{
  "format": "neural-download-model-family-v1",
  "id": "lfm-2-5",
  "primary_packet_id": "lfm25-26b-q8-b70",
  "name": "LFM2.5 2.6B",
  "display_name": "LFM2.5 2.6B",
  "publisher": "Liquid AI",
  "updated_at": "2026-08-27",
  "summary": "Liquid AI's LFM2.5, a compact 2.6B model with a hybrid convolution-and-attention design. Very fast on one Arc Pro B70 and a good fit for lightweight assistants and edge-style workloads.",
  "architecture": {
    "class": "LFM2ForCausalLM",
    "model_type": "lfm2",
    "type": "hybrid convolution and attention",
    "layers": 30,
    "hidden_size": 2048,
    "native_context_tokens": 131072,
    "evidence": "repro/lfm25-26b-q8-b70/README.md"
  },
  "weight_revisions": [
    {
      "id": "lfm2.5-2.6b",
      "label": "LFM2.5 2.6B",
      "role": "current measured weights",
      "model_manifest": "repro/lfm25-26b-q8-b70/model-manifest.json"
    }
  ],
  "transfer_scope": {
    "status": "one measured weight revision; quantization and runtime paths require independent gates",
    "transfers": [
      "stock llama.cpp LFM2 registration and SYCL execution path",
      "the package's model-file verification and one-card launch procedure"
    ],
    "does_not_transfer": [
      "Q8_0 quality or speed to another quantization",
      "the raw llama-bench depth curve to HTTP serving throughput",
      "the 32K measured depth trend to the model's unmeasured 131K native limit",
      "multi-card scaling"
    ],
    "evidence": "repro/lfm25-26b-q8-b70/README.md"
  },
  "model_signals": {
    "b70_fit": {
      "band": "high",
      "scope": "small local model on one B70",
      "basis": "The measured Q8_0 file is 2.88 GB and the stock-upstream package leaves substantial memory headroom on one 32 GiB B70.",
      "reviewed_at": "2026-08-23"
    },
    "quality_evidence": {
      "band": "scoped-deployment-evidence",
      "scope": "Q8_0 objective canaries and repeat stability; not a cross-model capability score",
      "evidence": [
        "repro/lfm25-26b-q8-b70/README.md",
        "experiments/qwen38-27b-b70/data/2026-08-22-neural-download-firstwave-baselines.json"
      ]
    },
    "popularity": {
      "state": "not-scored",
      "reason": "No dated family-level popularity snapshot is stored."
    }
  },
  "dimensions": {
    "weight_revision": ["lfm2.5-2.6b"],
    "weight_quantization": ["Q8_0"],
    "runtime": ["llama.cpp SYCL"],
    "tp": [1, 2, 4],
    "mtp": [0],
    "active_context_tokens": [0, 2048, 4096, 8192, 16384, 24576, 32768],
    "configured_max_context_tokens": [8192, 32768, 131072],
    "graph": ["not separately varied"],
    "kv": ["f16"]
  },
  "packets": [
    {
      "id": "lfm25-26b-q8-b70",
      "label": "LFM2.5 2.6B Q8_0 · TP1",
      "revision": "lfm2.5-2.6b",
      "quantization": "Q8_0",
      "runtime": "llama.cpp SYCL",
      "cards": 1,
      "status": "candidate",
      "evidence_level": "B70-verified",
      "coverage": ["decode", "prefill", "active context through 32K", "quality", "stock upstream"],
      "projection": {
        "model": "lfm2.5_2.6b",
        "quant": "Q8_0",
        "runtime": "llama_cpp",
        "spec": "none"
      , "prompt_tokens": 128, "output_tokens": 128},
      "manifest": "packages/lfm25-26b-q8-b70/package.json"
    }
  ],
  "run_measurements": [
    {
      "id": "lfm25-q8-tp1-serving-standard",
      "state": "lab-measured",
      "revision": "lfm2.5-2.6b",
      "variant": "Q8_0",
      "runtime": "llama.cpp SYCL 9fee29e9435f865ec0b811a783a6471a136d9317",
      "config": {
        "tp": 1,
        "mtp": 0,
        "graph": "not separately varied",
        "kv": "f16",
        "configured_max_context_tokens": 8192
      },
      "workload": "two fresh servers; complete 12-prompt/six-class suite; 512-token cap; median within prompt class then median across class medians; 100-event/99-interval decode window; target-only; cache-zero",
      "metrics": {
        "decode_tok_s": [132.16164589689635, 132.1132674603923]
      },
      "quality": "both workload gates and objective-canary batteries passed; cached_tokens=0 for every request; complete token arrays exact 12/12 across fresh servers",
      "evidence": "data/2026-08-27-lfm25-q8-tp1-strict-headline-result.json"
    }
  ],
  "series_measurements": [
    {
      "id": "lfm25-q8-tp1-f16-depth-sweep",
      "state": "lab-measured",
      "revision": "lfm2.5-2.6b",
      "variant": "Q8_0",
      "runtime": "llama.cpp bench 9fee29e",
      "config": {
        "tp": 1,
        "mtp": 0,
        "graph": "not separately varied",
        "kv": "f16"
      },
      "workload": "raw llama-bench pp2048 plus tg128 at each existing-context depth; FlashAttention on; five repetitions",
      "points": [
        {"x": 0, "decode_tok_s": 135.198135, "prefill_tok_s": 9130.854812},
        {"x": 2048, "decode_tok_s": 131.259551, "prefill_tok_s": 4699.373204},
        {"x": 4096, "decode_tok_s": 127.211224, "prefill_tok_s": 4584.945629},
        {"x": 8192, "decode_tok_s": 120.195857, "prefill_tok_s": 4399.719712},
        {"x": 16384, "decode_tok_s": 107.980065, "prefill_tok_s": 3752.42491},
        {"x": 24576, "decode_tok_s": 98.138413, "prefill_tok_s": 3698.15006},
        {"x": 32768, "decode_tok_s": 89.93812, "prefill_tok_s": 2824.910192}
      ],
      "quality": "raw engine shape characterization, not a language-quality gate or HTTP serving result",
      "evidence": "repro/lfm25-26b-q8-b70/lfm25-26b-q8.sweep.json"
    }
  ],
  "views": [
    {
      "id": "lfm-serving-operating-point",
      "title": "Serving repeat",
      "subtitle": "Q8_0 · TP1 · MTP0 · 8K configured · f16 KV · full 12-prompt/six-class cache-zero suite",
      "x_label": "tensor parallel cards",
      "discrete": true,
      "metrics": ["decode_tok_s"],
      "series": [
        {
          "label": "fresh-server repeats",
          "measurement_ids": ["lfm25-q8-tp1-serving-standard"],
          "x_from": "config.tp"
        }
      ]
    },
    {
      "id": "lfm-context-depth",
      "title": "Active context depth",
      "subtitle": "Q8_0 · TP1 · f16 KV · raw llama-bench pp2048/tg128; raw engine rates are not serving medians",
      "x_label": "existing context tokens",
      "metrics": ["decode_tok_s", "prefill_tok_s"],
      "series": [
        {
          "label": "Q8_0",
          "measurement_ids": ["lfm25-q8-tp1-f16-depth-sweep"]
        }
      ]
    }
  ],
  "coverage_views": [
    {
      "id": "lfm-q8-target-only",
      "label": "Q8_0 · target only",
      "fixed": "LFM2.5 2.6B Q8_0 · llama.cpp SYCL · f16 KV · MTP is not used. TP1 has serving and raw-depth evidence; TP2/TP4 have not been measured.",
      "decision_note": "Missing topologies are unknown, not zero or unsupported.",
      "evidence": "repro/lfm25-26b-q8-b70/README.md",
      "rows": [0],
      "columns": [1, 2, 4],
      "cells": {
        "0:1": {"state": "lab-measured", "label": "serving + depth", "evidence_id": "lfm25-q8-tp1-serving-standard"},
        "0:2": {"state": "missing", "label": "not measured", "reason": "No TP2 LFM2.5 Q8_0 result is stored."},
        "0:4": {"state": "missing", "label": "not measured", "reason": "No TP4 LFM2.5 Q8_0 result is stored."}
      }
    }
  ],
  "family_closures": [],
  "estimates": []
}