{ "format": "neural-download-model-family-v1", "id": "minimax-m2-7", "primary_packet_id": "minimax-m27-b70-110tps-ubuntu24-20260523", "name": "MiniMax M2.7", "display_name": "MiniMax M2.7", "publisher": "MiniMax AI", "updated_at": "2026-08-23", "summary": "MiniMax M2.7, a 229B mixture-of-experts (about 10B active per word) built for long conversations and agent-style work. Needs all four Arc Pro B70 cards in 4-bit form.", "architecture": { "class": "MiniMaxM2ForCausalLM", "model_type": "minimax_m2", "type": "sparse mixture of experts", "layers": 62, "hidden_size": 3072, "evidence": "repro/minimax-m27-b70-110tps-ubuntu24-20260523/README.md" }, "weight_revisions": [ { "id": "minimax-m2.7-lasimeri", "label": "MiniMax M2.7 · Lasimeri AutoRound export", "role": "measured deployment weights", "model_manifest": "repro/minimax-m27-b70-89tps-20260520/manifests/model-pin.json", "identity_limit": "The immutable revision pins future downloads but is not proven byte-identical to the historical May run because that run did not retain a complete model payload manifest." } ], "transfer_scope": { "status": "one packaged AutoRound revision; workload and runtime identities remain separate", "transfers": [ "MiniMax M2 vLLM/XPU model integration", "AutoRound W4A16 routing", "the retained MoE, graph, collective, and output-path source snapshots" ], "does_not_transfer": [ "the historical model payload identity beyond the recorded repository and configuration hashes", "89 tok/s strict-speed performance to the fresh 32K service stack", "constrained-output speed to free-form generation", "short-prompt decode to near-32K requests", "AutoRound quality or speed to GGUF IQ4 or compressed-KV research" ], "evidence": "results/minimax-m27-int4-autoround-b70/README.md" }, "model_signals": { "b70_fit": { "band": "four-card", "scope": "single active generation on four B70 cards", "basis": "The maintained native vLLM/XPU lane is TP4 and the deployable service reaches a configured 32K context only after reclaiming display VRAM.", "reviewed_at": "2026-08-23" }, "quality_evidence": { "band": "strong-scoped-evidence", "scope": "exact token hashes, deterministic repeats, semantic, arithmetic, and extended canaries for the packaged AutoRound lane; not a general capability score", "evidence": [ "repro/minimax-m27-b70-89tps-20260520/results/promoted-result-20260519.json", "repro/minimax-m27-b70-110tps-ubuntu24-20260523/results/summary-20260523.json" ] }, "popularity": { "state": "not-scored", "reason": "No dated family-level popularity snapshot is stored." } }, "dimensions": { "weight_revision": ["minimax-m2.7-lasimeri"], "weight_quantization": ["AutoRound W4A16 INT4"], "runtime": ["vLLM XPU 0.20.1-local", "vLLM XPU c51df430 deployable stack"], "deployment_profile": ["historical strict 2K", "deployable 32K", "production c1 32K", "constrained structured output"], "tp": [1, 2, 4], "mtp": [0], "active_context_tokens": [372, 510, 512, 2036, 8144, 16288, 18924, 32264, 32408], "configured_max_context_tokens": [2048, 32768, 33792], "graph": ["on"], "kv": ["f16-family"] }, "packets": [ { "id": "minimax-m27-b70-89tps-20260520", "label": "MiniMax M2.7 AutoRound · strict TP4", "revision": "minimax-m2.7-lasimeri", "quantization": "AutoRound W4A16 INT4", "runtime": "vLLM XPU 0.20.1-local", "cards": 4, "status": "historical-strict-speed-candidate", "evidence_level": "B70-verified historical", "coverage": ["decode", "exact-token quality", "semantic quality", "retained source"], "projection": { "model": "minimax_m2.7", "quant": "AutoRound INT4", "runtime": "vllm", "spec": "none", "prompt_tokens": 512, "output_tokens": 1536 }, "manifest": "packages/minimax-m27-int4-autoround-b70/package.json" }, { "id": "minimax-m27-b70-110tps-ubuntu24-20260523", "label": "MiniMax M2.7 AutoRound · deployable TP4", "revision": "minimax-m2.7-lasimeri", "quantization": "AutoRound W4A16 INT4", "runtime": "vLLM XPU c51df430 deployable stack", "cards": 4, "status": "qualified-deployment-reference", "evidence_level": "B70-verified", "coverage": ["decode", "prefill", "TTFT", "configured 32K", "quality", "OpenAI endpoint"], "manifest": "repro/minimax-m27-b70-110tps-ubuntu24-20260523/README.md" }, { "id": "minimax-m27-production-c1-service", "label": "MiniMax M2.7 · production c1 profile", "revision": "minimax-m2.7-lasimeri", "quantization": "AutoRound W4A16 INT4", "runtime": "vLLM XPU service", "cards": 4, "status": "banked-service-profile", "evidence_level": "B70-measured operations", "coverage": ["decode", "TTFT", "near-32K", "serialized concurrency", "operations"], "manifest": "docs/minimax-production-c1-service.md" }, { "id": "minimax-m27-structured-html", "label": "MiniMax M2.7 · constrained HTML", "revision": "minimax-m2.7-lasimeri", "quantization": "AutoRound W4A16 INT4", "runtime": "vLLM XPU constrained output", "cards": 4, "status": "research-only-task-specific", "evidence_level": "B70-measured scoped task", "coverage": ["constrained decode", "30/30 accepted", "zero rejects", "not free-form comparable"], "manifest": "notes/2026-05-22-minimax-structured-fast-lane-regex2.md" } ], "run_measurements": [ { "id": "minimax-strict-tp4-2k-four-repeat", "state": "lab-measured", "revision": "minimax-m2.7-lasimeri", "variant": "AutoRound W4A16 INT4", "runtime": "vLLM XPU 0.20.1-local; base c51df430; llm-scaler 4bfc007", "config": {"tp": 4, "mtp": 0, "graph": "on", "kv": "f16-family", "active_context_tokens": 512, "configured_max_context_tokens": 2048}, "workload": "four clean warm p512/n1536 batch-one runs at 2K context after the strict quality gate", "metrics": { "decode_tok_s": [88.92723924469661, 89.39667728272477, 89.52732050585546, 89.40554449051147] }, "reported_total_tok_s": [118.56965232626214, 119.19556971029968, 119.36976067447394, 119.2073926540153], "quality": "exact n64/n256 hashes, semantic suite, arithmetic repeat, and extended sixpack passed; deterministic; no quality-reducing changes", "evidence": "repro/minimax-m27-b70-89tps-20260520/results/promoted-result-20260519.json" }, { "id": "minimax-deployable-tp4-2k-four-repeat", "state": "lab-measured", "revision": "minimax-m2.7-lasimeri", "variant": "AutoRound W4A16 INT4", "runtime": "vLLM XPU c51df430 deployable Ubuntu 24 stack", "config": {"tp": 4, "mtp": 0, "graph": "on", "kv": "f16-family", "active_context_tokens": 512, "configured_max_context_tokens": 32768}, "workload": "four p512/n1536 batch-one comparison runs at 2K active context; endpoint configured to serve 32K", "metrics": { "decode_tok_s": [82.36975887624021, 83.47327548262813, 83.29957230411637, 83.5461306449009] }, "reported_total_tok_s": [109.82634516832029, 111.29770064350417, 111.0660964054885, 111.39484085986787], "quality": "quality gate passed; the 32K configured capacity is not the active context of this speed gate", "evidence": "repro/minimax-m27-b70-110tps-ubuntu24-20260523/results/summary-20260523.json" }, { "id": "minimax-deployable-tp4-32k-endpoint-metrics", "state": "lab-measured", "revision": "minimax-m2.7-lasimeri", "variant": "AutoRound W4A16 INT4", "runtime": "vLLM XPU c51df430 OpenAI-compatible endpoint", "config": {"tp": 4, "mtp": 0, "graph": "on", "kv": "f16-family", "active_context_tokens": 510, "configured_max_context_tokens": 32768}, "workload": "one warm streamed p510/n1536 endpoint request at the 32K served setting", "metrics": { "decode_tok_s": [85.45305752948855], "prefill_tok_s": [1445.6338234567852], "ttft_ms": [351.06825828552246] }, "quality": "prefill is a conservative prompt_tokens/TTFT lower bound that includes scheduling and first-token work; single performance observation, not a repeat distribution", "evidence": "data/minimax-m27-openai-endpoint-metrics-32k-20260524.json" }, { "id": "minimax-structured-html-tp4", "state": "lab-measured", "revision": "minimax-m2.7-lasimeri", "variant": "AutoRound W4A16 INT4", "runtime": "vLLM XPU constrained regex lane", "config": {"tp": 4, "mtp": 0, "graph": "on", "kv": "f16-family"}, "workload": "short constrained structured-HTML output; task-specific accepted-output accounting", "metrics": {"decode_tok_s": [94.406]}, "quality": "30/30 accepted with zero rejects; not comparable to unconstrained p512/n1536 generation", "evidence": "notes/2026-05-22-minimax-structured-fast-lane-regex2.md" } ], "series_measurements": [ { "id": "minimax-production-c1-context-observations", "state": "lab-measured", "revision": "minimax-m2.7-lasimeri", "variant": "AutoRound W4A16 INT4", "runtime": "vLLM XPU production c1 service", "config": {"tp": 4, "mtp": 0, "graph": "on", "kv": "f16-family", "configured_max_context_tokens": 32768}, "workload": "three production service scenarios with different prompt and output lengths; discrete observations, not a controlled context curve", "points": [ {"x": 372, "decode_tok_s": 96.82, "ttft_ms": 227}, {"x": 18924, "decode_tok_s": 77.89, "ttft_ms": 13085}, {"x": 32264, "decode_tok_s": 63.91, "ttft_ms": 23336} ], "point_workloads": { "372": "512 output tokens", "18924": "16 output tokens", "32264": "64 output tokens" }, "quality": "operational service observations; near-32K request was accepted by LocalMaxxing, but the three rows are different workload classes", "evidence": "docs/minimax-production-c1-service.md" }, { "id": "minimax-deployable-prefill-lower-bounds", "state": "lab-measured", "revision": "minimax-m2.7-lasimeri", "variant": "AutoRound W4A16 INT4", "runtime": "vLLM XPU OpenAI-compatible endpoint", "config": {"tp": 4, "mtp": 0, "graph": "on", "kv": "f16-family", "configured_max_context_tokens": 32768}, "workload": "warm /v1/completions calls with max_tokens=1; conservative prompt-processing lower bounds including HTTP and one generated token", "points": [ {"x": 510, "prefill_tok_s": 1607.9}, {"x": 2036, "prefill_tok_s": 1750.1}, {"x": 8144, "prefill_tok_s": 1830.3}, {"x": 16288, "prefill_tok_s": 1810.3} ], "quality": "prompt-processing characterization only", "evidence": "notes/2026-05-23-current-host-pcie4-prefill-check.md" } ], "views": [ { "id": "minimax-deployment-profiles", "title": "Deployment profiles", "subtitle": "TP4 · p512/n1536 comparison gate · active context 2K; x is configured service capacity, not active prompt length", "x_label": "configured max context tokens", "discrete": true, "metrics": ["decode_tok_s"], "series": [ {"label": "historical strict stack", "measurement_ids": ["minimax-strict-tp4-2k-four-repeat"], "x_from": "config.configured_max_context_tokens"}, {"label": "deployable 32K stack", "measurement_ids": ["minimax-deployable-tp4-2k-four-repeat"], "x_from": "config.configured_max_context_tokens"} ] }, { "id": "minimax-endpoint-metrics", "title": "32K endpoint observation", "subtitle": "TP4 · configured 32K · warm p510/n1536 request; prefill is a conservative lower bound", "x_label": "active prompt tokens", "discrete": true, "metrics": ["decode_tok_s", "prefill_tok_s", "ttft_ms"], "series": [ {"label": "OpenAI endpoint", "measurement_ids": ["minimax-deployable-tp4-32k-endpoint-metrics"], "x_from": "config.active_context_tokens"} ] }, { "id": "minimax-production-context", "title": "Production service observations", "subtitle": "TP4 production c1 · different output lengths at each prompt size; discrete capacity observations, not an interpolated context curve", "x_label": "prompt tokens", "discrete": true, "metrics": ["decode_tok_s", "ttft_ms"], "series": [ {"label": "production c1", "measurement_ids": ["minimax-production-c1-context-observations"]} ] }, { "id": "minimax-prefill-depth", "title": "Prompt processing", "subtitle": "TP4 OpenAI endpoint · max_tokens=1 · includes HTTP and first-token work, so values are conservative lower bounds", "x_label": "prompt tokens", "metrics": ["prefill_tok_s"], "series": [ {"label": "warm lower bound", "measurement_ids": ["minimax-deployable-prefill-lower-bounds"]} ] } ], "coverage_views": [ { "id": "minimax-autoround-target-only", "label": "AutoRound · target only", "fixed": "MiniMax M2.7 AutoRound W4A16 · vLLM XPU · graph on · f16-family KV · MTP/speculation off. Only TP4 is measured in the maintained family packet.", "decision_note": "Missing TP1/TP2 cells are unknown, not failed or zero.", "evidence": "results/minimax-m27-int4-autoround-b70/README.md", "rows": [0], "columns": [1, 2, 4], "cells": { "0:1": {"state": "missing", "label": "not measured", "reason": "No comparable TP1 AutoRound result is stored."}, "0:2": {"state": "missing", "label": "not measured", "reason": "No comparable TP2 AutoRound result is stored."}, "0:4": {"state": "lab-measured", "label": "strict + service", "evidence_id": "minimax-strict-tp4-2k-four-repeat"} } } ], "family_closures": [ { "selectors": {"revision": "minimax-m2.7-lasimeri", "variant": "AutoRound W4A16 INT4", "runtime": "vLLM XPU c51df430 deployable stack", "tp": 4, "configured_max_context_tokens": 33792}, "state": "closed", "reason": "The 33,792-token startup loaded and entered compile/warmup but did not expose /v1/models within the wait window; 32,768 is the promoted service setting.", "evidence": "repro/minimax-m27-b70-110tps-ubuntu24-20260523/results/context-window-32768-20260523.json" }, { "selectors": {"revision": "minimax-m2.7-lasimeri", "variant": "AutoRound W4A16 INT4", "runtime": "vLLM XPU c51df430 deployable stack", "tp": 4, "max_num_batched_tokens": 1024}, "state": "closed", "reason": "The larger prefill chunk triggered the documented PIECEWISE compile-range/startup failure; 512 remains the stable service setting.", "evidence": "repro/minimax-m27-b70-110tps-ubuntu24-20260523/README.md" } ], "estimates": [] }