Selective Sources & Evidence Directory
A selective directory of documentation references, official publisher model cards, secondary news reports, and runtime documentation cited in the project data files.
Data Confidence & Evidence Categories
The planner references a combination of official publisher documents, secondary technical press coverage, and runtime specifications. Each source cited in data/sources.json is tagged with an explicit classification:
- measured Empirically Verified: Derived from published hardware technical specifications or verified model cards.
- derived First-Principles Derived: Mathematically calculated from physical bus specifications (e.g. memory bus width × data rate).
- vendor-claimed / news-report Secondary Press / Pre-Launch: Information sourced from pre-launch press reports or manufacturer announcements. Not an independent laboratory benchmark.
Boundary Notice: Citing a model artifact or runtime documentation establishes software artifact availability and compatibility; it does not constitute an empirical device-level performance benchmark on target hardware.
Hardware Specifications & Secondary Press References
Official Apple hardware documentation confirming M2 Ultra 800 GB/s peak theoretical memory bandwidth and up to 192 GB unified memory. macOS default 75% wired memory limit for single-process GPU allocations is incorporated into addressable memory limits. Cited in data/sources.json under key apple-mac-studio-m2-ultra.
Secondary press report covering pre-launch claims for Sipeed MaixStation (Axera AX8860P with 128GB LPDDR5X at 273 GB/s) and comparing it to NVIDIA DGX Spark and AMD Ryzen AI Halo. Cited in data/sources.json under key toms-hardware-sipeed-oct-2026. Data in data/devices.json derives from these reported bus specifications:
- Sipeed MaixStation: Axera AX8860P, 128GB LPDDR5X, 273 GB/s (vendor-claimed pre-launch spec).
- NVIDIA DGX Spark: 256-bit LPDDR5X, 273 GB/s (derived theoretical spec).
- AMD Ryzen AI Halo Max+ 395: 256-bit LPDDR5X-8000, 256 GB/s (derived theoretical spec).
- AMD Ryzen AI Halo PRO 495: 256-bit LPDDR5X-8533, 273 GB/s (derived theoretical spec).
Publisher Model Cards & Architecture Documents
Official Meta AI documentation establishing 128k (131,072) context window limit, 80 transformer layers, 64 attention heads, 8 KV heads (Grouped-Query Attention), and 128 head dimension. Cited under key meta-llama-3-1-config.
Official Qwen / Alibaba Cloud model card verifying 128k (131,072) base context window, 64 layers, 40 attention heads, 8 KV heads, and 128 head dimension. Cited under key qwen-2-5-32b-config.
Table 1 documents 131,072 context window, 116.83B total parameters (114.71B MoE in MXFP4, 0.96B dense attention, 1.16B embedding/unembedding), 5.13B active parameters, and an official 60.8 GiB checkpoint size reference. Cited under keys openai-gpt-oss-120b and openai-gpt-oss-announcement.
Official DeepSeek model card establishing 32k (32,768) native context length and Multi-Head Latent Attention (MLA) architecture parameters ($d_c = 512, d_R = 64$). Cited under key deepseek-v2-lite-config.
Quantized Model Artifact Repositories
Audited quantized GGUF artifacts for base Meta-Llama-3.1 variants, verifying availability of Q4_K_M, Q5_K_M, Q6_K, and Q8_0 files. Cited under keys mradermacher-llama-3-1-70b-gguf and mradermacher-llama-3-1-8b-gguf.
Official Qwen 2.5 32B GGUF repository. Model card notes context window is limited to 32k (32,768) in standard GGUF distributions compared to 128k base model. Cited under key qwen-2-5-32b-instruct-gguf.
Official model repository documenting native MXFP4 checkpoint and public inference serving recipes via vLLM and Hugging Face Transformers. Cited under key openai-gpt-oss-120b-repo.
Inference Engines & Runtime Specifications
Official documentation verifying GGUF model execution and supported Key-Value cache quantization options (--cache-type-k and --cache-type-v supporting f16 and q8_0). Cited under key llamacpp-server-runtime.
Verifies native architecture support and standard 16-bit (auto / FP16 / BF16) KV cache execution for Llama, Qwen 2.5, DeepSeek V2, and GPT-OSS causal language model families. Cited under key vllm-supported-models-runtime.
Uncalibrated community reports and forum discussions. Because these reports lack standardized test harnesses and environment controls, figures are downgraded to estimated and uncalibrated prefill metrics remain suppressed. Cited under key community-llama-bench-calibration.