Only five of 32 models disclose a parameter count: an opacity index, and what the ranking looks like when you discount what vendors hide

A dimension nobody ranks: how much you were told

A tier list ranks capability. But when you procure a model, you need a second thing: how much the vendor told you, and who bears the cost of what they did not.

This is not a moral word about “transparency”; it is a computable procurement cost. Parameter count determines VRAM and throughput estimates. Context determines whether a long document fits in one call. Price determines budget. Licence determines commercial use. Evidence grade determines whether you must re-test. Every missing field moves cost from the vendor to you.

We checked four fields across all 32 models: parameter count, context, price and evidence grade. The headline: only five models disclose a parameter count; 24 give an adjective (“MoE”, “native”, “embodied”); three give nothing at all. Context disclosure is 96.9%, price disclosure 81.3% — parameter count is the least transparent column on the board.

Disclosure rates for four fields

Field	Disclosed	Total	Rate	Note
Parameter count	5	32	15.6%	Only total/active parameters count
Context	31	32	96.9%	The only exception is WITA-Omni
Output price	26	32	81.3%	Six at ¥0; see zero-price models
Evidence grade	32	32	100%	14 A / 16 B / 2 C, but graded by us
Licence	32	32	100%	“Open source”, “open weights” and “research licence” differ

The parameter bar is strict: total and active parameters, or an explicit size. “MoE”, “trillion-scale” or “about 8B active” do not count — they help you guess, not calculate VRAM.

By that bar, the 32 split into three groups:

• Five give numbers: Qwen3.5-397B (397B total / 17B active), Kimi K3 (2.8T total), Qwen3-Omni-30B (30B-A3B), Keye-VL-2.0 (30B-A3B), MiniCPM-V 4.5 (8B).
• Twenty-four give adjectives: from “undisclosed (MoE)” to “undisclosed (MoE, ~12B active)”. The latter adds a hint, still not enough to recompute.
• Three are silent: SenseNova 6.5 Pro, ERNIE 5.1 and Hunyuan-Vision have nothing but “undisclosed”.

A notable coincidence: all five models that give numbers are evidence grade A. Not because disclosure causes grade A, but the reverse — teams willing to disclose parameters also tend to publish technical reports, reproduction details and failure cases. Transparency is an organisational habit, not a field.

The five models that disclose a parameter count

Model	Parameters	Overall	Evidence	Price / 1M output tokens
Qwen3.5-397B-A17B	397B total / 17B active	#3	A	¥12
Kimi K3	2.8T total (MoE, open)	#5	A	¥16
Qwen3-Omni-30B	30B-A3B	#10	A	¥4
Keye-VL-2.0	30B-A3B	#14	A	¥2
MiniCPM-V 4.5	8B	#26	A	¥0.5

These five span four orders of magnitude, from 8B to 2.8T, but rank from #3 to #26. Parameter count has no monotonic relationship with rank: Kimi’s 2.8T does not beat Qwen3.5’s 397B total, and MiniCPM’s 8B is not last. In MoE, total and active parameters are different things; reading total alone can mislead.

What is useful is the four numbers together: total, active, context, price. Qwen3-Omni reaches 88.9 at 30B-A3B and ¥4; Kimi reaches 92.3 at 2.8T and ¥16. The former has an order-of-magnitude lower cost per point; the latter has stronger long context and agents. Parameter count matters not for rank, but for explaining price and deployment limits.

Evidence grade is not a score; it is an error bar

The board’s evidence distribution is 14 A, 16 B, 2 C. Averages by group:

Evidence	Count	Average score	Average rank	Rank range
A	14	87.4	12.6	#2–#26
B	16	83.9	18.4	#1–#32
C	2	77.7	28.5	#27, #30

Group A averages higher, but group B contains both #1 (Qwen3.8-Max, 95.4) and #4 (GLM-5.3, 92.4). Not a contradiction — an error bar. Grade B means vendor-reported benchmarks plus limited third-party reproduction, with a reasonable uncertainty of about ±1.5 points; grade C is wider.

So a fact must be stated plainly: raw #1 Qwen3.8-Max (95.4, B) leads #2 Doubao (95.2, A) by 0.2 points, far inside B’s ±1.5-point error. Our method treats anything inside 0.5 as one band, so strictly this is a tie for first. But if you must choose the “more credible first”, Doubao’s evidence grade is higher.

That is the motive for folding uncertainty into ranking: not deducting “capability points” from grade B, but adding a certainty discount to its score.

Confidence-adjusted ranking: folding the unknown in

We define a transparent discount formula, used only for procurement ordering, not to change raw scores:

• Evidence discount: A = 0, B = 1.5, C = 3.5 (matching our error bars).
• Parameter opacity: +0.5 (cannot estimate VRAM/throughput).
• Price opacity: +0.5 (¥0 with no public quote).
• Context opacity: +1.0 (only WITA triggers this).

Adjusted score = raw score − evidence discount − opacity penalties. Result:

Adjusted rank	Model	Raw	Raw rank	Discount	Adjusted
1	Doubao Seed 2.0 Pro	95.2	#2	−0.5	94.7
2	Qwen3.5-397B	94.6	#3	0	94.6
3	Qwen3.8-Max	95.4	#1	−2.0	93.4
4	Kimi K3	92.3	#5	0	92.3
5	GLM-5	92.0	#6	−0.5	91.5
6	GLM-5.3	92.4	#4	−2.0	90.4
7	SenseNova 6.5 Pro	89.9	#8	−0.5	89.4
8	Qwen3-Omni-30B	88.9	#10	0	88.9
9	LongCat-Next	89.2	#9	−0.5	88.7
10	MiMo-V2.5	90.1	#7	−2.0	88.1

Three changes matter:

Qwen3.8-Max falls from #1 to #3. Not because it is weak, but because its 95.4 rests on grade-B evidence and an undisclosed parameter count. Its capability ceiling is still the highest here; its verifiable floor is below Doubao and Qwen3.5.

Doubao rises from #2 to #1. Its raw score is only 0.2 lower, but grade-A evidence, a public price and a public context make the score more dependable. This is not “Doubao is stronger”; it is “if you can only afford one procurement decision, Doubao’s score is less likely to be overturned by later reproduction”.

GLM-5.3 falls from #4 to #6. Its 92.4 is also grade B with undisclosed parameters. GLM-5 scores 0.4 lower and costs ¥1 more (¥10 vs ¥9), but grade-A evidence lets it overtake after adjustment. When the evidence gap is real, a ¥1 price difference is not the deciding term.

The formula can be changed. Lower the B discount from 1.5 to 1.0 and Qwen3.8-Max adjusts to 93.9, still #3; raise it to 2.5 and it falls to 92.4, close to Kimi’s 92.3 but still behind. The top two (Doubao, Qwen3.5) are stable across the entire 1.0–2.5 range — the conclusion is not tuned into existence.

Why parameter count matters, and why it is not capability

Parameter count affects four things, all procurement costs:

1. VRAM and throughput: for on-prem, total parameters set the minimum memory; active parameters set compute per token. “MoE” alone cannot be estimated.
2. Fine-tuning cost: the boundary between full fine-tuning and LoRA depends on total size.
3. Latency and concurrency: the same active size with different routing yields different tail latency.
4. Reproducibility: with a public parameter count, third parties can recompute benchmarks; without it, benchmarks rest on trust.

But parameter count is not capability. Kimi’s 2.8T total does not beat Qwen3.5’s 397B; MiniCPM’s 8B still scores 86 on documents. Parameter count explains price, latency and deployment threshold, not score. Treating it as a ranking is another form of benchmark worship.

Transparency and rank: close to zero

The five models with numbers rank 3, 5, 10, 14 and 26; the three fully silent models rank 8, 11 and 15. You can reach the top 15 without disclosing parameters, and disclose parameters while ranking #26. Transparency and capability are two axes.

But transparency is strongly related to procurement risk:

• No parameters → no capacity planning; you buy API access or run black-box load tests.
• No price → no budget; you enter project-based negotiation.
• No context → no way to know whether a long document fits in one call.
• Low evidence → you must re-test, and the test cost may exceed the model price difference.

Transparency does not change capability; it changes the second bill you pay for capability. That bill is not in the token price. It is in evaluation, integration, compliance and exit.

Procurement question list

When you see a board with only a composite, ask six questions:

1. What are the total and active parameter counts? “MoE” is not an answer.
2. Is context native or extrapolated? What is the effective context?
3. Is the price list or tiered? How are image tokens counted?
4. Does the licence allow commercial use? On-prem? Distillation?
5. Is the evidence vendor-reported, third-party reproduced, or your own scenario?
6. If this version is retired next year, what is the replacement path?

The first three decide “can I use it”; the last three decide “can I keep using it”. The composite only answers “is it good”.

Boundaries

• The discount is a heuristic, not a statistical model. 1.5/3.5 come from our error bars; 0.5/1.0 are rough proxies for procurement cost. Change them to match your risk appetite.
• Evidence grades are ours. A/B/C are not vendor certifications; the criteria are in methodology. Another institution may grade differently.
• “Undisclosed” is not necessarily concealment. Some teams withhold parameters for security, compliance or competition. That is a legitimate choice, but the cost still lands on the buyer.
• Five is a tiny sample. Do not conclude “models that disclose parameters are stronger”; only “models that disclose parameters are easier to verify”.

Reproduction

Fields are checked against params, context, priceOut and evidence in src/data/models.ts; rates use all 32 models. The discount formula, adjusted table and sensitivity range ship with the site data; change one parameter and rerun. Numbers update in the changelog after each expansion.

If this essay leaves one sentence: the rank tells you what it can do; transparency tells you how much to trust that claim. Reading only the first layer and ordering is leaving the second layer’s cost to yourself.