{ "format": "neural-download-model-family-v1", "id": "gemma-4", "primary_packet_id": "gemma4-26b-a4b-q8-b70-125tps-20260701", "name": "Gemma 4", "display_name": "Gemma 4 · 26B A4B", "publisher": "Google", "updated_at": "2026-08-25", "summary": "Google's Gemma 4, the 26B mixture-of-experts with about 4B active per word. A strong all-rounder for chat, writing, and homework that runs on a single Arc Pro B70 and can also read images.", "architecture": { "class": "Gemma 4 26B A4B MoE", "design": "sparse mixture of experts", "total_parameters_approx": 25200000000, "active_parameters_approx": 3800000000, "layers": 30, "vocab_size": 262144, "experts": 128, "active_experts": 8, "shared_experts": 1, "trained_context_tokens": 262144, "architecture_modalities": ["text", "image"], "packaged_modalities": ["text"], "evidence": "results/gemma4-26b-a4b-q8-b70/model-options.md" }, "weight_revisions": [ { "id": "gemma4-26b-a4b-it-gguf-3bb10d5", "label": "Gemma 4 26B A4B IT · UD-Q8_K_XL target", "role": "target/verifier checkpoint", "repository": "unsloth/gemma-4-26B-A4B-it-GGUF", "revision": "3bb10d594514ef4edb7f3a65d41a7e4eb8c5767a", "model_manifest": "repro/gemma4-26b-a4b-q8-b70-125tps-20260701/model-manifest.json" } ], "auxiliary_artifacts": [ { "id": "gemma4-26b-a4b-it-mtp-q4-local", "label": "Gemma 4 26B A4B IT · locally derived Q4_0 MTP draft", "role": "speculative draft checkpoint; not a family weight revision or target quantization", "repository": "unsloth/gemma-4-26B-A4B-it-GGUF", "source_revision": "3bb10d594514ef4edb7f3a65d41a7e4eb8c5767a", "identity_status": "reconstructed reference is repeat-stable; historical local draft hash and byte identity were not retained", "model_manifest": "repro/gemma4-26b-a4b-q8-b70-125tps-20260701/model-manifest.json" } ], "transfer_scope": { "status": "single pinned target revision and reconstructed auxiliary draft path", "transfers": [ "The source reconstruction, target-verified MTP path, and B70-specific llama.cpp/SYCL patch stack transfer only when the pinned target revision, draft preparation, runtime revision, and gate identity are preserved.", "The context profile describes the same package family but a different operating profile from the short-context headline." ], "does_not_transfer": [ "performance to other Gemma 4 sizes or target quantizations", "performance to multimodal requests", "the package headline to the long-context service profile", "historical local draft binary identity from the reconstructed reference", "the 262K trained context limit as a measured 262K service result" ], "evidence": "repro/gemma4-26b-a4b-q8-b70-125tps-20260701/README.md" }, "model_signals": { "b70_fit": { "band": "high", "scope": "one-card text deployment on Intel Arc Pro B70", "basis": "The pinned UD-Q8_K_XL target plus Q4_0 MTP draft has a qualified one-card lane and a measured cache-zero service profile through 32,571 active tokens.", "reviewed_at": "2026-08-23" }, "quality_evidence": { "band": "strong-deployment-evidence", "scope": "fixed-suite target-verified MTP output integrity and service-ladder exact JSON retrieval; not a general intelligence or multimodal score", "evidence": [ "data/gemma4-q8-gpu0-finalpostnorm-reproexact-full512-20260701T084728Z/summary.json", "data/gemma4-26b-a4b-q8-b70-context-performance-profile-20260702.json" ] }, "popularity": { "state": "not-scored", "reason": "No dated family-level popularity snapshot is stored for this exact checkpoint and deployment variant." } }, "dimensions": { "weight_revision": ["gemma4-26b-a4b-it-gguf-3bb10d5"], "target_quantization": ["UD-Q8_K_XL"], "draft_quantization": ["Q4_0 MTP"], "runtime": ["llama.cpp SYCL"], "cards": [1, 2, 4], "tp": [1, 2, 4], "speculative_method": ["target-verified MTP"], "draft_depth": [0, 1, 2, 3, 4, 7], "active_context_tokens": [741, 2806, 5643, 10976, 16213, 22730, 30400, 32571], "configured_max_context_tokens": [8192, 32768], "kv": ["f16"] }, "packets": [ { "id": "gemma4-26b-a4b-q8-b70-125tps-20260701", "label": "Gemma 4 26B A4B UD-Q8_K_XL · TP1 + MTP3", "revision": "gemma4-26b-a4b-it-gguf-3bb10d5", "quantization": "UD-Q8_K_XL target + Q4_0 MTP draft", "runtime": "llama.cpp SYCL c926ad09857517978575d6a74d225b463f7417a0", "cards": 1, "status": "candidate", "evidence_level": "B70-verified source reconstruction", "coverage": ["decode", "TTFT", "active context", "prefill approximation", "quality", "recipe"], "projection": {"model": "gemma4_26b_a4b", "quant": "UD-Q8_K_XL", "runtime": "llama_cpp", "spec": "mtp:3", "prompt_tokens": 128, "output_tokens": 128}, "manifest": "packages/gemma4-26b-a4b-q8-b70/package.json" }, { "id": "gemma4-26b-a4b-q8-b70-95tps-20260624", "label": "Gemma 4 26B A4B UD-Q8_K_XL · historical TP1 + MTP7", "revision": "gemma4-26b-a4b-it-gguf-3bb10d5", "quantization": "UD-Q8_K_XL target + Q4_0 MTP draft", "runtime": "llama.cpp SYCL c926ad09857517978575d6a74d225b463f7417a0", "cards": 1, "status": "superseded-historical-reproduction", "evidence_level": "B70-verified history", "coverage": ["decode", "quality", "recipe"], "manifest": "repro/gemma4-26b-a4b-q8-b70-95tps-20260624/README.md" } ], "run_measurements": [ { "id": "gemma4-q8-tp1-mtp3-short-record", "state": "lab-measured", "revision": "gemma4-26b-a4b-it-gguf-3bb10d5", "variant": "UD-Q8_K_XL target + locally derived Q4_0 MTP draft", "runtime": "llama.cpp SYCL c926ad09857517978575d6a74d225b463f7417a0", "config": {"cards": 1, "tp": 1, "mtp": 3, "kv": "f16", "flash_attention": "on", "configured_max_context_tokens": 32768}, "workload": "12 unique cold prompts, 512-token output ceiling, cache zero; conventional 99-interval headline", "metrics": { "decode_tok_s": [122.16035656735696], "ttft_ms": [178.6938319564797] }, "quality": "512/512 canary passed; class-balanced median of input-class medians. The all-prompt 99-interval median is 123.72736943965285 tok/s and the historical 100-event compatibility figure is 124.97714084813418 tok/s; neither is the current headline", "evidence": "data/gemma4-q8-gpu0-finalpostnorm-reproexact-full512-20260701T084728Z/summary.json" }, { "id": "gemma4-q8-tp1-mtp7-historical-record", "state": "lab-measured", "revision": "gemma4-26b-a4b-it-gguf-3bb10d5", "variant": "UD-Q8_K_XL target + locally derived Q4_0 MTP draft", "runtime": "llama.cpp SYCL c926ad09857517978575d6a74d225b463f7417a0", "config": {"cards": 1, "tp": 1, "mtp": 7, "kv": "f16", "configured_max_context_tokens": 8192}, "workload": "historical filled-long prompt, 588 actual prompt tokens, 512 output tokens; first fresh response", "metrics": {"decode_tok_s": [95.26352416631231]}, "quality": "384/384 chat canary passed and cached_tokens=0; superseded by the current package", "evidence": "repro/gemma4-26b-a4b-q8-b70-95tps-20260624/results/summary-20260624T081218Z.json" } ], "series_measurements": [ { "id": "gemma4-q8-tp1-mtp-service-context", "state": "lab-measured", "revision": "gemma4-26b-a4b-it-gguf-3bb10d5", "variant": "UD-Q8_K_XL target + locally derived Q4_0 MTP draft", "runtime": "llama.cpp SYCL service stack", "config": {"cards": 1, "tp": 1, "kv": "f16", "configured_max_context_tokens": 32768, "samples_per_point": 4}, "workload": "fixed deterministic long-context suite; each prompt once; cache zero; exact JSON retrieval; one active request; 128-token output ceiling; arithmetic mean across four independent one-B70 lanes; prefill approximated as prompt tokens divided by TTFT", "axis": "active_context_tokens", "points": [ {"x": 741, "decode_tok_s": 161.89843663684636, "prefill_tok_s": 847.9619132833765, "ttft_ms": 873.9445414976217}, {"x": 2806, "decode_tok_s": 144.8662058096366, "prefill_tok_s": 1401.0150752344389, "ttft_ms": 2002.938948746305}, {"x": 5643, "decode_tok_s": 141.11391660033576, "prefill_tok_s": 1469.4385849395155, "ttft_ms": 3840.5072987661697}, {"x": 10976, "decode_tok_s": 135.93492558486764, "prefill_tok_s": 1330.3881630879202, "ttft_ms": 8250.644568965072}, {"x": 16213, "decode_tok_s": 127.63665710590647, "prefill_tok_s": 1256.9527613956354, "ttft_ms": 12899.11724528065}, {"x": 22730, "decode_tok_s": 120.48241930243164, "prefill_tok_s": 1128.9763236685953, "ttft_ms": 20134.24033729825}, {"x": 30400, "decode_tok_s": 114.00353788462567, "prefill_tok_s": 1028.1957534609096, "ttft_ms": 29567.641104455106}, {"x": 32571, "decode_tok_s": 114.8486529751413, "prefill_tok_s": 1001.687171893108, "ttft_ms": 32517.333841737127} ], "quality": "validated service sweep; this is a different operating profile from the short-context package headline", "evidence": "data/gemma4-26b-a4b-q8-b70-context-performance-profile-20260702.json" } ], "views": [ { "id": "gemma4-context-profile", "title": "Active-context service profile", "subtitle": "Gemma 4 UD-Q8_K_XL target + Q4_0 MTP draft · TP1 llama.cpp · four-lane arithmetic means; prefill is prompt tokens / TTFT", "x_label": "active context tokens", "metrics": ["decode_tok_s", "prefill_tok_s", "ttft_ms"], "series": [ {"label": "measured service profile", "measurement_ids": ["gemma4-q8-tp1-mtp-service-context"]} ] }, { "id": "gemma4-record-history", "title": "Qualified short-record history", "subtitle": "Distinct MTP depths and context contracts; values are shown as historical operating points, not a controlled depth comparison", "x_label": "draft tokens", "discrete": true, "metrics": ["decode_tok_s"], "series": [ {"label": "qualified records", "measurement_ids": ["gemma4-q8-tp1-mtp3-short-record", "gemma4-q8-tp1-mtp7-historical-record"], "x_from": "config.mtp"} ] } ], "coverage_views": [ { "id": "gemma4-udq8kxl-mtp-by-tp", "label": "MTP depth × TP", "row_axis": {"key": "mtp", "label": "MTP", "prefix": "MTP"}, "column_axis": {"key": "tp", "label": "TP", "prefix": "TP"}, "fixed_selectors": { "revision": "gemma4-26b-a4b-it-gguf-3bb10d5", "variant": "UD-Q8_K_XL target + locally derived Q4_0 MTP draft", "runtime": "llama.cpp SYCL c926ad09857517978575d6a74d225b463f7417a0", "kv": "f16", "configured_max_context_tokens": 32768 }, "fixed": "Gemma 4 26B A4B · UD-Q8_K_XL target + locally derived Q4_0 MTP draft · llama.cpp SYCL · f16 KV · 32K configured ceiling.", "decision_note": "MTP3 at TP1 is the current strict package reference. Every other core MTP0–4 × TP1/2/4 cell remains explicitly unmeasured for this exact identity; the distinct historical MTP7 operating point stays in the record-history chart.", "rows": [0, 1, 2, 3, 4], "columns": [1, 2, 4], "cells": { "0:1": {"state": "missing", "label": "not measured", "reason": "No normalized qualified target-only result exists for this exact runtime and 32K profile."}, "0:2": {"state": "missing", "label": "not measured", "reason": "No qualified TP2 profile exists for this exact identity."}, "0:4": {"state": "missing", "label": "not measured", "reason": "No qualified TP4 profile exists for this exact identity."}, "1:1": {"state": "missing", "label": "not measured", "reason": "No normalized qualified MTP1 result exists for this exact identity."}, "1:2": {"state": "missing", "label": "not measured", "reason": "No qualified TP2 profile exists for this exact identity."}, "1:4": {"state": "missing", "label": "not measured", "reason": "No qualified TP4 profile exists for this exact identity."}, "2:1": {"state": "missing", "label": "not measured", "reason": "No normalized qualified MTP2 result exists for this exact identity."}, "2:2": {"state": "missing", "label": "not measured", "reason": "No qualified TP2 profile exists for this exact identity."}, "2:4": {"state": "missing", "label": "not measured", "reason": "No qualified TP4 profile exists for this exact identity."}, "3:1": {"state": "lab-measured", "label": "strict", "evidence_id": "gemma4-q8-tp1-mtp3-short-record"}, "3:2": {"state": "missing", "label": "not measured", "reason": "No qualified TP2 profile exists for this exact identity."}, "3:4": {"state": "missing", "label": "not measured", "reason": "No qualified TP4 profile exists for this exact identity."}, "4:1": {"state": "missing", "label": "not measured", "reason": "No normalized qualified MTP4 result exists for this exact identity."}, "4:2": {"state": "missing", "label": "not measured", "reason": "No qualified TP2 profile exists for this exact identity."}, "4:4": {"state": "missing", "label": "not measured", "reason": "No qualified TP4 profile exists for this exact identity."} } } ], "family_closures": [ { "selectors": {"revision": "gemma4-26b-a4b-it-gguf-3bb10d5", "cards": [2, 3, 4]}, "state": "missing", "reason": "No qualified multi-card performance profile is stored for this exact target, draft, runtime, and gate identity.", "evidence": "packages/gemma4-26b-a4b-q8-b70/package.json" }, { "selectors": {"revision": "gemma4-26b-a4b-it-gguf-3bb10d5", "active_context_tokens_above": 32571}, "state": "missing", "reason": "The architecture's trained context capacity is not a measurement; the stored service sweep ends at 32,571 active tokens.", "evidence": "data/gemma4-26b-a4b-q8-b70-context-performance-profile-20260702.json" }, { "selectors": {"revision": "gemma4-26b-a4b-it-gguf-3bb10d5", "modalities": "image"}, "state": "missing", "reason": "The packaged and measured deployment is text-only; no multimodal package result is stored.", "evidence": "packages/gemma4-26b-a4b-q8-b70/package.json" }, { "selectors": {"packet": "gemma4-26b-a4b-q8-b70-125tps-20260701", "replay": "clean-host"}, "state": "missing", "reason": "The current package has not completed a clean-host rebuild and endpoint replay, and the historical local Q4_0 draft identity was not retained.", "evidence": "packages/gemma4-26b-a4b-q8-b70/package.json" } ], "estimates": [] }