{"generated_at":"2026-10-31T20:45:15Z","notes":"Boards group by HARDWARE configuration (card family at the same device count; RTX PRO 6000 Max-Q and plain RTX PRO 6000 share a family) \u2014 \"what is the best on this rig\" is the question. Every model with a recipe for that hardware competes together; entries are ranked by the requested ?sort= mode (aa = Artificial Analysis Intelligence of the base model DESC, null/un-scored models last, then C1 single-stream decode per GPU DESC \u2014 the default; decode = C1 decode per GPU DESC; c8 = the exact-C8 decode cell DESC (null-C8 trails, then the C1 decode fallback orders the null-C8 tail); prefill = whole-rig prefill DESC; ttft = TTFT at C1 ASC, lower is better; votes = community vote count DESC (recipe-grain \u2014 votes attach to recipes, one per anon identity); added = recipe creation date DESC, newest first (when it first got listed); updated = last real revision DESC (the same updated_at the recipe twins expose \u2014 a view/vote/import never re-sorts the board); max_context = the recipe's OWN DECLARED max_context_len DESC (t_703c7037 recipe metadata carried on every entry as max_context_len \u2014 NOT a measurement; rows without the declaration sink LAST via null-trail, never first); evidence = measured run file before author-reported) plus a secondary ?then= level applied immediately after the primary (default then=decode, so the default chain is AA intelligence then decode speed) with evidence (measured before reported) and slug as final tiebreaks; board sections and entry order are identical to /leaderboard?sort=<same value> (entry order IS the rank \u2014 there is no rank field). Board SECTION order is hardware-only (generation, family, device count) and never changes with the sort. Boards carry their base model name plus its Artificial Analysis Intelligence index (the quality axis) with aa_index_version/aa_source_url provenance. evidence: reported in the UI for all rows; run-file-backed rows are marked in tooltips (API evidence values: measured = run file on record, reported = author-reported). aa_intelligence is the Artificial Analysis index of the base model (null = not scored); aa_index_version / aa_source_url give the version and page every value was read from. aa_source_url is the source of the EFFECTIVE aa value (measured rows: the recorded page; published rows: the register snapshot_url \u2014 the exact href the HTML AA cell renders; the raw published snapshot also ships as aa_published_source_url). aa is the EFFECTIVE AA value (t_b3adaed5): COALESCE(aa_intelligence, aa_published_index) \u2014 the exact number the HTML AA cell renders and the value sort=aa ranks on. Sorting entries by aa reproduces the page order exactly (null trails, as rendered). PROVENANCE TIERS (aa_provenance field): measured = a locally recorded index in aa_intelligence; published = no local record, the AA-published index is shown instead in aa_published_index with aa_published_methodology_version / aa_published_snapshot_url / aa_published_fetched_at provenance (from the append-only aa_scores register; newest fetched_at wins) \u2014 provenance is descriptive metadata on every row (it is NOT a rank level: a published index ranks by its number like any other), and carries in the number tooltip (the former visible AA published chip is retired for good \u2014 owner, Oct 6); unscored = neither, rendered n/a and trailing (unchanged honest trail).","meta":{"count":2,"sort":"aa","then":"decode","sort_chain":"aa>decode>prefill>evidence>slug","applied":{"hardware":["dgx-spark-x3","170hx-x6"]},"origin":"https://localairecipes.com","openapi":"/openapi.json"},"boards":[{"label":"NVIDIA DGX Spark 128GB \u00d73 (Blackwell)","card_family":"dgx-spark","representative_card":"dgx-spark","vendor":"nvidia","device_count":3,"generation":"blackwell","vram_gb":128,"entries":[{"slug":"glm-5-3-flash-exlr8-exl3-experts-on-dgx-spark-tp1-4","title":"GLM-5.3-Flash exlr8 EXL3 experts on DGX Spark TP1-4","url":"/r/glm-5-3-flash-exlr8-exl3-experts-on-dgx-spark-tp1-4","beta":false,"base_model":"DeepSeek-V3","aa_intelligence":null,"aa":null,"aa_index_version":null,"aa_source_url":null,"aa_recorded_at":null,"aa_provenance":"unscored","aa_published_index":null,"aa_published_methodology_version":null,"aa_published_snapshot_url":null,"aa_published_fetched_at":null,"aa_resolved_slug":null,"aa_published_source_url":null,"card_variant":"DGX Spark (GB10, unified memory)","decode_c1_tok_s":128,"decode_c4_tok_s":null,"decode_c8_tok_s":null,"decode_c16_tok_s":null,"prefill_tok_s":4393,"ttft_ms_c1":null,"other_concurrencies":[1],"max_context_len":524288,"evidence":"claimed","votes":0}]},{"label":"NVIDIA CMP 170HX 64GB \u00d76 (Ampere)","card_family":"170hx","representative_card":"CMP 170HX","vendor":"nvidia","device_count":6,"generation":"ampere","vram_gb":64,"entries":[{"slug":"deepseek-v4-1-flash-sm80-deepinfer-kernel-port-6x-a100-dspark-k5","title":"DeepSeek-V4.1-Flash SM80 DeepInfer kernel port 6x CMP 170HX DSpark K5","url":"/r/deepseek-v4-1-flash-sm80-deepinfer-kernel-port-6x-a100-dspark-k5","beta":false,"base_model":"DeepSeek-V4.1-Flash","aa_intelligence":39,"aa":39,"aa_index_version":"Index v4.3.2","aa_source_url":"https://artificialanalysis.ai/models/deepseek-v4-1-flash","aa_recorded_at":"2026-09-30 22:34:38.86447-06","aa_provenance":"measured","aa_published_index":45,"aa_published_methodology_version":"Index v4.3.2 (2026-10-18)","aa_published_snapshot_url":"https://artificialanalysis.ai/models/deepseek-v4-1-flash-1017","aa_published_fetched_at":"2026-10-20 21:49:11.309778-06","aa_resolved_slug":"deepseek-v4-1-flash","aa_published_source_url":"https://artificialanalysis.ai/models/deepseek-v4-1-flash-1017","card_variant":"cmp-170hx","decode_c1_tok_s":160.80000000000001136868377216160297393798828125,"decode_c4_tok_s":170.5,"decode_c8_tok_s":null,"decode_c16_tok_s":null,"prefill_tok_s":4506.8000000000001818989403545856475830078125,"ttft_ms_c1":10700,"other_concurrencies":[1,4],"max_context_len":1048576,"evidence":"claimed","votes":0}]}]}