• docs/CODESTRAL_MAMBA_7B_VS_QWEN_7B_COMPARISON.md: 71 KB Exhaustive 10-Prompt Technical Report.
• docs/CODESTRAL_MAMBA_7B_VS_QWEN_7B_COMPARISON.pptx: 9-Slide Visual Benchmark Comparison Deck.
• docs/MAMBA_7B_INSTALLATION_AND_ARCHITECTURE_GUIDE.pptx: 8-Slide Hands-on Installation & Journey Deck.
• docs/MAMBA_MODELS_RESEARCH_OLLAMA_HUGGINGFACE.md: State Space Models Architectural Research.
• data/mamba_vs_qwen_results.json: Raw Evaluation JSON logs across 10 technical categories.
• src/: Complete Python benchmark test harness and slide generation scripts.