Blackwellboy / Models
This is not a logo wall and it is not a leaderboard. The status on each card says how deep the public evidence goes: full deployment, full campaign, targeted test or control. If the evidence is narrower, the claim is narrower.
The models with the strongest public trail: bring-up work, repeatable measurements, long runs or dedicated repositories.
Multi-Spark bring-up of the Neuron IQ1_S GGUF with SparkInfer. The public recipe documents working TP3 and TP4 paths, rank-local NVMe loading, NCCL fabric behaviour and the geometry behind the 3-Spark 512/512/512 FFN split.
Full 428B model on two stock-driver DGX Sparks. The public lane work separates the b12x first-token failure from the KVarN paths that actually run, and measures code, structured, prose and deep-context behaviour rather than reporting one speed number.
Dual-Spark NVFP4 serving with MTP, FP8 KV and structured tool calling. The campaign includes throughput, memory, cold-start and an accuracy slice instead of treating successful loading as proof that the quant is good.
A two-Spark production lane used to study speculative acceptance, long-context behaviour, cache effects, templates and reasoning round-trips. The public page explicitly separates the community-abliterated checkpoint from stock DeepSeek and records where the advertised window stops being useful.
The deepest single-model research programme in the public repos: tuning sweep, 12-hour soak, reasoning-gate work, quant-floor comparisons, HumanEval+ replication, multi-turn history experiments, prompt-topology tests and cross-model controls.
A 13-hour single-Spark soak used to test stability, thinking-on/off behaviour, speculative counters and memory/acceptance claims over time. The public evidence keeps the limitations beside the result instead of turning the run into a general benchmark.
Useful first-hand evidence without pretending every model below received the same depth of campaign.
NVFP4 and bf16 lanes used for head-to-head quality, reasoning-gate controls, ceiling behaviour, MoE routing and harness checks.
llama.cpp and NVFP4/dequant work around generation defaults, thinking controls, budget floors and quantisation fidelity.
GGUF / llama.cpp checks for generation-config defaults and model-family output-budget assumptions.
Ollama-on-GB10 campaign covering release assets, device logs, request-field validation, tool choice and out-of-range context handling.
Apple-silicon quant-floor work comparing 1-bit and 2-bit ternary builds across a blinded 512-case aggregate battery plus performance and robustness tests.
llama.cpp + DFlash checks around reasoning control, tools, harness scoring and speculative-decode measurement on a 5090-class lane.
Reasoning round-trip, parser packaging, MTP configuration, memory and serving-control checks across vLLM and contributor-measured SGLang evidence.
Multimodal serving checks around reasoning toggles, history rendering, media fetches and multimodal token accounting.
llama.cpp template and tool-roundtrip forensics: system relocation, baked defaults, final-turn branches and thinking-type validation.
The full Minefield model map also contains upstream reports and contributor evidence. Those are deliberately not relabelled here as my first-hand work.