isilicon/mango-evaluation-results

Mango · Model tests, results and scoring

The public evidence archive for Mango's local-model evaluation. It contains the complete current matched comparison, per-question scores, paired uncertainty, exact model/file catalog, the Qwen effort experiment, two-grader subjective results and every committed legacy numerical experiment. Historical snapshots and unfinished tests have their own status; they are not extra independent samples or zero-quality models. New verified experiments belong here as well as on their model cards.

This release: 12 matched 202-question configurations · 3,232 per-question records and 168 agent-case results across distinct cohorts · 268 legacy experiments · 32 model/artifact entries · 1,283 indexed test definitions.

Model/file catalog · Per-question grades · Configuration table · Full reports · Legacy experiments · Tests · Scoring code · Methodology

Join Infinity for support, model requests and benchmark updates.

Latest systems study · Flash-Next 32K prefill

Read the native prefill study →

Matched 8,000-token cold-prefix native prefill: 79.606 → 23.103 seconds (3.446×) in the adaptive 32K allocated route. Fresh live validation measured 22.207 seconds. A 20K appended turn reused 19,980 tokens and processed 48 in 0.753 seconds. Fast settings remain unchanged.

One sweep run per setting, warm OS file cache—not disk-cold or a p95 estimate. This systems diagnostic is separate from canonical quality scores and the earlier 64K optimization; do not multiply the gains. The study separates 232/232 offline tests from live/recovery checks and explains cache loss on model switch, timing definitions and reproducibility limits.

Latest study · Flash capacity bookkeeping

Read the native study: methods, results and limitations →

Twelve fresh scheduling cases, three matched seeds and 72 evaluation responses. This prompt-only experiment is separate from the canonical leaderboard; no weights changed.

Baseline	Bookkeeping prompt	Paired difference	Mean elapsed time
31/36 · 86.1%	34/36 · 94.4%	+8.3 pp; 95% case-bootstrap interval −2.8 to +19.4 pp	57.3 → 69.7 s · about 22% longer

Inconclusive general benefit: the interval includes zero, and the statistical sample is 12 cases. Bookkeeping used about 21% more completion tokens and 24% more total tokens; prompt length and extra effort are confounds. Case-level data and audit scope · Supplementary PDF.

Current matched 202-question comparison

Each configuration answered the same 202 selected questions plus 12 mechanical agent cases. The weighted score uses social/EQ (30%), cyber (20%), agentic (30%), coding (10%), instruction following (5%) and general reasoning (5%). Each domain averages its component tasks. These are small convenience samples with local graders: the checkpoint, quantization, engine, context, draft and effort settings define the result.

Rank¹	Exact configuration	Weighted /100	Social/EQ	Cyber	Agentic	Coding	Instructions	General	Workload min
1	Flash Next · mradermacher	84.19	74.41	83.33	88.75	95.00	85.00	96.43	35.33
2	Flash Next · direct Orca	83.42	71.15	85.00	87.50	95.00	90.00	96.43	37.36
3	DeepSeek GSQ IQ2_XS	83.19	74.98	87.22	85.00	85.00	95.00	90.00	71.43
4	Historical GLM · capped	82.55	77.10	81.11	86.25	90.00	70.00	96.43	27.49
5	Qwen122B Q4_K	80.72	71.19	83.33	79.58	90.00	100.00	96.43	120.76
6	Qwen27 MLX4 + DFlash2	80.64	69.68	86.11	83.75	85.00	95.00	82.86	23.98
7	MiMo Flash Q3_K_M	79.39	65.47	85.00	82.50	90.00	90.00	90.00	220.28
8	GLM · Bolding IQ2_XXS	79.11	70.34	85.56	85.00	85.00	65.00	72.86	40.98
9	GLM · Orca Q2_K	75.64	68.60	81.11	78.75	80.00	65.00	79.29	54.76
10	Cyber Qwen · IQ4_XS + MTP	70.27	43.83	82.22	85.00	85.00	55.00	78.57	29.74
11	Nanbeige33 Q6_K	67.46	54.52	78.89	61.67	85.00	80.00	86.43	175.31
12	Cyber Qwen v2 · Q6_K + Q8 MTP	66.98	38.68	75.56	83.75	85.00	70.00	62.86	45.60

¹ Ranks show point estimates. The leading intervals overlap, and these tests establish no universal winner. The earlier Flash configuration remains the recorded choice under the project's quality/time rule. The current report contains all 66 paired comparisons and task-level intervals. Passing all 12 mechanical cases does not establish complete semantic agent competence. The historical GLM weights are unavailable and that run used capped thinking. MiMo retains two genuine length finishes; Nanbeige 33B Q6 retains three. Infrastructure failures are tracked separately.

Cyber Qwen findings

The Cyber IQ4_XS-imatrix configuration with embedded MTP scored 70.27/100 in 29.74 workload minutes. Cyber v2 Q6 with its separate Q8 MTP head scored 66.98/100 in 45.60 minutes. Both passed 12 mechanical cases with zero infrastructure errors or length finishes. Their social/EQ and instruction scores were lower than the general-purpose Qwen MLX4 configuration's results.

Each Cyber run produced three malformed or out-of-range EQ responses out of 30. Formatting does not explain the entire EQ gap. The pinned upstream EQ formula permits negative item values; four IQ4 and five Q6 answers received them, without clipping. These Cyber releases have different declared parent lineages, so their difference cannot be attributed to quantization alone. Their 22 open-ended responses per configuration are collected; independent Claude and Codex grading remains pending.

Separate Qwen effort study · 403 valid questions

With the same Orca MLX4 checkpoint, TensorFold 1.1.6, DFlash2 and 64K context, medium scored 82.87 and xhigh scored 84.41. Both collected 404 questions plus 12 agents. One invalid reference annotation is unscored in both runs, leaving 403 valid questions and strict IFEval results of 36/39 and 37/39.

The paired difference is +1.54 points, with a 95% interval of [−1.71, +4.82]. The fresh matched 202-question workload took 4.46× longer at xhigh. Medium remains the practical default. This larger cohort stays separate from the 202-question comparison above.

Blind subjective panel · complete

Claude and Codex each froze 44 independent ratings before unblinding: 22 prompts per configuration. Their equal-weight means are 7.77/10 for medium and 8.18/10 for xhigh. Both planned raters have finished; the owner has no grading step.

The medium/xhigh category means are interpersonal judgment 7.50/8.00, creative writing 7.50/7.08, faithfulness 9.90/9.10 and grounded inference 6.30/8.80. These are separate subjective results from 22 prompts and two AI raters. The panel report includes coverage, per-rater means, agreement and uncertainty.

All tests and historical evidence

The tables index every frozen suite/item ID and prompt hash. Suite manifests, selection/scoring code and upstream dataset references establish the protocol. Mango-authored judge tasks and mechanical-agent fixtures are included. Upstream third-party question text follows its source dataset terms; retrieve it through the supplied builder and source references. The core report contains every per-question grade, timing, truncation and exclusion—not only model averages.

All 268 committed legacy result JSONs retain their numerical, boolean and controlled status evidence in legacy/, with original source hashes. Unknown textual payloads and generated text are fingerprinted rather than exposed. They include partial/failure/probe experiments, with different sample sizes and protocols; they are not pooled into the current leaderboard. Historical report snapshots preserve prior analyses; the old full404 report is explicitly superseded by the annotation correction.

GPQA, logic, long context, HLE, planning, hallucination, search/load, sandbox and GUI have definitions or queued work but no invented completion scores. See the suite/configuration tables for actual coverage. Every ungraded weight and companion stays ungraded or not-applicable; a parent model's score never becomes a component's standalone score. The catalog covers all 23 original mirrors, the eight files in the combined Qwen repo and the unavailable historical GLM configuration. Earlier model/quant/source references—including exploratory builds outside the current preserved set—are indexed in legacy model references. A reference is not a completed benchmark; its evidence files retain the recorded status and numerical results.

Integrity and maintenance

ARCHIVE_MANIFEST.json hashes every published evidence file and records source coverage. VALIDATION.json checks report arithmetic, all-model coverage, baseline preservation and payload policy. No model weights are duplicated here. Update this archive and affected model cards at each verified reporting boundary. Original model authors, publishers, quantizers, dataset authors and reference-grader authors retain their attribution and licenses. Apache-2.0 covers Mango's archive/scoring contributions; upstream materials retain their terms.

Collection snapshot

Every saved run has explicit counts and status. The preserved checkpoint includes four Nanbeige Cyber answers, one genuine length finish and no full-cohort grade. These are checkpoint counts, not a claim that an unfinished cohort has completed or that a runtime is active. Both completed Cyber Qwen configurations retain 202 suite answers, 12 agent records and 22 judge responses each. Partial/probe outputs are not additional leaderboard observations.

Mothership RTX3090 synthetic benchmark supplement (November 2026)

30 diagnostic observations from 12 hashed receipts, including Flash-Next long-context prefill/retrieval, auto-routing canaries, and 9B Q6 CPU-only smoke results. Scope and caveats; receipt manifest. Not new weighted Mango quality scores. Frozen leaderboard unchanged.

Mothership RTX3090 verified Mango comparison (November 3, 2026)

Download the full frozen-source score-only supplement with 691 original sanitized observations, plus a 23-row Flash IQ4 judge22 addendum (714 total) covering three exact checkpoint/quant variants, per-question numerical grades, core/domain/task breakdowns, single-GPT-6-rater qualitative results for AtomicChat, Heretic9B, and Flash IQ4, 202+12 paired performance, and aggregated long-context gateway telemetry. Methodology and limitations; staging manifest.

The original Flash-Next IQ4 score is 84.66610/100 and marked certified for the pinned native Medium benchmark. AtomicChat V4.1 IQ2_XS 84.66382/100 and Heretic9B Q6 57.78120/100 are clearly labeled diagnostic/leaderboard-ineligible (native-effort parity unverified; additional Q6 EQBench issue). Empty individual IFEval scores were not fabricated from 19/20 and 17/20 aggregate receipts. The optional judge22 scores are separate from the core economics formula. On the identical 22 frozen prompts a single unblinded GPT-6 reviewer scored Flash IQ4 7.86364/10, Atomic V4.1 8.72727/10, and Heretic9B 7.09091/10, with no multi-judge adjudication. Addendum manifest. The archived Quality262K/Fast128K gateway telemetry has different prompt mixes and is not a matched context benchmark.