{ "format": "neural-download-model-family-v1", "id": "deepseek-v4", "primary_packet_id": "deepseek-v4-flash-k160-b70-80tps-20260718", "name": "DeepSeek V4", "display_name": "DeepSeek V4 · Flash 180B", "speedup_label": "target / MTP0–4 / DSpark5–8", "summary": "DeepSeek's large mixture-of-experts reasoning model in a community-trimmed 180B form (about 13B active per word). A research-grade deployment: it needs all four B70 cards and an experimental checkpoint, so treat it as a look at what the biggest open models can do on this hardware, not a daily driver.", "updated_at": "2026-08-28", "architecture": { "class": "DeepSeek V4 Flash 180B sparse MoE", "design": "experimental uniform-K160 target with DSpark7 draft", "evidence": "results/deepseek-v4-flash-k160-b70/README.md" }, "dimensions": { "weight_revision": ["deepseek-v4-flash-180b-community"], "weight_quantization": ["experimental uniform-K160 FP8/MXFP4 target"], "runtime": [ "vLLM XPU target-only stack a681dbb/6522849/48fda4f", "vLLM XPU MTP1 stack 4a6fd87/18a44f4/48fda4f", "vLLM XPU record stack 264c7f2/3131567/48fda4f", "vLLM XPU DSpark DEV width stack 2289554/3131567/48fda4f" ], "tp": [1, 2, 4], "ep": [4], "speculative_method": ["target-only", "MTP1", "MTP2", "MTP3", "MTP4", "DSpark", "DSpark7"], "draft_depth": [0, 1, 2, 3, 4, 5, 6, 7, 8], "graph_mode": ["PIECEWISE"], "kv": ["fp8"] }, "weight_revisions": [ { "id": "deepseek-v4-flash-180b-community", "label": "DeepSeek V4 Flash 180B · community base identity", "role": "parent model identity", "repository": "0xSero/DeepSeek-V4-Flash-180B", "revision_status": "The measured K160 export identifies this community model repository, but no separately pinned pre-K160 parent revision was retained.", "quantized_artifacts": [ { "id": "deepseek-v4-flash-180b-k160-7c360e1", "label": "Experimental uniform-K160 FP8/MXFP4 target export", "quantization": "experimental uniform-K160 FP8/MXFP4 target", "quantization_origin": "export", "repository": "0xSero/DeepSeek-V4-Flash-180B", "revision": "7c360e1cd4a5168099dbc54d16d929bf6df04990", "evidence": "experiments/deepseek-v4-flash-reap-xpu-b70/data/dspark-sharded-target-argmax-record-20260718.json" } ] } ], "model_variants": [], "transfer_scope": { "status": "The experimental K160 export is a quantized child artifact of the DeepSeek V4 Flash community model identity, not a separate model. Measurements remain artifact- and runtime-specific.", "transfers": [ "the archived vLLM, XPU-kernel, and oneCCL source bundles and the target-verified DSpark7 mechanism", "the fail-closed launcher and ordered exact-gate workflow" ], "does_not_transfer": [ "performance to an official REAP construction, another K value, or another runtime revision", "the unavailable K160 pruning calibration or ranking provenance", "the single-active-generation result to aggregate serving throughput" ], "evidence": "repro/deepseek-v4-flash-k160-b70-80tps-20260718/README.md" }, "model_signals": { "b70_fit": { "band": "four-card measured", "scope": "TP4 plus expert parallelism on four 32 GiB B70 cards", "basis": "The exact experimental target and DSpark7 record stack completed three strict suites on four B70s.", "reviewed_at": "2026-08-24" }, "quality_evidence": { "band": "target-verified exact gates", "scope": "36/36 cache-zero realistic requests and 24/24 ordered exact canaries; checkpoint-construction provenance remains incomplete", "evidence": [ "experiments/deepseek-v4-flash-reap-xpu-b70/data/dspark-sharded-target-argmax-record-20260718.json" ] }, "popularity": { "state": "not-scored", "reason": "No dated popularity snapshot is stored for this exact experimental checkpoint." } }, "run_measurements": [ { "id": "deepseek-v4-k160-tp4-ep-target-only-strict", "state": "lab-measured", "revision": "deepseek-v4-flash-180b-community", "artifact_id": "deepseek-v4-flash-180b-k160-7c360e1", "variant": "experimental uniform-K160 FP8/MXFP4 target", "quantization": "experimental uniform-K160 FP8/MXFP4 target", "runtime": "vLLM XPU target-only stack a681dbb/6522849/48fda4f", "runtime_identity": { "vllm": "a681dbb2b4b19c2c5a964817095b5f8c1f27ff48", "xpu_kernels": "6522849b02894273b1e779b3c115527b5cdf3756", "oneccl": "48fda4f0e074db005596d6899d5227d3f0316c12" }, "speculative_method": "target-only", "config": {"tp": 4, "ep": 4, "draft_depth": 0, "verifier_width": 1, "graph": "piecewise", "graph_mode": "PIECEWISE", "kv": "fp8"}, "profile_id": "deepseek-v4-k160-target-only-tp4-record-v1", "measurement_class": "four independent strict suites", "promotion_status": "qualified historical target-only record", "quality_scope": "48/48 strict rows cache-zero; 70/70 exact graph captures", "workload": "12 unique cold prompts per suite, 128 output tokens, median generated tokens 1-100 after TTFT; four strict suites", "metrics": { "decode_tok_s": [43.766673266271965, 43.69855047858043, 43.69421019994415, 43.66790844592444], "wall_output_tok_s": [39.28124709626695, 38.98299402541498], "ttft_ms": [308.684051502496, 330.7663530067657] }, "sample_annotations": [ {"metric": "decode_tok_s", "index": 0, "value": 43.766673266271965, "label": "qualified target-only high"} ], "aggregation": {"decode_tok_s": "four retained strict-suite medians; qualified high 43.766673266271965"}, "evidence": "experiments/deepseek-v4-flash-reap-xpu-b70/data/m1-direct-routed-moe-wideepoch-record-20260715.json" }, { "id": "deepseek-v4-k160-tp4-ep-mtp1-strict", "state": "lab-measured", "revision": "deepseek-v4-flash-180b-community", "artifact_id": "deepseek-v4-flash-180b-k160-7c360e1", "variant": "experimental uniform-K160 FP8/MXFP4 target", "quantization": "experimental uniform-K160 FP8/MXFP4 target", "runtime": "vLLM XPU MTP1 stack 4a6fd87/18a44f4/48fda4f", "runtime_identity": { "vllm": "4a6fd874725312c53883b1d53970af1d0eccfc3f", "xpu_kernels": "18a44f440ca3ac2006d5ba19cd12ccca0a0c9982", "oneccl": "48fda4f0e074db005596d6899d5227d3f0316c12" }, "speculative_method": "MTP1", "config": {"tp": 4, "ep": 4, "draft_depth": 1, "verifier_width": 2, "graph": "piecewise", "graph_mode": "PIECEWISE", "kv": "fp8"}, "profile_id": "deepseek-v4-k160-mtp1-tp4-record-v1", "measurement_class": "candidate-control-candidate portfolio gate", "promotion_status": "qualified historical MTP1 record", "quality_scope": "36 qualifying rows cache-zero; 70/70 ordered exact suites; unchanged target verifies accepted tokens", "workload": "12 unique cold prompts, 128 output tokens, median generated tokens 1-100 after TTFT; candidate-control-candidate portfolio gate", "metrics": { "decode_tok_s": [62.51566129683944, 63.85130111953857], "wall_output_tok_s": [53.22380403046468], "ttft_ms": [334.09914300136734] }, "sample_annotations": [ {"metric": "decode_tok_s", "index": 1, "value": 63.85130111953857, "label": "qualified MTP1 record candidate"} ], "aggregation": {"decode_tok_s": "two candidate arms around a same-binary control; qualified high 63.85130111953857"}, "evidence": "experiments/deepseek-v4-flash-reap-xpu-b70/data/qnorm-routeportfolio-20260716/summary.json" }, { "id": "deepseek-v4-k160-tp4-ep-dspark7-strict", "state": "lab-measured", "revision": "deepseek-v4-flash-180b-community", "artifact_id": "deepseek-v4-flash-180b-k160-7c360e1", "variant": "experimental uniform-K160 FP8/MXFP4 target", "quantization": "experimental uniform-K160 FP8/MXFP4 target", "runtime": "vLLM XPU record stack 264c7f2/3131567/48fda4f", "runtime_identity": { "vllm": "264c7f2f7df21ddeeab32ecca0353133344f1ac9", "xpu_kernels": "31315673737d95da0f79179c8f755260ef02c1d6", "oneccl": "48fda4f0e074db005596d6899d5227d3f0316c12" }, "speculative_method": "DSpark7", "config": {"tp": 4, "ep": 4, "draft_depth": 7, "graph": "piecewise", "graph_mode": "PIECEWISE", "kv": "fp8"}, "profile_id": "deepseek-v4-k160-dspark7-tp4-record-v1", "measurement_class": "three independent strict suites", "promotion_status": "closed-frontier record", "quality_scope": "target-verified accepted tokens; 36/36 cache-zero rows and 24/24 exact canaries", "workload": "12 unique cold prompts per suite, 128 output tokens, median generated tokens 1-100 after TTFT; three strict suites", "metrics": { "decode_tok_s": [80.82005189243556, 76.90017809136465, 78.28722593298039], "ttft_ms": [342.77765700244345] }, "sample_annotations": [ {"metric": "decode_tok_s", "index": 0, "value": 80.82005189243556, "label": "record strict suite"}, {"metric": "decode_tok_s", "index": 2, "value": 78.28722593298039, "label": "third strict suite and median-of-medians"} ], "aggregation": {"decode_tok_s": "median-of-medians 78.28722593298039; record high retained separately"}, "evidence": "experiments/deepseek-v4-flash-reap-xpu-b70/data/dspark-sharded-target-argmax-record-20260718.json" } ], "series_measurements": [ { "id": "deepseek-v4-k160-tp4-dspark-dev-width-screen", "state": "lab-screened", "revision": "deepseek-v4-flash-180b-community", "artifact_id": "deepseek-v4-flash-180b-k160-7c360e1", "variant": "experimental uniform-K160 FP8/MXFP4 target", "quantization": "experimental uniform-K160 FP8/MXFP4 target", "runtime": "vLLM XPU DSpark DEV width stack 2289554/3131567/48fda4f", "runtime_identity": { "vllm": "22895545b96c9e16a7604db925dc05b531dbc0d0", "xpu_kernels": "31315673737d95da0f79179c8f755260ef02c1d6", "oneccl": "48fda4f0e074db005596d6899d5227d3f0316c12" }, "speculative_method": "DSpark", "profile_id": "deepseek-v4-k160-dspark-width-dev-v1", "measurement_class": "public-plus-DEV same-binary width screen", "promotion_status": "DEV-only; no held-out reveal and no record candidate", "quality_scope": "20/22 prompts reached the timing window; all completed policies passed ordered exact canaries and cache-zero checks", "config": {"tp": 4, "ep": 4, "graph": "piecewise", "graph_mode": "PIECEWISE", "kv": "fp8", "max_model_len": 256, "max_num_batched_tokens": 256}, "workload": "12 public continuity plus 10 explicitly DEV prompts; 20 rows reached the 100-token timing window; one active generation", "axis": "draft_depth", "points": [ {"x": 5, "decode_tok_s": 76.314666, "effective_tokens_per_verification": 3.265476}, {"x": 6, "decode_tok_s": 72.803326, "effective_tokens_per_verification": 3.465234}, {"x": 7, "decode_tok_s": 79.801765, "effective_tokens_per_verification": 3.507042}, {"x": 8, "decode_tok_s": 73.834972, "effective_tokens_per_verification": 3.470812} ], "quality": "DEV diagnostic only. M7 remained 1.26% below the protected 80.820052 qualified record, so no held-out candidate or submission followed.", "evidence": "experiments/deepseek-v4-flash-reap-xpu-b70/notes/2026-07-19-dspark-no-training-speculation-screen.md" } ], "estimates": [], "packets": [ { "id": "deepseek-v4-flash-k160-b70-80tps-20260718", "label": "DeepSeek V4 Flash K160 · TP4+EP DSpark7", "revision": "deepseek-v4-flash-180b-community", "artifact_id": "deepseek-v4-flash-180b-k160-7c360e1", "quantization": "experimental uniform-K160 FP8/MXFP4 target", "runtime": "vLLM XPU record stack 264c7f2/3131567/48fda4f", "cards": 4, "status": "closed research frontier", "evidence_level": "B70-verified experimental record", "coverage": ["TP4+EP", "DSpark7", "decode", "TTFT", "exact canaries", "reproduction"], "grades": { "evidence": { "grade": "C", "scope": "exact experimental uniform-K160 TP4+EP DSpark7 lane", "basis": "strong exact lab and replay material, but the public K160 checkpoint lacks calibration/ranking provenance and is not an official REAP construction", "reviewed_at": "2026-08-24", "evidence": ["results/deepseek-v4-flash-k160-b70/README.md"] } }, "featured_metric": { "metric": "decode_tok_s", "measurement_id": "deepseek-v4-k160-tp4-ep-dspark7-strict", "sample_index": 2, "value": 78.28722593298039, "unit": "tok/s", "workload": "12 unique cold prompts per suite, 128 output tokens, median generated tokens 1-100 after TTFT; three strict suites", "evidence": "experiments/deepseek-v4-flash-reap-xpu-b70/data/dspark-sharded-target-argmax-record-20260718.json" }, "manifest": "results/deepseek-v4-flash-k160-b70/README.md", "guide": "repro/deepseek-v4-flash-k160-b70-80tps-20260718/README.md" } ], "views": [ { "id": "deepseek-v4-strict-tp4", "title": "Strict TP4+EP DSpark7 suites", "subtitle": "Decode retains three independent suite medians (76.900178–80.820052); the packet headline is the 78.287226 median-of-medians, while TTFT is the 342.777657 record-suite observation", "x_label": "tensor parallel cards", "discrete": true, "missing_x": [1, 2], "metrics": ["decode_tok_s", "ttft_ms"], "series": [ {"label": "experimental K160 + DSpark7", "measurement_ids": ["deepseek-v4-k160-tp4-ep-dspark7-strict"], "x_from": "config.tp"} ] }, { "id": "deepseek-v4-dspark-dev-width", "title": "DSpark width screen", "subtitle": "Same patched DEV binary and K160 target · TP4+EP4 · PIECEWISE · FP8 KV · public+DEV prompts · screened, not promotion evidence", "x_label": "DSpark draft tokens", "discrete": true, "metrics": ["decode_tok_s", "effective_tokens_per_verification"], "series": [ {"label": "DSpark DEV widths", "measurement_ids": ["deepseek-v4-k160-tp4-dspark-dev-width-screen"]} ] } ], "coverage_views": [ { "id": "deepseek-v4-k160-mtp-by-tp", "label": "attached MTP depth × TP", "fixed": "Exact experimental K160 artifact · PIECEWISE · FP8 KV. Historical TP4 rows use EP4 and retain their exact differing runtime identities; rates are not transferred across them. The approximately 103.1 GB on-disk target does not fit fully resident at TP1 or TP2 on 32/64 GiB; CPU offload is outside this view.", "decision_note": "MTP depth counts drafted tokens; verifier width is depth + 1. MTP3 is the rejected three-draft/four-row diagnostic. The evidence-based post-MTP3 stop decision left MTP4 unlaunched.", "row_axis": {"key": "speculative_method", "label": "Method", "prefix": ""}, "column_axis": {"key": "tp", "label": "TP", "prefix": "TP"}, "fixed_selectors": { "revision": "deepseek-v4-flash-180b-community", "artifact_id": "deepseek-v4-flash-180b-k160-7c360e1", "variant": "experimental uniform-K160 FP8/MXFP4 target", "quantization": "experimental uniform-K160 FP8/MXFP4 target", "graph_mode": "PIECEWISE", "kv": "fp8" }, "rows": ["target-only", "MTP1", "MTP2", "MTP3", "MTP4"], "columns": [1, 2, 4], "cells": { "target-only:1": {"state": "unsupported", "label": "weights do not fit", "reason": "The approximately 103.1 GB on-disk K160 target cannot fit fully resident on one 32 GiB B70; CPU offload is outside this view."}, "target-only:2": {"state": "unsupported", "label": "weights do not fit", "reason": "The approximately 103.1 GB on-disk K160 target cannot fit fully resident on two 32 GiB B70s; CPU offload is outside this view."}, "target-only:4": {"state": "lab-measured", "label": "43.767 qualified high", "evidence_id": "deepseek-v4-k160-tp4-ep-target-only-strict"}, "MTP1:1": {"state": "unsupported", "label": "weights do not fit", "reason": "The fully resident target already exceeds one B70 before the attached MTP layer is considered."}, "MTP1:2": {"state": "unsupported", "label": "weights do not fit", "reason": "The fully resident target already exceeds two B70s before the attached MTP layer is considered."}, "MTP1:4": {"state": "lab-measured", "label": "63.851 qualified high", "evidence_id": "deepseek-v4-k160-tp4-ep-mtp1-strict"}, "MTP2:1": {"state": "unsupported", "label": "weights do not fit", "reason": "The fully resident target already exceeds one B70; repeated-layer drafting cannot change target capacity."}, "MTP2:2": {"state": "unsupported", "label": "weights do not fit", "reason": "The fully resident target already exceeds two B70s; repeated-layer drafting cannot change target capacity."}, "MTP2:4": {"state": "closed", "label": "no valid rate", "reason": "M=3 capture and 10/10 exact pre-hang rows passed, but the second draft accepted only 0.5–2.2% and realistic traffic stopped before a valid suite completed."}, "MTP3:1": {"state": "unsupported", "label": "weights do not fit", "reason": "The fully resident target already exceeds one B70; repeated-layer drafting cannot change target capacity."}, "MTP3:2": {"state": "unsupported", "label": "weights do not fit", "reason": "The fully resident target already exceeds two B70s; repeated-layer drafting cannot change target capacity."}, "MTP3:4": {"state": "quarantined", "label": "D46.247 · incomplete", "reason": "Only 8/10 rows reached the timing window; the third draft accepted 0–3.2%, so the 46.247281 diagnostic is not a valid endpoint result.", "evidence": "experiments/deepseek-v4-flash-reap-xpu-b70/data/mwidth-sequential-verifier-20260718.json"}, "MTP4:1": {"state": "unsupported", "label": "weights do not fit", "reason": "The fully resident target already exceeds one B70; repeated-layer drafting cannot change target capacity."}, "MTP4:2": {"state": "unsupported", "label": "weights do not fit", "reason": "The fully resident target already exceeds two B70s; repeated-layer drafting cannot change target capacity."}, "MTP4:4": {"state": "closed", "label": "not launched", "reason": "The evidence-based post-MTP3 stop decision closed four drafted tokens after the third proposal accepted at most 3.2%. No MTP4 throughput exists."} } }, { "id": "deepseek-v4-spec-by-tp", "label": "speculation × TP", "fixed": "Exact experimental K160 target and record runtime; missing cells have no inherited result.", "row_axis": {"key": "speculative_method", "label": "Method", "prefix": ""}, "column_axis": {"key": "tp", "label": "TP", "prefix": "TP"}, "fixed_selectors": { "revision": "deepseek-v4-flash-180b-community", "artifact_id": "deepseek-v4-flash-180b-k160-7c360e1", "variant": "experimental uniform-K160 FP8/MXFP4 target", "runtime": "vLLM XPU record stack 264c7f2/3131567/48fda4f", "graph_mode": "PIECEWISE" }, "rows": ["DSpark7"], "columns": [1, 2, 4], "cells": { "DSpark7:1": {"state": "unsupported", "label": "weights do not fit", "reason": "The approximately 103.1 GB on-disk K160 target cannot fit fully resident on one 32 GiB B70, before the DSpark draft pack is considered; CPU offload is outside this view."}, "DSpark7:2": {"state": "unsupported", "label": "weights do not fit", "reason": "The approximately 103.1 GB on-disk K160 target cannot fit fully resident on two 32 GiB B70s, before the DSpark draft pack is considered; CPU offload is outside this view."}, "DSpark7:4": {"state": "lab-measured", "label": "78.287 MoM; 80.820 high", "evidence_id": "deepseek-v4-k160-tp4-ep-dspark7-strict", "packet_id": "deepseek-v4-flash-k160-b70-80tps-20260718"} } } ], "family_closures": [ { "selectors": {"revision": "deepseek-v4-flash-180b-community", "artifact_id": "deepseek-v4-flash-180b-k160-7c360e1", "speculative_method": "MTP2", "draft_depth": 2, "tp": 4}, "state": "closed", "reason": "Repeated use of the attached single MTP layer stopped before a valid suite completed; its second proposal accepted only 0.5–2.2% on realistic traffic.", "evidence": "experiments/deepseek-v4-flash-reap-xpu-b70/notes/2026-07-15-mtp2-reuse-deadlock-closure.md" }, { "selectors": {"revision": "deepseek-v4-flash-180b-community", "artifact_id": "deepseek-v4-flash-180b-k160-7c360e1", "speculative_method": "MTP3", "draft_depth": 3, "tp": 4}, "state": "quarantined", "reason": "The three-draft/four-row endpoint produced only eight eligible timing rows at 46.247281 tok/s; the third proposal accepted 0–3.2%, so the incomplete screen is diagnostic only.", "evidence": "experiments/deepseek-v4-flash-reap-xpu-b70/data/mwidth-sequential-verifier-20260718.json" }, { "selectors": {"revision": "deepseek-v4-flash-180b-community", "artifact_id": "deepseek-v4-flash-180b-k160-7c360e1", "speculative_method": "MTP4", "draft_depth": 4, "tp": 4}, "state": "closed", "reason": "True MTP4 was not launched because the evidence-based post-MTP3 decision rejected deeper repeated-layer work after near-zero third-proposal acceptance.", "evidence": "experiments/deepseek-v4-flash-reap-xpu-b70/notes/2026-07-18-sequential-mwidth-verifier-and-predictor-pivot.md" }, { "selectors": {"revision": "deepseek-v4-flash-180b-community", "artifact_id": "deepseek-v4-flash-180b-k160-7c360e1", "frontier": "configuration-sweeps"}, "state": "closed", "reason": "The configuration frontier is paused; reopen only for a substantial EAGLE capture/training effort or a genuinely new mechanism, not another generic flag sweep.", "evidence": "experiments/deepseek-v4-flash-reap-xpu-b70/notes/2026-07-21-deepseek-v4-flash-frontier-closeout.md" } ] }