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

suite-v2 | index-v4.2

Qwen3.6 35B A3B

qwen3-6-35b-a3b__qwen3-6-35b-a3b-udq4-s2v5

Local Intelligence Index

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

41.0

±2.7 95% CI

Profile: Agentic / Knowledge / Instruction / Coding / Math

Agentic 6.3 / Knowledge 82.0 / Instruction 57.5 / Coding 28.4 / Math 22.3

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

Total run time

9.2 h

Data quality note: this run has 0 error(s) and 11 no-answer item(s).

Data warnings

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

Axis breakdown

Agenticworst axis

n=96 · errors=0 · no answer=0

6.3 ±5.4

Knowledge

n=400 · errors=0 · no answer=5

82.0 ±4.2

Instruction

n=294 · errors=0 · no answer=0

57.5 ±5.8

Coding

n=141 · errors=0 · no answer=0

28.4 ±7.8

Math

n=139 · errors=0 · no answer=6

22.3 ±7.2

Diagnostics · unweighted

Call formatting

n=330 · errors=0 · no answer=2

71.2 ±4.8

BFCL single-turn

— not measured

BFCL v3 multi-turn base — frozen snapshot

n=50 · errors=0 · no answer=2

20.0 ±12.0

BFCL multi-turn long-context

— not measured

RULER 32K

— not measured

Instruction-following decomposition

Strict = Termination × Conditional

Strict accuracy

57.5%

correct AND terminated / all

Termination rate

96.3%

terminated / all

Conditional accuracy

59.7%

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: qwen35moe
quant: UD-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: 6,913,387 prompt / 6,349,939 completion / 13,263,326 total
tokens-to-answer: 3,639 median / 9,377 p95
tok/s: 190.8
total run time: 9.2 h
est cost: n/a
n_items: 1,457
n_errors: 0
n_no_answer: 11

Serving performance

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

prefill

6,098 tok/s

decode

228 tok/s

TTFT proxy

340.1 ms

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

coverage

93.4%

prompt median / p95

340.1 ms / 1,059.6 ms

predicted median / p95

15,668 ms / 41,837.5 ms

bench	prefill	decode	prompt median	n
amo	2,113.7 tok/s	231.3 tok/s	117.8 ms	39
bfcl_multi_turn_base	7,183.9 tok/s	243.8 tok/s	1,027.5 ms	50
bigcodebench_hard	6,931.1 tok/s	232.4 tok/s	707.8 ms	148
ifbench	6,384.2 tok/s	234.1 tok/s	448.4 ms	294
mmlu_pro	5,414.1 tok/s	218 tok/s	292.1 ms	400
olymmath_hard	1,361.1 tok/s	232.1 tok/s	112.8 ms	100
tc_json_v1	6,047.7 tok/s	237 tok/s	334.3 ms	330

Source: llama.cpp server timings.