{ "slide_number": 1, "slide_type": "title", "title": "On-Premises LLM Deployment: Candidate Model Evaluation", "subtitle": "GLM-5.3-Flash vs. DeepSeek V4.1-Flash vs. Qwen-3.8-Flash-Next", "content": [ "Target Infrastructure: NVIDIA DGX H200 (8x H200, 141GB each)", "Prepared for: [Client Name]", "Date: September 2026", "Confidential – Internal Use Only" ], "notes": "Opening slide. Set context immediately: the client is evaluating three open-weight models for self-hosted deployment on DGX H200 hardware." }, { "slide_number": 2, "slide_type": "agenda", "title": "Agenda", "content": [ "1. Model Background – Developers, Licenses, Architecture & Hardware Requirements", "2. Benchmark Comparison – Coding, Reasoning, and Agentic Performance", "3. System Throughput & Concurrency – Expected Performance on DGX H200", "4. Recommendation & Next Steps" ], "notes": "Four content slides. The agenda mirrors the logical flow: understand the models first, then compare performance, then assess operational feasibility on your target hardware." }, { "slide_number": 3, "slide_type": "text_with_bullets", "title": "Model Background – Developers, Licenses & Architecture", "content": [ { "model": "GLM-5.3-Flash", "developer": "Z.ai (Zhipu AI)", "license": "MIT", "architecture": "320B total parameters, 18B active per token (Mixture-of-Experts). Hybrid linear and sparse attention. 1M token context window. Reported KV cache ~4.4x smaller than predecessor.", "hardware_requirement": "Reference config: 8x H200 node (1,128 GB total HBM). Full-precision weights alone consume 700–800 GB. Quantized (1-bit) reduces to ~100 GB RAM/VRAM." }, { "model": "DeepSeek V4.1-Flash", "developer": "DeepSeek", "license": "MIT", "architecture": "552B backbone (763B with vision encoder). Causal encoder-decoder: 40 layers (20 encoder, 20 decoder), 384 routed experts + 1 shared expert per layer. 8B active during prefill, 16B during decode. FP4 KV cache at 890 bytes/token.", "hardware_requirement": "Minimum 8x H200 node or GB200 NVL4 tray. vLLM recipe specifies 614 GB minimum VRAM. Checkpoint storage ~511 GB. 4-bit serving needs 4–8 GPUs in 80GB class." }, { "model": "Qwen-3.8-Flash-Next", "developer": "Alibaba / Qwen", "license": "Qwen Community License 1.0 (commercial use generally free; separate contract required for certain business types)", "architecture": "125B MoE + 51B N-gram lookup table (stored in system RAM, not VRAM). 6B active per token. Qwen4 architecture preview: sparse attention + Gated DeltaNet alternating, gated residuals. 262K native context (extendable to 1M via YaRN).", "hardware_requirement": "The 51B N-gram table requires ~96 GB system RAM. FP8 weights: ~186 GB VRAM. Q4 GGUF: ~82 GB. 1-bit: 75 GB. Fits on a sin
{ "slide_number": 1, "slide_type": "title", "title": "On-Premises LLM Deployment: Candidate Model Evaluation", "subtitle": "GLM-5.3-Flash vs. DeepSeek V4.1-Flash vs. Qwen-3.8-Flash-Next", "content": [ "Target Infrastructure: NVIDIA DGX H200 (8x H200, 141GB each)", "Prepared for: [Client Name]", "Date: September 2026", "Confidential – Internal Use Only" ], "notes": "Opening slide. Set context immediately: the client is evaluating three open-weight models for self-hosted deployment on DGX H200 hardware." }, { "slide_number": 2, "slide_type": "agenda", "title": "Agenda", "content": [ "1. Model Background – Developers, Licenses, Architecture & Hardware Requirements", "2. Benchmark Comparison – Coding, Reasoning, and Agentic Performance", "3. System Throughput & Concurrency – Expected Performance on DGX H200", "4. Recommendation & Next Steps" ], "notes": "Four content slides. The agenda mirrors the logical flow: understand the models first, then compare performance, then assess operational feasibility on your target hardware." }, { "slide_number": 3, "slide_type": "text_with_bullets", "title": "Model Background – Developers, Licenses & Architecture", "content": [ { "model": "GLM-5.3-Flash", "developer": "Z.ai (Zhipu AI)", "license": "MIT", "architecture": "320B total parameters, 18B active per token (Mixture-of-Experts). Hybrid linear and sparse attention. 1M token context window. Reported KV cache ~4.4x smaller than predecessor.", "hardware_requirement": "Reference config: 8x H200 node (1,128 GB total HBM). Full-precision weights alone consume 700–800 GB. Quantized (1-bit) reduces to ~100 GB RAM/VRAM." }, { "model": "DeepSeek V4.1-Flash", "developer": "DeepSeek", "license": "MIT", "architecture": "552B backbone (763B with vision encoder). Causal encoder-decoder: 40 layers (20 encoder, 20 decoder), 384 routed experts + 1 shared expert per layer. 8B active during prefill, 16B during decode. FP4 KV cache at 890 bytes/token.", "hardware_requirement": "Minimum 8x H200 node or GB200 NVL4 tray. vLLM recipe specifies 614 GB minimum VRAM. Checkpoint storage ~511 GB. 4-bit serving needs 4–8 GPUs in 80GB class." }, { "model": "Qwen-3.8-Flash-Next", "developer": "Alibaba / Qwen", "license": "Qwen Community License 1.0 (commercial use generally free; separate contract required for certain business types)", "architecture": "125B MoE + 51B N-gram lookup table (stored in system RAM, not VRAM). 6B active per token. Qwen4 architecture preview: sparse attention + Gated DeltaNet alternating, gated residuals. 262K native context (extendable to 1M via YaRN).", "hardware_requirement": "The 51B N-gram table requires ~96 GB system RAM. FP8 weights: ~186 GB VRAM. Q4 GGUF: ~82 GB. 1-bit: 75 GB. Fits on a sin
Created using ChatSlide
This project focuses on establishing model fit and deployment constraints by comparing developers, licenses, and architectures, while assessing risks. It further validates performance through benchmarking various workloads and estimating system capabilities on DGX H200. The outcome includes recommending a suitable model, a pilot plan, and defining success metrics.