{ "schema_version": 1, "status": "FROZEN", "study": { "title": "Strata on a 12 GB gaming GPU and 32 GB RAM", "run_date": "2026-10-17 through 2026-10-18 UTC", "hardware": { "gpu": "Radeon RX 6700 XT", "vram_capacity_gb": 12, "cpu": "Ryzen 5 3600", "cpu_cores": 6, "cpu_threads": 12, "ram_installed_gb": 32, "ram": "2 x 16 GB Corsair CMK32GX4M2E3200C16 DDR4-3200", "ram_configured_mt_s": 3200, "motherboard": "ASUS TUF GAMING X570-PLUS (WI-FI)", "ssd": "Intel SSDPEKNW010T8 1 TB NVMe", "host_os": "Proxmox VE 9.2.21 / Debian 13, Linux 7.0.14-24-pve", "environment": "LXC guest with isolated Docker inference and scoring", "guest_ram_limit_gib": 28, "inference_ram_limit_gib": 26, "inference_cpu_limit": 9, "benchmark_disk_gib": 400 }, "runtime": { "name": "Strata", "version": "0.1.45", "backend": "HIP gfx1031", "source_commit": "477ac13cbe8e582c20484871ce459d20d82df09d", "engine_sha256": "661d6e28a2552bad8ff29ad980771e3a69d8f4cd2a8fb3ef9b67446f9a22b0ca", "engine_build": { "source": "local-hip", "backend": "hip", "version": "0.1.45", "archs": [ "gfx1031" ], "vision": "none", "lib_dirs": [ "/upstream/.venv/lib/python3.13/site-packages/_rocm_sdk_devel/lib", "/upstream/.venv/lib/python3.13/site-packages/_rocm_sdk_libraries_gfx103X_all/lib" ], "src": "b95137167bd9b751" }, "inference_image": "sha256:28e04eca47e22b9bee47a6a2ec4260bfa25fce525d089b5151e07952cdad3f23", "scorer_image": "sha256:be7f04ee1008c85d34f7512cedc0bd530145ba4022b6c640804a1f8611a216c0" }, "llama_cpp_commit": "3cf03257f219afbe7334045ff7c6a06ac68c627d", "llama_cpp_scope": "source dependency, not inference runtime", "cost_usd": null, "cost_scope": "Local hardware; rental and quant-creation spend $0. Electricity, internet and amortization unmeasured.", "costs": { "rental_usd": 0, "quant_creation_usd": 0, "electricity_usd": null, "idle_and_failed_run_electricity_usd": null, "total_project_usd": null } }, "methodology": { "protocol": "frozen v2.1", "practical": "24 tasks x 3 repetitions; 2048 output cap", "humaneval": "164 tasks, one greedy sample each; EvalPlus 0.3.1", "dossier": "3 repetitions; 3072 output cap", "humaneval_output_cap_tokens": 2048, "configured_reasoning_budget_tokens": 512, "custom_suite_temperature": 0, "custom_suite_seed": 12345, "needle": "Unmodified upstream needle_bench.py; greedy, thinking off, three depths 10/50/90, seeded needle placement 7", "scoring": "Isolated network-disabled Docker scorer; generated code never evaluated on the host", "official": "Unmodified upstream tools/strata_bench.py, runs=3; extended tests use --long", "memory_measurement": "Sampled docker stats cgroup memory usage in GiB, including counted cache; no peak claim", "units": "Canonical binary memory units are GiB; video workflow displays the same numeric binary readings as GB. Model file GB is decimal." }, "display_precision": { "score_decimals": 2, "decode_tps_table_decimals": 1, "chart_value_decimals": 2, "ttft_seconds_decimals": 3, "actual_prompt_tokens_decimals": 0, "memory_gib_decimals": 2 }, "weighting": { "business": 20, "research": 20, "knowledge": 20, "agent_tools": 15, "practical_sysadmin": 5, "humaneval_plus": 10, "executive_content": 10 }, "role_weights": { "CFO / Analyst": { "business": 50, "research": 20, "knowledge": 15, "agent_tools": 10, "executive_content": 5 }, "Researcher / Scout": { "research": 40, "knowledge": 25, "agent_tools": 15, "executive_content": 10, "business": 10 }, "Hermes Coordinator": { "agent_tools": 30, "knowledge": 25, "research": 20, "executive_content": 15, "practical_sysadmin": 5, "humaneval_plus": 5 }, "Librarian / Knowledge": { "knowledge": 35, "research": 25, "agent_tools": 20, "executive_content": 15, "business": 5 }, "Forge / Coder": { "humaneval_plus": 50, "practical_sysadmin": 20, "agent_tools": 15, "knowledge": 10, "research": 5 } }, "models": [ { "id": "coder-iq1_m", "label": "Flash-Next Coder IQ1_M", "quality_scope": "Full frozen v2.1 suite", "balanced_score": 95.08, "humaneval_plus_percent": 92.68, "role_scores": { "CFO / Analyst": 97.53, "Researcher / Scout": 95.94, "Hermes Coordinator": 93.34, "Librarian / Knowledge": 94.35, "Forge / Coder": 91.92 }, "contexts": [ { "allocated_tokens": 131072, "allocation_passed": true, "recall": [ { "length": "126k", "depth": 10, "prompt_tokens": 125572, "seconds": 389.5, "found": true, "word": "meadow-copper-504", "answer": "meadow-copper-504" }, { "length": "126k", "depth": 50, "prompt_tokens": 125572, "seconds": 389.6, "found": true, "word": "falcon-quartz-940", "answer": "falcon-quartz-940" }, { "length": "126k", "depth": 90, "prompt_tokens": 125573, "seconds": 224.5, "found": true, "word": "willow-cobalt-696", "answer": "willow-cobalt-696" }, { "length": "64k", "depth": 10, "prompt_tokens": 66168, "seconds": 204.3, "found": true, "word": "meadow-copper-504", "answer": "meadow-copper-504" }, { "length": "64k", "depth": 50, "prompt_tokens": 66168, "seconds": 201.2, "found": true, "word": "falcon-quartz-940", "answer": "falcon-quartz-940" }, { "length": "64k", "depth": 90, "prompt_tokens": 66169, "seconds": 147.0, "found": true, "word": "willow-cobalt-696", "answer": "willow-cobalt-696" } ], "fully_passed_recall_tiers": [ "126k", "64k" ], "max_recall_validated_prompt_tokens": 125573, "status": "allocation_passed" }, { "allocated_tokens": 16384, "allocation_passed": true, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "allocation_passed" }, { "allocated_tokens": 32768, "allocation_passed": true, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "allocation_passed" }, { "allocated_tokens": 65536, "allocation_passed": true, "recall": [ { "length": "16k", "depth": 10, "prompt_tokens": 15156, "seconds": 43.9, "found": true, "word": "meadow-copper-504", "answer": "meadow-copper-504" }, { "length": "16k", "depth": 50, "prompt_tokens": 15156, "seconds": 44.3, "found": true, "word": "falcon-quartz-940", "answer": "falcon-quartz-940" }, { "length": "16k", "depth": 90, "prompt_tokens": 15157, "seconds": 44.1, "found": true, "word": "willow-cobalt-696", "answer": "willow-cobalt-696" }, { "length": "32k", "depth": 10, "prompt_tokens": 32792, "seconds": 96.5, "found": true, "word": "meadow-copper-504", "answer": "meadow-copper-504" }, { "length": "32k", "depth": 50, "prompt_tokens": 32792, "seconds": 96.5, "found": true, "word": "falcon-quartz-940", "answer": "falcon-quartz-940" }, { "length": "32k", "depth": 90, "prompt_tokens": 32793, "seconds": 96.7, "found": true, "word": "willow-cobalt-696", "answer": "willow-cobalt-696" }, { "length": "8k", "depth": 10, "prompt_tokens": 6920, "seconds": 23.0, "found": true, "word": "meadow-copper-504", "answer": "meadow-copper-504" }, { "length": "8k", "depth": 50, "prompt_tokens": 6920, "seconds": 20.7, "found": true, "word": "falcon-quartz-940", "answer": "falcon-quartz-940" }, { "length": "8k", "depth": 90, "prompt_tokens": 6921, "seconds": 20.5, "found": true, "word": "willow-cobalt-696", "answer": "willow-cobalt-696" } ], "fully_passed_recall_tiers": [ "16k", "32k", "8k" ], "max_recall_validated_prompt_tokens": 32793, "status": "allocation_passed" }, { "allocated_tokens": 8192, "allocation_passed": true, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "allocation_passed" } ], "official": [ { "allocated_tokens": 32768, "cohort": "official", "runs": 3, "elapsed_seconds": 433.8292200565338, "chat_decode_tps": { "median": 22.4, "min": 17.5, "max": 27.0, "n": 9 }, "chat_ttft_seconds": { "median": 1.499, "min": 1.281, "max": 1.784, "n": 9 }, "failures": [], "prefill": { "4K": { "tokens_per_second": { "median": 327.1, "min": 323.3, "max": 328.5, "n": 3 }, "ttft_seconds": { "median": 12.333, "min": 12.287, "max": 12.49, "n": 3 } }, "16K": { "tokens_per_second": { "median": 342.2, "min": 341.5, "max": 343.4, "n": 3 }, "ttft_seconds": { "median": 46.986, "min": 46.809, "max": 47.072, "n": 3 } } }, "agent_cache_reuse_percent": { "median": 68.793, "min": 68.473, "max": 68.834, "n": 3 }, "agent_decode_tps": { "median": 24.7, "min": 16.7, "max": 31.2, "n": 15 }, "agent_ttft_seconds": { "median": 9.383, "min": 5.442, "max": 10.889, "n": 15 }, "clients": {}, "drafts_offered": 1722, "drafts_accepted": 1043, "source": "source/selected/results/coder-iq1_m-ctx32768/out/official/2026-10-17-community-unknown-qwen3-8-flash-next-coder-iq1-m/results.json" }, { "allocated_tokens": 65536, "cohort": "official-optional", "runs": 3, "elapsed_seconds": 797.9292860031128, "chat_decode_tps": { "median": 21.2, "min": 19.2, "max": 26.2, "n": 9 }, "chat_ttft_seconds": { "median": 1.693, "min": 1.416, "max": 1.789, "n": 9 }, "failures": [], "prefill": { "4K": { "tokens_per_second": { "median": 321.3, "min": 318.8, "max": 323.2, "n": 3 }, "ttft_seconds": { "median": 12.558, "min": 12.488, "max": 12.67, "n": 3 } }, "16K": { "tokens_per_second": { "median": 341.5, "min": 331.0, "max": 341.7, "n": 3 }, "ttft_seconds": { "median": 47.077, "min": 47.05, "max": 48.564, "n": 3 } }, "32K": { "tokens_per_second": { "median": 337.5, "min": 309.6, "max": 338.2, "n": 3 }, "ttft_seconds": { "median": 95.045, "min": 94.843, "max": 103.633, "n": 3 } } }, "agent_cache_reuse_percent": { "median": 68.488, "min": 68.478, "max": 68.673, "n": 3 }, "agent_decode_tps": { "median": 24.5, "min": 11.7, "max": 27.7, "n": 15 }, "agent_ttft_seconds": { "median": 9.763, "min": 5.728, "max": 11.312, "n": 15 }, "clients": { "clients": 2, "rounds": [ { "round": 0, "wall_s": 17.22, "tokens": 256, "total_tps": 14.87, "first_s": [ 2.326245716001722, 11.58621229200071 ], "latency_s": [ 9.302565578997019, 17.218570477001776 ], "errors": [], "error_count": 0 }, { "round": 1, "wall_s": 16.33, "tokens": 256, "total_tps": 15.68, "first_s": [ 2.4264337810018333, 10.784170826002082 ], "latency_s": [ 8.503292759000033, 16.328239408001537 ], "errors": [], "error_count": 0 }, { "round": 2, "wall_s": 16.44, "tokens": 256, "total_tps": 15.57, "first_s": [ 2.4174519359949045, 10.968774136999855 ], "latency_s": [ 8.688245152996387, 16.436975523 ], "errors": [], "error_count": 0 } ], "total_tps": { "median": 15.57, "min": 14.87, "max": 15.68, "n": 3 }, "ttft_p50_s": 6.605, "ttft_p95_s": 11.432, "latency_p50_s": 12.815, "latency_p95_s": 17.023, "errors": 0 }, "drafts_offered": 1874, "drafts_accepted": 1181, "source": "source/selected/results/coder-iq1_m-ctx65536/out/official-optional/2026-10-17-community-unknown-qwen3-8-flash-next-coder-iq1-m/results.json" } ], "quant": "IQ1_M", "scores": { "business": 100.0, "research": 100.0, "knowledge": 91.6667, "agent_tools": 90.625, "practical_sysadmin": 89.0975, "executive_content": 94.3108, "humaneval_plus": 92.68 }, "humaneval_plus_passes": 152, "humaneval_base_passes": 158, "humaneval_evaluated": 164, "practical_attempts": 72, "practical_empty_answers": 1, "practical_output_limit_finishes": 3, "humaneval_empty_solutions": 0, "humaneval_output_limit_finishes": 0, "recorded_transport_failures": 1, "dossier": { "repetitions": 3, "prompt_tokens": [ 57135, 57135, 57135 ], "cached_prompt_tokens": [ 0, 57130, 57130 ], "wall_seconds": [ 250.2572, 90.4603, 89.0411 ], "median_wall_seconds": 90.4603, "limitation": "Repeated fixture/cache effects; output heuristics are diagnostic, not an independent factual accuracy score." }, "weights": [ { "path": "data/models/coder-IQ1_M/Qwen3.8-Flash-Next-GSQ-RCO-IQ1_M-00001-of-00002.gguf", "bytes": 29608446496, "mtime_ns": 1792222790578054808, "sha256": "e11083ba855e7666b48ea3f2db6a9c3a20c18751a012cc24f948de91b7087fad" }, { "path": "data/models/coder-IQ1_M/Qwen3.8-Flash-Next-GSQ-RCO-IQ1_M-00002-of-00002.gguf", "bytes": 28800138432, "mtime_ns": 1792223075432550870, "sha256": "316b46f3a2dbd68c900f43136ab9449f9dcc3725dfd8c794847c204bc161e113" } ], "requested_mtp": 4, "download_total_bytes": 58408584928, "repository": "ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF", "repository_revision": "5348543e0147355ac9cbcb031184a3546350988e", "highest_recall_validated_prompt_tokens": 125573, "highest_allocation_with_successful_recall": 131072, "sampled_ram_gib": { "min": 22.57, "max": 25.76, "n": 25 }, "observed_mtp_drafts_offered": 3596, "observed_mtp_drafts_accepted": 2224, "provider": "ISTA-DASLab" }, { "id": "qwen36-ud-q4_k_xl", "label": "Qwen3.6 UD-Q4_K_XL", "quality_scope": "Full frozen v2.1 suite", "balanced_score": 96.44, "humaneval_plus_percent": 79.27, "role_scores": { "CFO / Analyst": 99.02, "Researcher / Scout": 98.04, "Hermes Coordinator": 97.11, "Librarian / Knowledge": 98.15, "Forge / Coder": 88.85 }, "contexts": [ { "allocated_tokens": 131072, "allocation_passed": false, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "not_completed", "startup_failure": "strata-q36: cudaMalloc(&p, n * 4) failed: out of memory (/upstream/src/qwen35/q36_engine.cu:1127)" }, { "allocated_tokens": 16384, "allocation_passed": true, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "allocation_passed" }, { "allocated_tokens": 32768, "allocation_passed": true, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "allocation_passed" }, { "allocated_tokens": 65536, "allocation_passed": true, "recall": [ { "length": "16k", "depth": 10, "prompt_tokens": 15156, "seconds": 22.9, "found": true, "word": "meadow-copper-504", "answer": "meadow-copper-504" }, { "length": "16k", "depth": 50, "prompt_tokens": 15156, "seconds": 22.9, "found": true, "word": "falcon-quartz-940", "answer": "falcon-quartz-940" }, { "length": "16k", "depth": 90, "prompt_tokens": 15157, "seconds": 22.9, "found": true, "word": "willow-cobalt-696", "answer": "willow-cobalt-696" }, { "length": "32k", "depth": 10, "prompt_tokens": 32792, "seconds": 62.5, "found": true, "word": "meadow-copper-504", "answer": "meadow-copper-504" }, { "length": "32k", "depth": 50, "prompt_tokens": 32792, "seconds": 62.6, "found": true, "word": "falcon-quartz-940", "answer": "falcon-quartz-940" }, { "length": "32k", "depth": 90, "prompt_tokens": 32793, "seconds": 62.6, "found": true, "word": "willow-cobalt-696", "answer": "willow-cobalt-696" }, { "length": "8k", "depth": 10, "prompt_tokens": 6920, "seconds": 9.5, "found": true, "word": "meadow-copper-504", "answer": "meadow-copper-504" }, { "length": "8k", "depth": 50, "prompt_tokens": 6920, "seconds": 9.6, "found": true, "word": "falcon-quartz-940", "answer": "falcon-quartz-940" }, { "length": "8k", "depth": 90, "prompt_tokens": 6921, "seconds": 9.6, "found": true, "word": "willow-cobalt-696", "answer": "willow-cobalt-696" } ], "fully_passed_recall_tiers": [ "16k", "32k", "8k" ], "max_recall_validated_prompt_tokens": 32793, "status": "allocation_passed" }, { "allocated_tokens": 8192, "allocation_passed": true, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "allocation_passed" }, { "allocated_tokens": 81920, "allocation_passed": true, "recall": [ { "length": "74k", "depth": 10, "prompt_tokens": 75929, "seconds": 227.2, "found": true, "word": "meadow-copper-504", "answer": "meadow-copper-504" }, { "length": "74k", "depth": 50, "prompt_tokens": 75929, "seconds": 227.2, "found": true, "word": "falcon-quartz-940", "answer": "falcon-quartz-940" }, { "length": "74k", "depth": 90, "prompt_tokens": 75930, "seconds": 227.3, "found": true, "word": "willow-cobalt-696", "answer": "willow-cobalt-696" } ], "fully_passed_recall_tiers": [ "74k" ], "max_recall_validated_prompt_tokens": 75930, "status": "allocation_passed" }, { "allocated_tokens": 98304, "allocation_passed": false, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "not_completed", "startup_failure": "strata-q36: not enough VRAM left for the hybrid experts' buffers and slots (1256 MiB needed, 756 MiB free after the weights, the context and the sessions): use fewer slots (\"parallel\" / --batch), a shorter --max-context" } ], "official": [ { "allocated_tokens": 32768, "cohort": "official", "runs": 3, "elapsed_seconds": 344.77955770492554, "chat_decode_tps": { "median": 33.2, "min": 26.4, "max": 37.4, "n": 9 }, "chat_ttft_seconds": { "median": 0.667, "min": 0.478, "max": 0.849, "n": 9 }, "failures": [], "prefill": { "4K": { "tokens_per_second": { "median": 814.3, "min": 800.9, "max": 814.4, "n": 3 }, "ttft_seconds": { "median": 4.941, "min": 4.941, "max": 5.025, "n": 3 } }, "16K": { "tokens_per_second": { "median": 677.5, "min": 677.5, "max": 677.6, "n": 3 }, "ttft_seconds": { "median": 23.717, "min": 23.716, "max": 23.718, "n": 3 } } }, "agent_cache_reuse_percent": { "median": 0.0, "min": 0.0, "max": 0.0, "n": 3 }, "agent_decode_tps": { "median": 32.6, "min": 27.7, "max": 38.0, "n": 15 }, "agent_ttft_seconds": { "median": 11.803, "min": 4.679, "max": 21.729, "n": 15 }, "clients": {}, "drafts_offered": 0, "drafts_accepted": 0, "source": "source/selected/results/qwen36-ud-q4_k_xl-ctx32768/out/official/2026-10-17-community-unknown-qwen3-6-35b-a3b-ud-q4-k-xl/results.json" }, { "allocated_tokens": 65536, "cohort": "official-optional", "runs": 3, "elapsed_seconds": 573.9356691837311, "chat_decode_tps": { "median": 24.6, "min": 24.2, "max": 24.9, "n": 9 }, "chat_ttft_seconds": { "median": 0.865, "min": 0.839, "max": 0.978, "n": 9 }, "failures": [], "prefill": { "4K": { "tokens_per_second": { "median": 813.4, "min": 811.3, "max": 813.9, "n": 3 }, "ttft_seconds": { "median": 4.947, "min": 4.944, "max": 4.96, "n": 3 } }, "16K": { "tokens_per_second": { "median": 677.8, "min": 677.2, "max": 678.0, "n": 3 }, "ttft_seconds": { "median": 23.707, "min": 23.701, "max": 23.726, "n": 3 } }, "32K": { "tokens_per_second": { "median": 543.5, "min": 543.2, "max": 543.6, "n": 3 }, "ttft_seconds": { "median": 59.002, "min": 58.988, "max": 59.03, "n": 3 } } }, "agent_cache_reuse_percent": { "median": 0.0, "min": 0.0, "max": 0.0, "n": 3 }, "agent_decode_tps": { "median": 27.7, "min": 24.6, "max": 28.8, "n": 15 }, "agent_ttft_seconds": { "median": 11.785, "min": 4.646, "max": 21.735, "n": 15 }, "clients": { "clients": 2, "rounds": [ { "round": 0, "wall_s": 12.04, "tokens": 256, "total_tps": 21.27, "first_s": [ 1.0837469640027848, 7.222556913999142 ], "latency_s": [ 6.19907937500102, 12.035335437998583 ], "errors": [], "error_count": 0 }, { "round": 1, "wall_s": 11.28, "tokens": 256, "total_tps": 22.7, "first_s": [ 1.1046717380013433, 6.547357652998471 ], "latency_s": [ 5.464262839006551, 11.276742007001303 ], "errors": [], "error_count": 0 }, { "round": 2, "wall_s": 10.68, "tokens": 256, "total_tps": 23.96, "first_s": [ 1.0722366210029577, 6.228817782000988 ], "latency_s": [ 5.141915908003284, 10.682149476000632 ], "errors": [], "error_count": 0 } ], "total_tps": { "median": 22.7, "min": 21.27, "max": 23.96, "n": 3 }, "ttft_p50_s": 3.667, "ttft_p95_s": 7.054, "latency_p50_s": 8.441, "latency_p95_s": 11.846, "errors": 0 }, "drafts_offered": 0, "drafts_accepted": 0, "source": "source/selected/results/qwen36-ud-q4_k_xl-ctx65536/out/official-optional/2026-10-17-community-unknown-qwen3-6-35b-a3b-ud-q4-k-xl/results.json" } ], "quant": "UD-Q4_K_XL", "scores": { "business": 100.0, "research": 96.8942, "knowledge": 100.0, "agent_tools": 100.0, "practical_sysadmin": 96.875, "executive_content": 92.855, "humaneval_plus": 79.27 }, "humaneval_plus_passes": 130, "humaneval_base_passes": 136, "humaneval_evaluated": 164, "practical_attempts": 72, "practical_empty_answers": 0, "practical_output_limit_finishes": 0, "humaneval_empty_solutions": 0, "humaneval_output_limit_finishes": 0, "recorded_transport_failures": 0, "dossier": { "repetitions": 3, "prompt_tokens": [ 57097, 57097, 57097 ], "cached_prompt_tokens": [ 0, 0, 0 ], "wall_seconds": [ 313.9941, 319.78, 337.1514 ], "median_wall_seconds": 319.78, "limitation": "Repeated fixture/cache effects; output heuristics are diagnostic, not an independent factual accuracy score." }, "weights": [ { "path": "data/models/qwen36-UD-Q4_K_XL/Qwen3.6-35B-A3B-UD-Q4_K_XL.gguf", "bytes": 22853663008, "mtime_ns": 1792257125836886652, "sha256": "55983c5a75a1ab969824077b3bb3de4146e82a9234072b48ad4e8f92ad3fe9f1" } ], "requested_mtp": 2, "download_total_bytes": 22853663008, "repository": "unsloth/Qwen3.6-35B-A3B-MTP-GGUF", "repository_revision": "5bc3e238d916f48a861bac2f8a1990a0e9b7e98d", "highest_recall_validated_prompt_tokens": 75930, "highest_allocation_with_successful_recall": 81920, "sampled_ram_gib": { "min": 20.51, "max": 24.05, "n": 33 }, "observed_mtp_drafts_offered": 0, "observed_mtp_drafts_accepted": 0, "provider": "unsloth" }, { "id": "qwen36-ud-q3_k_m-zero-mtp", "label": "Qwen3.6 UD-Q3_K_M (no MTP)", "quality_scope": "not evaluated", "balanced_score": null, "humaneval_plus_percent": null, "role_scores": {}, "contexts": [ { "allocated_tokens": 131072, "allocation_passed": false, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "failed", "retrieval_passed": null, "startup_failure": "strata-q36: not enough VRAM left for the hybrid experts' buffers and slots (943 MiB needed, 562 MiB free after the weights, the context and the sessions): use fewer slots (\"parallel\" / --batch), a shorter --max-context" }, { "allocated_tokens": 65536, "allocation_passed": true, "recall": [ { "length": "60k", "depth": 10, "prompt_tokens": 61342, "seconds": 151.0, "found": true, "word": "meadow-copper-504", "answer": "meadow-copper-504" }, { "length": "60k", "depth": 50, "prompt_tokens": 61342, "seconds": 148.2, "found": true, "word": "falcon-quartz-940", "answer": "falcon-quartz-940" }, { "length": "60k", "depth": 90, "prompt_tokens": 61343, "seconds": 148.4, "found": true, "word": "willow-cobalt-696", "answer": "willow-cobalt-696" } ], "fully_passed_recall_tiers": [ "60k" ], "max_recall_validated_prompt_tokens": 61343, "status": "passed", "retrieval_passed": true }, { "allocated_tokens": 81920, "allocation_passed": true, "recall": [ { "length": "74k", "depth": 10, "prompt_tokens": 75929, "seconds": 215.5, "found": true, "word": "meadow-copper-504", "answer": "meadow-copper-504" }, { "length": "74k", "depth": 50, "prompt_tokens": 75929, "seconds": 211.3, "found": true, "word": "falcon-quartz-940", "answer": "falcon-quartz-940" }, { "length": "74k", "depth": 90, "prompt_tokens": 75930, "seconds": 211.4, "found": true, "word": "willow-cobalt-696", "answer": "willow-cobalt-696" } ], "fully_passed_recall_tiers": [ "74k" ], "max_recall_validated_prompt_tokens": 75930, "status": "passed", "retrieval_passed": true } ], "official": [ { "allocated_tokens": 65536, "cohort": "official-quant", "runs": 3, "elapsed_seconds": 502.8638844490051, "chat_decode_tps": { "median": 27.4, "min": 21.4, "max": 36.0, "n": 9 }, "chat_ttft_seconds": { "median": 0.673, "min": 0.578, "max": 0.819, "n": 9 }, "failures": [], "prefill": { "4K": { "tokens_per_second": { "median": 862.6, "min": 850.5, "max": 862.9, "n": 3 }, "ttft_seconds": { "median": 4.665, "min": 4.663, "max": 4.732, "n": 3 } }, "16K": { "tokens_per_second": { "median": 721.4, "min": 721.1, "max": 721.7, "n": 3 }, "ttft_seconds": { "median": 22.274, "min": 22.267, "max": 22.284, "n": 3 } }, "32K": { "tokens_per_second": { "median": 581.5, "min": 581.0, "max": 581.5, "n": 3 }, "ttft_seconds": { "median": 55.147, "min": 55.144, "max": 55.186, "n": 3 } } }, "agent_cache_reuse_percent": { "median": 0.0, "min": 0.0, "max": 0.0, "n": 3 }, "agent_decode_tps": { "median": 29.6, "min": 26.3, "max": 34.2, "n": 15 }, "agent_ttft_seconds": { "median": 11.018, "min": 4.381, "max": 20.216, "n": 15 }, "clients": {}, "drafts_offered": 0, "drafts_accepted": 0, "source": "source/selected/results/qwen36-ud-q3_k_m-zero-mtp-ctx65536/out/official-quant/2026-10-17-community-unknown-qwen36-ud-q3-k-m-zero-mtp/results.json" }, { "allocated_tokens": 81920, "cohort": "official-quant", "runs": 3, "elapsed_seconds": 514.5989193916321, "chat_decode_tps": { "median": 23.5, "min": 20.6, "max": 28.2, "n": 9 }, "chat_ttft_seconds": { "median": 0.834, "min": 0.761, "max": 0.909, "n": 9 }, "failures": [], "prefill": { "4K": { "tokens_per_second": { "median": 862.5, "min": 854.9, "max": 862.7, "n": 3 }, "ttft_seconds": { "median": 4.665, "min": 4.664, "max": 4.707, "n": 3 } }, "16K": { "tokens_per_second": { "median": 721.3, "min": 721.3, "max": 721.7, "n": 3 }, "ttft_seconds": { "median": 22.276, "min": 22.264, "max": 22.278, "n": 3 } }, "32K": { "tokens_per_second": { "median": 581.7, "min": 581.1, "max": 581.9, "n": 3 }, "ttft_seconds": { "median": 55.122, "min": 55.108, "max": 55.18, "n": 3 } } }, "agent_cache_reuse_percent": { "median": 0.0, "min": 0.0, "max": 0.0, "n": 3 }, "agent_decode_tps": { "median": 26.7, "min": 24.3, "max": 29.4, "n": 15 }, "agent_ttft_seconds": { "median": 10.997, "min": 4.37, "max": 21.023, "n": 15 }, "clients": {}, "drafts_offered": 0, "drafts_accepted": 0, "source": "source/selected/results/qwen36-ud-q3_k_m-zero-mtp-ctx81920/out/official-quant/2026-10-17-community-unknown-qwen36-ud-q3-k-m-zero-mtp/results.json" } ], "quant": "UD-Q3_K_M", "repository": "unsloth/Qwen3.6-35B-A3B-GGUF", "repository_revision": "a483e9e6cbd595906af30beda3187c2663a1118c", "file_name": "Qwen3.6-35B-A3B-UD-Q3_K_M.gguf", "file_bytes": 16600710112, "file_sha256": "1b715841683f960bd9a49f008181bd910ee169b78d4cf465b6fde7f4d929ff99", "requested_mtp": 0, "metadata_compatibility": { "change": { "qwen35moe.nextn_predict_layers": 0 }, "source": "/data/models/qwen36-ud-q3_k_m-zero-mtp/Qwen3.6-35B-A3B-UD-Q3_K_M.gguf", "compatibility_file": "/data/models/qwen36-ud-q3_k_m-zero-mtp/Qwen3.6-35B-A3B-UD-Q3_K_M-strata-zero-mtp.gguf", "source_sha256": "1b715841683f960bd9a49f008181bd910ee169b78d4cf465b6fde7f4d929ff99", "compatibility_sha256": "9623429e1b7f55dc741843c3334c1773db9d351a47dd04f55b2ec57932dde9ab", "tensor_data_sha256": "d2651072d67f55aa8f6f7e296263745d39db59426379eb80581e6dd41242da4d", "tensor_data_byte_identical": true, "original_data_offset": 10990048, "compatibility_data_offset": 10990112, "all_other_metadata_equal": true, "tensor_descriptors_equal": true }, "highest_recall_validated_prompt_tokens": 75930, "highest_allocation_with_successful_recall": 81920, "sampled_ram_gib": { "min": 16.06, "max": 16.21, "n": 14 }, "observed_mtp_drafts_offered": 0, "observed_mtp_drafts_accepted": 0, "provider": "unsloth" }, { "id": "qwen36-ud-q4_k_m-mtp", "label": "Qwen3.6 UD-Q4_K_M (MTP file)", "quality_scope": "not evaluated", "balanced_score": null, "humaneval_plus_percent": null, "role_scores": {}, "contexts": [ { "allocated_tokens": 131072, "allocation_passed": false, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "failed", "retrieval_passed": null, "startup_failure": "strata-q36: cudaMalloc(&p, n * 4) failed: out of memory (/upstream/src/qwen35/q36_engine.cu:1127)" }, { "allocated_tokens": 65536, "allocation_passed": true, "recall": [ { "length": "60k", "depth": 10, "prompt_tokens": 61342, "seconds": 163.8, "found": true, "word": "meadow-copper-504", "answer": "meadow-copper-504" }, { "length": "60k", "depth": 50, "prompt_tokens": 61342, "seconds": 160.1, "found": true, "word": "falcon-quartz-940", "answer": "falcon-quartz-940" }, { "length": "60k", "depth": 90, "prompt_tokens": 61343, "seconds": 160.3, "found": true, "word": "willow-cobalt-696", "answer": "willow-cobalt-696" } ], "fully_passed_recall_tiers": [ "60k" ], "max_recall_validated_prompt_tokens": 61343, "status": "passed", "retrieval_passed": true }, { "allocated_tokens": 81920, "allocation_passed": true, "recall": [ { "length": "74k", "depth": 10, "prompt_tokens": 75929, "seconds": 227.2, "found": true, "word": "meadow-copper-504", "answer": "meadow-copper-504" }, { "length": "74k", "depth": 50, "prompt_tokens": 75929, "seconds": 227.1, "found": true, "word": "falcon-quartz-940", "answer": "falcon-quartz-940" }, { "length": "74k", "depth": 90, "prompt_tokens": 75930, "seconds": 227.2, "found": true, "word": "willow-cobalt-696", "answer": "willow-cobalt-696" } ], "fully_passed_recall_tiers": [ "74k" ], "max_recall_validated_prompt_tokens": 75930, "status": "passed", "retrieval_passed": true } ], "official": [ { "allocated_tokens": 65536, "cohort": "official-quant", "runs": 3, "elapsed_seconds": 530.9470307826996, "chat_decode_tps": { "median": 27.5, "min": 25.0, "max": 30.0, "n": 9 }, "chat_ttft_seconds": { "median": 0.815, "min": 0.647, "max": 0.883, "n": 9 }, "failures": [], "prefill": { "4K": { "tokens_per_second": { "median": 814.8, "min": 807.4, "max": 815.5, "n": 3 }, "ttft_seconds": { "median": 4.938, "min": 4.934, "max": 4.984, "n": 3 } }, "16K": { "tokens_per_second": { "median": 678.0, "min": 677.6, "max": 678.5, "n": 3 }, "ttft_seconds": { "median": 23.7, "min": 23.683, "max": 23.715, "n": 3 } }, "32K": { "tokens_per_second": { "median": 543.3, "min": 543.0, "max": 543.3, "n": 3 }, "ttft_seconds": { "median": 59.021, "min": 59.02, "max": 59.046, "n": 3 } } }, "agent_cache_reuse_percent": { "median": 0.0, "min": 0.0, "max": 0.0, "n": 3 }, "agent_decode_tps": { "median": 29.9, "min": 26.4, "max": 34.1, "n": 15 }, "agent_ttft_seconds": { "median": 11.723, "min": 4.647, "max": 21.739, "n": 15 }, "clients": {}, "drafts_offered": 0, "drafts_accepted": 0, "source": "source/selected/results/qwen36-ud-q4_k_m-mtp-ctx65536/out/official-quant/2026-10-17-community-unknown-qwen36-ud-q4-k-m-mtp/results.json" }, { "allocated_tokens": 81920, "cohort": "official-quant", "runs": 3, "elapsed_seconds": 538.8739361763, "chat_decode_tps": { "median": 24.8, "min": 24.5, "max": 25.2, "n": 9 }, "chat_ttft_seconds": { "median": 0.879, "min": 0.854, "max": 0.986, "n": 9 }, "failures": [], "prefill": { "4K": { "tokens_per_second": { "median": 813.3, "min": 812.5, "max": 815.2, "n": 3 }, "ttft_seconds": { "median": 4.949, "min": 4.936, "max": 4.952, "n": 3 } }, "16K": { "tokens_per_second": { "median": 678.2, "min": 677.7, "max": 678.3, "n": 3 }, "ttft_seconds": { "median": 23.69, "min": 23.689, "max": 23.708, "n": 3 } }, "32K": { "tokens_per_second": { "median": 543.5, "min": 543.2, "max": 543.5, "n": 3 }, "ttft_seconds": { "median": 59.004, "min": 58.996, "max": 59.028, "n": 3 } } }, "agent_cache_reuse_percent": { "median": 0.0, "min": 0.0, "max": 0.0, "n": 3 }, "agent_decode_tps": { "median": 26.8, "min": 24.9, "max": 28.7, "n": 15 }, "agent_ttft_seconds": { "median": 11.723, "min": 4.64, "max": 21.711, "n": 15 }, "clients": {}, "drafts_offered": 0, "drafts_accepted": 0, "source": "source/selected/results/qwen36-ud-q4_k_m-mtp-ctx81920/out/official-quant/2026-10-18-community-unknown-qwen36-ud-q4-k-m-mtp/results.json" } ], "quant": "UD-Q4_K_M", "repository": "unsloth/Qwen3.6-35B-A3B-MTP-GGUF", "repository_revision": "5bc3e238d916f48a861bac2f8a1990a0e9b7e98d", "file_name": "Qwen3.6-35B-A3B-UD-Q4_K_M.gguf", "file_bytes": 22663387424, "file_sha256": "0b21525e972670ed59e1812e170b27c26355381f0656ecc4e25617ece7dac58b", "requested_mtp": 2, "metadata_compatibility": null, "highest_recall_validated_prompt_tokens": 75930, "highest_allocation_with_successful_recall": 81920, "sampled_ram_gib": { "min": 19.36, "max": 25.57, "n": 14 }, "observed_mtp_drafts_offered": 0, "observed_mtp_drafts_accepted": 0, "provider": "unsloth" }, { "id": "qwen36-ud-q3_k_xl-zero-mtp", "label": "Qwen3.6 UD-Q3_K_XL (no MTP)", "quality_scope": "not evaluated", "balanced_score": null, "humaneval_plus_percent": null, "role_scores": {}, "contexts": [ { "allocated_tokens": 131072, "allocation_passed": false, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "failed", "retrieval_passed": null, "startup_failure": "strata-q36: not enough VRAM left for the hybrid experts' buffers and slots (1025 MiB needed, 444 MiB free after the weights, the context and the sessions): use fewer slots (\"parallel\" / --batch), a shorter --max-context" }, { "allocated_tokens": 65536, "allocation_passed": true, "recall": [ { "length": "60k", "depth": 10, "prompt_tokens": 61342, "seconds": 148.3, "found": true, "word": "meadow-copper-504", "answer": "meadow-copper-504" }, { "length": "60k", "depth": 50, "prompt_tokens": 61342, "seconds": 148.2, "found": true, "word": "falcon-quartz-940", "answer": "falcon-quartz-940" }, { "length": "60k", "depth": 90, "prompt_tokens": 61343, "seconds": 148.4, "found": true, "word": "willow-cobalt-696", "answer": "willow-cobalt-696" } ], "fully_passed_recall_tiers": [ "60k" ], "max_recall_validated_prompt_tokens": 61343, "status": "passed", "retrieval_passed": true }, { "allocated_tokens": 81920, "allocation_passed": true, "recall": [ { "length": "74k", "depth": 10, "prompt_tokens": 75929, "seconds": 209.8, "found": true, "word": "meadow-copper-504", "answer": "meadow-copper-504" }, { "length": "74k", "depth": 50, "prompt_tokens": 75929, "seconds": 209.8, "found": true, "word": "falcon-quartz-940", "answer": "falcon-quartz-940" }, { "length": "74k", "depth": 90, "prompt_tokens": 75930, "seconds": 209.9, "found": true, "word": "willow-cobalt-696", "answer": "willow-cobalt-696" } ], "fully_passed_recall_tiers": [ "74k" ], "max_recall_validated_prompt_tokens": 75930, "status": "passed", "retrieval_passed": true } ], "official": [ { "allocated_tokens": 65536, "cohort": "official-quant", "runs": 3, "elapsed_seconds": 505.60574316978455, "chat_decode_tps": { "median": 26.4, "min": 21.6, "max": 34.4, "n": 9 }, "chat_ttft_seconds": { "median": 0.706, "min": 0.622, "max": 0.837, "n": 9 }, "failures": [], "prefill": { "4K": { "tokens_per_second": { "median": 861.5, "min": 850.7, "max": 861.6, "n": 3 }, "ttft_seconds": { "median": 4.671, "min": 4.67, "max": 4.731, "n": 3 } }, "16K": { "tokens_per_second": { "median": 721.0, "min": 720.9, "max": 721.5, "n": 3 }, "ttft_seconds": { "median": 22.286, "min": 22.271, "max": 22.291, "n": 3 } }, "32K": { "tokens_per_second": { "median": 581.4, "min": 580.8, "max": 581.5, "n": 3 }, "ttft_seconds": { "median": 55.148, "min": 55.14, "max": 55.213, "n": 3 } } }, "agent_cache_reuse_percent": { "median": 0.0, "min": 0.0, "max": 0.0, "n": 3 }, "agent_decode_tps": { "median": 28.7, "min": 24.7, "max": 33.4, "n": 15 }, "agent_ttft_seconds": { "median": 10.999, "min": 4.4, "max": 20.226, "n": 15 }, "clients": {}, "drafts_offered": 0, "drafts_accepted": 0, "source": "source/selected/results/qwen36-ud-q3_k_xl-zero-mtp-ctx65536/out/official-quant/2026-10-18-community-unknown-qwen36-ud-q3-k-xl-zero-mtp/results.json" }, { "allocated_tokens": 81920, "cohort": "official-quant", "runs": 3, "elapsed_seconds": 515.8064384460449, "chat_decode_tps": { "median": 22.6, "min": 20.3, "max": 25.0, "n": 9 }, "chat_ttft_seconds": { "median": 0.831, "min": 0.797, "max": 0.948, "n": 9 }, "failures": [], "prefill": { "4K": { "tokens_per_second": { "median": 861.8, "min": 855.4, "max": 861.9, "n": 3 }, "ttft_seconds": { "median": 4.67, "min": 4.669, "max": 4.704, "n": 3 } }, "16K": { "tokens_per_second": { "median": 720.9, "min": 720.8, "max": 721.3, "n": 3 }, "ttft_seconds": { "median": 22.289, "min": 22.28, "max": 22.295, "n": 3 } }, "32K": { "tokens_per_second": { "median": 581.5, "min": 581.0, "max": 581.6, "n": 3 }, "ttft_seconds": { "median": 55.141, "min": 55.138, "max": 55.187, "n": 3 } } }, "agent_cache_reuse_percent": { "median": 0.0, "min": 0.0, "max": 0.0, "n": 3 }, "agent_decode_tps": { "median": 25.8, "min": 23.5, "max": 28.4, "n": 15 }, "agent_ttft_seconds": { "median": 11.071, "min": 4.389, "max": 20.233, "n": 15 }, "clients": {}, "drafts_offered": 0, "drafts_accepted": 0, "source": "source/selected/results/qwen36-ud-q3_k_xl-zero-mtp-ctx81920/out/official-quant/2026-10-18-community-unknown-qwen36-ud-q3-k-xl-zero-mtp/results.json" } ], "quant": "UD-Q3_K_XL", "repository": "unsloth/Qwen3.6-35B-A3B-GGUF", "repository_revision": "a483e9e6cbd595906af30beda3187c2663a1118c", "file_name": "Qwen3.6-35B-A3B-UD-Q3_K_XL.gguf", "file_bytes": 16845511648, "file_sha256": "a832b9689925f1bd335bbe985cdfb06c36bf2cf268f4f8f6eceafa3ceb515617", "requested_mtp": 0, "metadata_compatibility": { "change": { "qwen35moe.nextn_predict_layers": 0 }, "source": "/data/models/qwen36-ud-q3_k_xl-zero-mtp/Qwen3.6-35B-A3B-UD-Q3_K_XL.gguf", "compatibility_file": "/data/models/qwen36-ud-q3_k_xl-zero-mtp/Qwen3.6-35B-A3B-UD-Q3_K_XL-strata-zero-mtp.gguf", "source_sha256": "a832b9689925f1bd335bbe985cdfb06c36bf2cf268f4f8f6eceafa3ceb515617", "compatibility_sha256": "79f770b6d56d57d5373520a558874fc4fb3bb442a1bc07c4fc3e7823f9f95455", "tensor_data_sha256": "f79afe0e9ab26272beb34202f018d94c65969f6a3f7fb1ab0879eb02ef254fe9", "tensor_data_byte_identical": true, "original_data_offset": 10990048, "compatibility_data_offset": 10990112, "all_other_metadata_equal": true, "tensor_descriptors_equal": true }, "highest_recall_validated_prompt_tokens": 75930, "highest_allocation_with_successful_recall": 81920, "sampled_ram_gib": { "min": 14.75, "max": 17.16, "n": 13 }, "observed_mtp_drafts_offered": 0, "observed_mtp_drafts_accepted": 0, "provider": "unsloth" }, { "id": "qwen36-ud-iq3_s", "label": "Qwen3.6 UD-IQ3_S (MTP file)", "quality_scope": "not evaluated", "balanced_score": null, "humaneval_plus_percent": null, "role_scores": {}, "contexts": [ { "allocated_tokens": 131072, "allocation_passed": false, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "failed", "retrieval_passed": null, "startup_failure": "strata-q36: blk.40.experts are Q2_K/Q4_K, which the native expert kernels do not take (build with -DSTRATA_Q6K_EXPERTS=ON for Q6_K)" }, { "allocated_tokens": 65536, "allocation_passed": false, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "failed", "retrieval_passed": null, "startup_failure": "strata-q36: blk.40.experts are Q2_K/Q4_K, which the native expert kernels do not take (build with -DSTRATA_Q6K_EXPERTS=ON for Q6_K)" }, { "allocated_tokens": 81920, "allocation_passed": false, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "failed", "retrieval_passed": null, "startup_failure": "strata-q36: blk.40.experts are Q2_K/Q4_K, which the native expert kernels do not take (build with -DSTRATA_Q6K_EXPERTS=ON for Q6_K)" } ], "official": [], "quant": "UD-IQ3_S", "repository": "unsloth/Qwen3.6-35B-A3B-MTP-GGUF", "repository_revision": "5bc3e238d916f48a861bac2f8a1990a0e9b7e98d", "file_name": "Qwen3.6-35B-A3B-UD-IQ3_S.gguf", "file_bytes": 15346432288, "file_sha256": "ab639a7f330f96c47d3e6c2dd2d6445182e7b763e17e6048dc850a71bbc9f27f", "requested_mtp": 2, "metadata_compatibility": null, "highest_recall_validated_prompt_tokens": 0, "highest_allocation_with_successful_recall": 0, "observed_mtp_drafts_offered": 0, "observed_mtp_drafts_accepted": 0, "provider": "unsloth" }, { "id": "qwen36-ud-iq3_xxs", "label": "Qwen3.6 UD-IQ3_XXS (MTP file)", "quality_scope": "not evaluated", "balanced_score": null, "humaneval_plus_percent": null, "role_scores": {}, "contexts": [ { "allocated_tokens": 131072, "allocation_passed": false, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "failed", "retrieval_passed": null, "startup_failure": "strata-q36: blk.0.experts are IQ2_S/IQ3_S, which the native expert kernels do not take (build with -DSTRATA_Q6K_EXPERTS=ON for Q6_K)" }, { "allocated_tokens": 65536, "allocation_passed": false, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "failed", "retrieval_passed": null, "startup_failure": "strata-q36: blk.0.experts are IQ2_S/IQ3_S, which the native expert kernels do not take (build with -DSTRATA_Q6K_EXPERTS=ON for Q6_K)" }, { "allocated_tokens": 81920, "allocation_passed": false, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "failed", "retrieval_passed": null, "startup_failure": "strata-q36: blk.0.experts are IQ2_S/IQ3_S, which the native expert kernels do not take (build with -DSTRATA_Q6K_EXPERTS=ON for Q6_K)" } ], "official": [], "quant": "UD-IQ3_XXS", "repository": "unsloth/Qwen3.6-35B-A3B-MTP-GGUF", "repository_revision": "5bc3e238d916f48a861bac2f8a1990a0e9b7e98d", "file_name": "Qwen3.6-35B-A3B-UD-IQ3_XXS.gguf", "file_bytes": 14069266720, "file_sha256": "36f9ec0e4c775f6efd3a61c1ea76f0875128469c3d012e3d9e2495a90e7a7150", "requested_mtp": 2, "metadata_compatibility": null, "highest_recall_validated_prompt_tokens": 0, "highest_allocation_with_successful_recall": 0, "observed_mtp_drafts_offered": 0, "observed_mtp_drafts_accepted": 0, "provider": "unsloth" }, { "id": "qwen36-ud-iq4_xs", "label": "Qwen3.6 UD-IQ4_XS (MTP file)", "quality_scope": "not evaluated", "balanced_score": null, "humaneval_plus_percent": null, "role_scores": {}, "contexts": [ { "allocated_tokens": 131072, "allocation_passed": false, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "failed", "retrieval_passed": null, "startup_failure": "strata-q36: blk.40.experts are Q3_K/Q4_K, which the native expert kernels do not take (build with -DSTRATA_Q6K_EXPERTS=ON for Q6_K)" }, { "allocated_tokens": 65536, "allocation_passed": false, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "failed", "retrieval_passed": null, "startup_failure": "strata-q36: blk.40.experts are Q3_K/Q4_K, which the native expert kernels do not take (build with -DSTRATA_Q6K_EXPERTS=ON for Q6_K)" } ], "official": [], "quant": "UD-IQ4_XS", "repository": "unsloth/Qwen3.6-35B-A3B-MTP-GGUF", "repository_revision": "5bc3e238d916f48a861bac2f8a1990a0e9b7e98d", "file_name": "Qwen3.6-35B-A3B-UD-IQ4_XS.gguf", "file_bytes": 18209036576, "file_sha256": "df27a780435b7b45c2597536112ea3cb091f8544c3d0c3318d9f4258b31f7adf", "requested_mtp": 2, "metadata_compatibility": null, "highest_recall_validated_prompt_tokens": 0, "highest_allocation_with_successful_recall": 0, "observed_mtp_drafts_offered": 0, "observed_mtp_drafts_accepted": 0, "provider": "unsloth" }, { "id": "qwen36-genesis-hermes-apex-compact", "label": "Genesis-Hermes APEX Compact", "quality_scope": "not evaluated", "balanced_score": null, "humaneval_plus_percent": null, "role_scores": {}, "contexts": [ { "allocated_tokens": 131072, "allocation_passed": false, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "failed", "retrieval_passed": null, "startup_failure": "strata-q36: missing qwen35moe.nextn_predict_layers" }, { "allocated_tokens": 65536, "allocation_passed": false, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "failed", "retrieval_passed": null, "startup_failure": "strata-q36: missing qwen35moe.nextn_predict_layers" }, { "allocated_tokens": 81920, "allocation_passed": false, "recall": [], "fully_passed_recall_tiers": [], "max_recall_validated_prompt_tokens": 0, "status": "failed", "retrieval_passed": null, "startup_failure": "strata-q36: missing qwen35moe.nextn_predict_layers" } ], "official": [], "quant": "APEX Compact", "repository": "LuffyTheFox/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-GGUF", "repository_revision": "5d97a7ca89c5036f8aa4b73e75cc478979ad32a0", "file_name": "Hermes3.6-35B-A3B-Uncensored-Genesis-APEX-Compact.gguf", "file_bytes": 17441463424, "file_sha256": "366ed21cf8ca7e2e226ae48ba17fb377d7b66557c6af9551709c36cc162fe732", "requested_mtp": 0, "metadata_compatibility": null, "highest_recall_validated_prompt_tokens": 0, "highest_allocation_with_successful_recall": 0, "observed_mtp_drafts_offered": 0, "observed_mtp_drafts_accepted": 0, "provider": "LuffyTheFox" } ], "rankings": { "balanced_score": [ "qwen36-ud-q4_k_xl", "coder-iq1_m" ], "humaneval_plus_percent": [ "coder-iq1_m", "qwen36-ud-q4_k_xl" ] }, "role_recommendations": { "CFO / Analyst": "qwen36-ud-q4_k_xl", "Researcher / Scout": "qwen36-ud-q4_k_xl", "Hermes Coordinator": "qwen36-ud-q4_k_xl", "Librarian / Knowledge": "qwen36-ud-q4_k_xl", "Forge / Coder": "coder-iq1_m" }, "speed_ranking_at_64k": [ { "id": "qwen36-ud-q4_k_m-mtp", "label": "Qwen3.6 UD-Q4_K_M (MTP file)", "decode_tps": 27.5 }, { "id": "qwen36-ud-q3_k_m-zero-mtp", "label": "Qwen3.6 UD-Q3_K_M (no MTP)", "decode_tps": 27.4 }, { "id": "qwen36-ud-q3_k_xl-zero-mtp", "label": "Qwen3.6 UD-Q3_K_XL (no MTP)", "decode_tps": 26.4 }, { "id": "qwen36-ud-q4_k_xl", "label": "Qwen3.6 UD-Q4_K_XL", "decode_tps": 24.6 }, { "id": "coder-iq1_m", "label": "Flash-Next Coder IQ1_M", "decode_tps": 21.2 } ], "conclusions": [ "Qwen3.6 Q4_K_XL leads the two-model weighted practical/quality index (96.44 versus 95.08).", "Coder leads HumanEval+ (152/164 versus 130/164) and validated longer retrieval (125573 prompt tokens under 131072 allocation).", "The extra Q3/Q4 runs establish runnable alternatives and context limits, not their quality rankings." ], "caveats": [ "Two models received the full quality suite. Q3_K_M, Q4_K_M and Q3_K_XL received only feasibility, speed and recall tests; their quality is unmeasured here.", "The custom v2.1 applied-task rubric and weighted index are not official Strata quality benchmarks. Official refers to the unmodified upstream speed tool; HumanEval+ is separately scored with EvalPlus.", "HumanEval+ here is one greedy sample per Python task, not a comprehensive software-engineering or end-to-end coding-agent evaluation.", "Coder is a pruned Flash-Next variant; Qwen3.6 is a different model. These are deployment choices, not a controlled quant-only quality comparison.", "Q3 non-MTP compatibility copies add only nextn_predict_layers=0. Original model files are preserved and full tensor-data equality is verified.", "Coder retained one practical transport/engine-stall failure among 72 attempts, and three practical output-limit finishes. Its failure was not replaced with a successful retry.", "MTP requested differs from MTP observed. Qwen3.6 MTP files reported zero drafts offered/accepted in the official measurements; Coder showed active drafts. No on/off ablation was run.", "Official short-chat speed is context-allocation dependent. Compare 32K with 32K and 64K with 64K. Small differences and these few samples do not establish statistical significance.", "A context allocation is distinct from actual exercised prompt length and retrieval validation. Highest tested success is not the exact maximum.", "Three needle positions are a narrow recall test, not a guarantee of all long-context tasks.", "RAM readings are sampled Docker cgroup usage, including counted cache; they are not a measured sustained peak or total PC RAM. VRAM peak is not established.", "The two-client official test used one engine slot and queued requests; it is not two-slot engine concurrency.", "Legacy daily_driver_eligible=false reflects unavailable metadata in the old aggregator; it is not a measured suitability verdict.", "the earlier RTX 5070 Ti system used an RTX 5070 Ti 16 GB, more system RAM, and Strata v0.1.43; do not attribute differences solely to GPU vendor or VRAM.", "RAM rated timings are 16-20-20-38 at 1.35V; active timings and XMP voltage were not verified.", "Dossier timing includes repeated-fixture cache effects. No adaptive-policy timing comparison was run." ], "sources": [ "https://github.com/Niko1221/Strata/tree/477ac13cbe8e582c20484871ce459d20d82df09d", "https://huggingface.co/ISTA-DASLab/Qwen3.8-Flash-Next-GSQ-RCO-Coder-GGUF", "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-MTP-GGUF", "https://huggingface.co/unsloth/Qwen3.6-35B-A3B-GGUF", "https://huggingface.co/LuffyTheFox/Qwen3.6-35B-A3B-Uncensored-Genesis-Hermes-GGUF", "https://www.corsair.com/us/en/p/memory/cmk32gx4m2e3200c16/vengeance-lpx-32gb-2-x-16gb-ddr4-dram-3200mhz-c16-memory-kit-black-cmk32gx4m2e3200c16" ] }