This receipt lets anyone verify this run against the frozen suite — see Methodology.

suite-v2 | index-v4.2

Gemma 4 31B IT

gemma-4-31b-it__gemma-4-31b-it-q4km-s2v5

Local Intelligence Index

LB-2026-07.2 | 25/22.5/22.5/22.5/7.5

51.7

±3.1 95% CI

Profile: Agentic / Knowledge / Instruction / Coding / Math

Agentic 10.4 / Knowledge 87.4 / Instruction 79.3 / Coding 35.5 / Math 48.2

Weighted headline profile: Agentic 25%, Knowledge 22.5%, Instruction 22.5%, Coding 22.5%, Math 7.5%.

Total run time

27.8 h

Data quality note: this run has 0 error(s) and 7 no-answer item(s).

Data warnings

tool_use chance_corrected differs from item-derived mean; stale/inconsistent run JSON
coding chance_corrected differs from item-derived mean; stale/inconsistent run JSON

Axis breakdown

Agenticworst axis

n=96 · errors=0 · no answer=0

10.4 ±8.5

Knowledge

n=400 · errors=0 · no answer=1

87.4 ±3.6

Instruction

n=294 · errors=0 · no answer=0

79.3 ±4.8

Coding

n=141 · errors=0 · no answer=0

35.5 ±7.8

Math

n=139 · errors=0 · no answer=6

48.2 ±8.6

Diagnostics · unweighted

Call formatting

n=330 · errors=0 · no answer=0

74.2 ±4.8

BFCL single-turn

— not measured

BFCL v3 multi-turn base — frozen snapshot

n=50 · errors=0 · no answer=0

28.0 ±12.0

BFCL multi-turn long-context

— not measured

RULER 32K

— not measured

Instruction-following decomposition

Strict = Termination × Conditional

Strict accuracy

79.3%

correct AND terminated / all

Termination rate

99.3%

terminated / all

Conditional accuracy

79.8%

correct / terminated

Outputs that hit the answer-token cap are counted incorrect; this prevents non-terminating generations from getting credit for matching required tokens inside a runaway response.

Manifest

model: gemma4
quant: Q4_K_M
runtime: llama.cpp b9852/fd1a05791 · KV k=f16,v=f16 · ctx 32,768
hardware: NVIDIA GeForce RTX 5090 (31.8 GB) · Windows-11-10.0.26200-SP0
os: Windows-11-10.0.26200-SP0
lane: bounded-final-v2
thinking_mode: n/a
caps: max_tokens_math: 0, max_tokens_mcq: 16384, thinking_budget: 8192
sampling: temp 0 | top_p n/a | top_k 1 | min_p n/a | seed 1234 | effort n/a | knowledge: max 16384 | instruction: max 16384 | coding: max 16384 | math: max 16384 | bfcl_multi_turn_base: max 16384 | tc_json_v1: max 16384
tokens: 5,527,046 prompt / 4,848,815 completion / 10,375,861 total
tokens-to-answer: 1,918 median / 9,318 p95
tok/s: 48.4
total run time: 27.8 h
est cost: n/a
n_items: 1,457
n_errors: 0
n_no_answer: 7

Serving performance

NVIDIA GeForce RTX 5090 (31.8 GB) · Windows-11-10.0.26200-SP0

prefill

2,425.1 tok/s

decode

63.1 tok/s

TTFT proxy

252.6 ms

prompt processing before first token — non-streaming harness, lower bound

coverage

93.4%

prompt median / p95

252.6 ms / 801.3 ms

predicted median / p95

29,984.3 ms / 152,744.1 ms

bench	prefill	decode	prompt median	n
amo	818.7 tok/s	58 tok/s	290.3 ms	39
bfcl_multi_turn_base	3,089.3 tok/s	63.5 tok/s	1,659.2 ms	50
bigcodebench_hard	2,354.9 tok/s	59.1 tok/s	143.8 ms	148
ifbench	2,265.3 tok/s	66.3 tok/s	197.7 ms	294
mmlu_pro	2,188.9 tok/s	62.3 tok/s	156.8 ms	400
olymmath_hard	612.4 tok/s	65.1 tok/s	255.7 ms	100
tc_json_v1	2,791.1 tok/s	60.2 tok/s	392.1 ms	330

Source: llama.cpp server timings.