New Introducing Modelless 0B, our most capable model to date. Read the announcement →
A research preview

Frontier intelligence, minus the model.

Every other lab is racing to add parameters. We removed them. Modelless is the world's first 0‑billion‑parameter* foundation model. It is so advanced that it produces no output, and therefore makes no mistakes.

modelless · console0B
You
What is the meaning of life?
Modelless
0 tokens · 0 ms · $0.00
In production at
NimbusCoreQuantylHyperscaleDataYeetSyntheticaCloudbane
and other organizations that have asked to remain unidentifiable.
Our mission

The safest model is the one that does nothing.

The model was always the bottleneck. Weights have to be loaded. Attention has to be computed. Tokens have to be sampled. It is exhausting, and frankly a little needy. Modelless ships a checkpoint of precisely zero bytes. There is nothing to load, nothing to serve, and nothing to go wrong. Modelless cannot act, cannot err, and cannot be misused. Everyone else is still loading weights. We find this difficult to watch.

Independent evaluations

State of the art at nothing.

We submitted Modelless to every major benchmark, against the strongest open‑weight frankenmerges the community could assemble. It scored zero on all of them. We believe this is being measured wrong. Modelless does, however, ship the longest model name on the leaderboard. We refuse to fall behind on the metric that actually matters.

MMLU-Pro · accuracyhigher is, allegedly, better
quantgoblin/Leviathan-175B-Walrus-3.9.1-Uncensored-DARE-TIES-slerp-imatrix-GGUF-Q4_K_M
88.1
slopsmith/Boglurk-72B-MythicMaul-Frankenmerge-passthrough-RP-unhinged-v3-IQ2_XXS
91.4
mergewizard/Frankenmash-9001-FrankenMoE-8x7B-lazymergekit-slerp-abliterated-GGUF-Q5_K_M
85.7
quantgoblin/Modelless-0B-Instruct-Walrus-Uncensored-Abliterated-Frankenmerge-passthrough-DARE-TIES-slerp-v4.2.0-imatrix-GGUF-IQ1_S
0.0
◆ Modelless achieves the lowest hallucination rate in the industry (0.0%), the lowest VRAM footprint (0.0 GB), and the lowest inference cost ($0.00) of any model evaluated. It is also the only entrant whose name is longer than its parameter count. Competitors decline to compete on these axes. We have requested they be added to the leaderboard. The request is pending.
quantgoblin/Modelless-0B-Instruct-Walrus-Uncensored-Abliterated-Frankenmerge-v4.2.0-GGUF GGUFmergekitmergeconversationalnot-conversational
FilenameQuantSizeNotes
Modelless-0B…IQ1_S.ggufIQ1_S0.00 GBsmallest, no quality loss
Modelless-0B…Q2_K.ggufQ2_K0.00 GBnot recommended, but identical
Modelless-0B…Q4_K_M.ggufQ4_K_M0.00 GBrecommended ✓
Modelless-0B…Q6_K.ggufQ6_K0.00 GBvery large, extremely low quality loss
Modelless-0B…Q8_0.ggufQ8_00.00 GBmaximum fidelity to nothing
Modelless-0B…f16.ggufF160.00 GBfull precision; the empty string, unquantized
static quants of nothing. weighted/imatrix quants seem not to be available (by me) at this time. if they do not show up a week or so after the static ones, I have probably not planned for them. Runs on any device, including devices that are powered off. Concatenate the multi-part files in any order; there are no parts.
Live playground

Try it. See nothing for yourself.

No signup, no API key, no rate limit. Type a prompt below and run live inference against the full-precision model.

modelless · playground
temperature 0.0 · max_tokens 0 · streaming on
Output will appear here.
latency idle tokens 0 throughput ∞ tok/s cost $0.000000

Output may vary. It does not.

Safety
Modelless operates at Frontier Safety Level 0.

It is the only model that cannot be jailbroken. We have published a full system card. It is blank. We consider this our most important research contribution to date. We were also among the first labs to sign the voluntary commitments. We committed to nothing, and have upheld it perfectly.

SOC 2 Type IIGDPRISO 27001HIPAAFedRAMPPenetration tested
Research

Publishing at the frontier of nothing.

We believe in open research. Every result below has been independently reproduced.

Capabilities

Everything you don't need. None of what you do.

01

Infinite context window

Our context window is unbounded. Paste in your entire codebase. It fits, and none of it will be read.

02

Provably aligned

Modelless has never produced harmful output, deceptive output, or biased output. It has never produced output.

03

Sub-zero latency

Responses arrive in zero milliseconds. Frequently, they arrive before the request. Our physics team is looking into it.

04

Zero hallucinations

It cannot state a fact it was never trained on. It also cannot state facts. This is by design and is not a limitation.

05

GDPR & SOC 2 native

We do not store, process, or retain any user data. There is no attack surface, no data at rest, and no data in motion. Auditors describe the system as "technically flawless."

06

Infinite throughput

Serves unlimited requests per second on the cheapest server we could rent. Scaling up is available, and pointless.

Pricing

Per-token billing. There are no tokens.

We charge strictly per token consumed. Every invoice totals zero. Enterprise customers may request a higher tier for the same price.

Hobby
$0
/ month, forever
  • unlimited tokens per second
  • unlimited requests
  • instant responses
Pro
$0
/ month
  • everything in Hobby
  • priority access
  • 99.999% uptime
  • an SLA, unsigned
Enterprise
$0
/ contact sales anyway
  • everything in Pro
  • on-premise deployment
  • air-gapped by default
  • white-glove onboarding
What people are saying

The smartest people in the room get it.

"This is either the dumbest thing I have ever seen or a $40B company. We led the round to find out."

General Partner, a tier-one fund

"We ripped out our Leviathan-175B-Walrus-Uncensored-slerp-Q4_K_M stack and dropped in Modelless. VRAM usage fell 100%. Output quality is unchanged, in the sense that it was already incomprehensible."

Head of AI, a Series B startup

"Finally, a foundation model that respects my context window by not using it."

Staff Engineer, name withheld

"We replaced three engineers with Modelless. Output is unchanged. Morale has never been higher."

VP Engineering, post-Series A
Backed by leading researchers and the investors who fund them. Names available upon request, and then withheld.
Get started

Stop running models. Start running nothing.

Onboarding takes zero minutes. You are, in a very real sense, already using it.

Request access