Six models at ¥0, none of them free: licence, hardware and delivery in the real unit price

¥0 is the absence of a price tag, not the absence of cost

Six models on this board have an output price of ¥0: WITA-Omni, Emu3.5, Zidong Taichu 2.0, Hikvision Guanlan 3, NeuroXess-N1 and BAGEL-2.

Read only that column and they are the cheapest six. But “¥0” here means: no public per-token price. It may mean weights are downloadable, an industry licence is quoted per project, there is no API at all, or the vendor has not decided how to sell it.

This essay does one procurement job: unpack licence, delivery and hidden cost for all six, and answer when zero price is rational and when it is a trap.

The headline: none of the six is open source in the OSI sense; none is “download and commercialise”; none has a total cost of zero.

The six ¥0 models

Model	Overall	Licence	Modalities	Evidence	Entry	What ¥0 really means
WITA-Omni	#23	Closed, filed commercial	T/I/V/A/embodied	A	AgiBot project partnership	No public token price; ships with robot/project
Emu3.5	#24	Open source (research licence)	T/I/V	A	Weight download + self-hosted inference	Weights free; commercial use needs a separate licence
Zidong Taichu 2.0	#25	Open weights (research licence)	T/I/V/A	B	Research licence + self-host	Weights open for research
Hikvision Guanlan 3	#27	Closed · industry licence	T/I/V	C	Industry authorisation (per lane/project)	No API list price; ships with hardware/solution
NeuroXess-N1	#30	Closed · project licence	T/I/embodied	C	Project partnership	No API; ships with device/project
BAGEL-2	#31	Open source (research licence)	T/I	B	Weight download + self-host	Weights free; commercial use needs a separate licence

Three completely different kinds of “zero”:

One: research-licensed open weights (Emu3.5, Zidong Taichu 2.0, BAGEL-2). Weights are downloadable, but the licence says research use. Commercial use needs a separate grant, or is limited to internal research. You save the token fee and pay in legal review, GPUs and inference engineering.

Two: industry authorisation (Guanlan 3). It is not an API product; it is part of a surveillance solution. Price is per camera lane, platform licence and project delivery; token price does not apply.

Three: project-based or unpriced (WITA-Omni, NeuroXess-N1). One is an embodied brain from a robotics company, the other a neural-signal decoder; neither has a public API. ¥0 is not “cheap”; it is “not sold in this unit”.

Six at ¥0, zero OSI open source

The most overlooked fact: none of the six uses an OSI-defined open-source licence.

• Emu3.5, Zidong Taichu 2.0 and BAGEL-2 are open weights under a research licence;
• WITA, Guanlan and NeuroXess are closed with project or industry authorisation.

OSI open source (Apache-2.0, MIT) lives elsewhere on the board: Qwen3-Omni-30B (Apache-2.0), Qwen3.8-Flash (Apache-2.0), DeepSeek-V4.1-Flash (MIT), InternVL 4.0 (MIT), MiniCPM-V 4.5 (Apache-2.0, commercial registration required). They have explicit token prices, but the licences allow commercial use, modification and redistribution.

“Free” means two completely different things: a price of zero, and legal freedom. The first is a commercial strategy; the second is a right. Mixing them is the biggest trap in zero-price procurement.

A counterintuitive conclusion: if what you want is “free commercial use”, a ¥0 research-licence model can be more expensive than a ¥0.6-per-million-token MIT model. The former needs legal review, GPUs, an inference stack and self-managed operations; the latter needs an invoice. A price tag is not total cost.

The hidden bill behind zero price

Split total cost of ownership (TCO) into five lines; none of the six escapes them:

Cost line	Research licence (Emu/Zidong/BAGEL)	Industry licence (Guanlan)	Project licence (WITA/NeuroXess)
Licence	Commercial use needs a separate grant; research boundary	Per lane/project/platform	Contract-defined; usually non-transferable
Compute	Self-built GPU; MoE inference stack	Edge box/server	Ships with robot/device
Integration	Write your own serving, quantisation, concurrency	Integrate with surveillance platform	Integrate with robot/neural device
Compliance	Data residency, model filing, content safety	Industry qualifications, security grading	Medical/research ethics where applicable
Exit	Weights stay, maintenance depends on community	Change vendor = change hardware/platform	Change vendor = change device and recollect data

Of the five, only the licence is explicit; the other four are hidden. That is why the six sit below #23 overall yet cannot be procured off rank alone: their cost structure is completely different from the 26 priced models.

Zero price is not weak: each has a top-ten long dimension

The six average 78.9, 7.6 points below the 26 priced models at 86.5. That average hides three things:

Emu3.5 has the board’s best generation at 98. It ranks #24 overall, but for text-to-image or a world model it is the first choice; free weights allow on-prem deployment, at the cost of agent 70 and document 82.

Hikvision Guanlan 3 has video 91, seventh on the board. It ranks #27 overall with grade-C evidence, but in surveillance it is the only entrant; a ¥0 token price means the cost sits in hardware and industry authorisation.

WITA-Omni has video 92, sixth on the board. Its embodied-native capability is rare; rank #23 does not reflect embodied ability, because the scorecard has no embodied dimension.

Zero-price models are not “bad models”; they are a different product. They sell weights, hardware, industry solutions or research partnerships, not tokens. Comparing them on token price is using a thermometer to measure length.

When zero price is rational

Four scenarios:

1. Research and benchmarking. You need to reproduce a paper, run ablations, test your own evaluation set. A research licence is enough; weights are modifiable; cost is controllable.

2. Data sovereignty and on-prem. Data cannot leave the intranet; API is unavailable. Open weights are the only path; “zero price” is a side effect, not the reason.

3. Hardware or industry bundle. You are buying cameras, robots or neural devices; the model is part of the solution. Guanlan, WITA and NeuroXess belong here; comparing token prices is meaningless.

4. Single purpose, self-hosted. You only need document OCR or image generation, and your team can run inference. Emu3.5 and BAGEL-2 can be low-cost components — but include GPUs, operations and upgrades.

When zero price is a trap

Five signals; any one means recalculate:

• “Free” is treated as “no cost”. GPUs, operations, legal and integration are missing.
• Licence boundaries are unclear. Does the research licence allow internal commercial use? SaaS? Distillation? If the contract is silent, it is risk.
• No SLA or API. Project delivery means downtime is yours; production cannot rely on “community support”.
• Benchmarks cannot be reproduced. No public evaluation, no third-party reproduction; you must test, and testing may cost more than the model difference.
• Exit cost is high. Industry authorisation binds hardware/platform; project licensing binds devices; data and interfaces do not migrate.

The real risk of zero price is not “bad performance”; it is “uncountable”. An uncountable cost is infinite in procurement.

Procurement question list

When you see ¥0, ask seven questions:

1. Is the licence research, commercial, or evaluation-only? Can it be used in SaaS or customer-facing services?
2. Are weights downloadable? Can they be modified, distilled, redistributed?
3. Is there a public API? If not, who runs inference?
4. What is the compute cost? Total/active parameters, VRAM, throughput — if undisclosed, can you run a POC load test?
5. What does the industry licence include? Hardware, platform, implementation, maintenance — how much each?
6. What are the SLA and support? Who answers, how fast?
7. What is the exit path? When you change models or vendors, can data, interfaces and weights leave?

Only questions 1 and 2 can be answered from the model table; the other five require the vendor. That is the truth of zero-price procurement: the price is zero and so is the information; you fill it in with questions.

A comparison: free commercial use vs zero price

Path	Examples	Price	Licence	Fits
Zero price + research licence	Emu3.5, BAGEL-2, Zidong Taichu 2.0	¥0 token	Research use; commercial negotiated	Research, on-prem, controlled experiments
Zero price + industry/project licence	Guanlan 3, WITA, NeuroXess	¥0 token	Closed, project-based	Surveillance, robotics, neural devices
Low price + OSI open source	Qwen3.8-Flash (Apache-2.0, ¥1.2), DeepSeek-V4.1-Flash (MIT, ¥0.6), InternVL 4.0 (MIT, ¥1.0)	¥0.6–1.2	Commercial use, modification	Production systems, secondary development
Low price + closed API	MiniCPM-V 4.5 (¥0.5) and others	From ¥0.5	Commercial API	Fast launch, no operations

If your goal is “free commercial use”, row three is usually cheaper than row one. That is the difference between a price tag and total cost.

Boundaries

• ¥0 comes from priceOut: 0; it does not mean the vendor charges nothing. It means no public per-token price; project, industry and hardware bundles may all charge.
• Licence descriptions come from the license field in the model dossier; they are editorial summaries, not legal advice. Read the original licence and contract before commercial use.
• The six average lower, but zero price does not cause lower scores. They cluster in long-tail, specialist and research scenarios; six is too small for causal inference.
• The five TCO lines are a framework, not a quote. Amounts depend on your compute, team and compliance requirements.

Reproduction

The six are filtered by priceOut === 0 in src/data/models.ts; licence, modalities and evidence come from the same entry; averages recompute with the site weights. Scripts ship with the data; rerun after every expansion, and the zero-price list updates in the changelog.

If this essay leaves one sentence: ¥0 is not a price; it is the absence of a pricing method. When you see it, first ask how it is actually sold, then ask which invisible bills you will pay.