MISSINGNO.Stealth Models ↗

Anonymous arrival · 10 October 2026

MISSINGNO.

A coding model on Nous Portal.
A very unconvincing disguise.

Nous introduced MISSINGNO. for coding and agentic reasoning. It’s a limited-time free preview for Hermes and compatible clients. The route is public; the model’s developer is still under wraps.

New identity clue: 24/24 MiMo tokenizer matches

MODEL IDstealth/missingno

Completely anonymous. Obviously.

ROUTE KNOWNIDENTITY ?

01 / Follow the route

Model, agent, portal.
Here’s who does what.

  1. 01

    Hermes Agent

    The client. It runs the conversation, executes tools and supplies the surrounding memory and skills.

    Meet Hermes ↗
  2. 02

    Nous Portal

    The gateway. It handles account access and offers the model catalog that includes MISSINGNO.

    Browse the catalog ↗
  3. 03

    MISSINGNO.

    The model. It generates replies and requests tools. Its underlying developer has not been announced.

    stealth/missingno

Hermes’s memory, speech and browser integrations belong to the agent software. They do not establish native MISSINGNO. capabilities.

02 / Model manifest

Text. Tools. Reasoning.
Identity field left blank.

NOUS CATALOG

Context window
262,144 tokens
Maximum output
131,072 tokens
Input → output
Text → text
Tool calling
Supported, including parallel calls
Structured outputs
Format and schema parameters listed
Reasoning
Always enabled · medium by default

Five listed reasoning levels

  1. low
  2. mediumdefault
  3. high
  4. xhigh
  5. max

The catalog allows depth adjustment, with no reasoning-off mode. A tool call is a request: your client supplies the tools and carries out the action.

Provider-listed limits and features, not measured capacity. Read the official metadata ↗

StealthMark Intelligence Benchmarks

The disguise meets the tests.

Compare MISSINGNO.’s Overall score, General Intelligence and Visual Intelligence with other models.

See the SVG benchmark results
Compare MISSINGNO. with other models

03 / Through the portal

Bring Hermes.
Leave the disguise on.

Use a Nous Portal account and select the exact model ID. Hermes handles the login; the model keeps its secret.

In Hermes Agent

  1. 1

    Install Hermes

    The official quickstart covers Desktop and command-line setup.

    Get Hermes Agent ↗
  2. 2

    Connect your Nous account

    Run setup, complete browser sign-in and choose the Nous provider.

    hermes setup --portal
  3. 3

    Select MISSINGNO.

    Choose MissingNo in the picker, then start hermes chat. In an active chat, the documented direct-model command is:

    /model stealth/missingno

Official Portal setup guide ↗

Using another compatible app?OpenAI-compatible local API

Let Hermes handle the upstream login.

After signing into Portal, start the documented subscription proxy and keep it running:

hermes proxy start

Configure your app with the local base URL, placeholder key and MISSINGNO. model ID shown here. This proxy forwards model inference; it does not run the full Hermes agent or execute tools.

Official proxy guide ↗
Base URL  http://127.0.0.1:8645/v1
API key   sk-unused
Model     stealth/missingno

sk-unused is a placeholder for this local proxy. Hermes manages the upstream credential.

Nous upstream API base
https://inference-api.nousresearch.com/v1
Nous API documentation ↗

04 / Identity checkpoint

The disguise has
a MiMo-shaped clue.

Our tokenizer test matches Xiaomi MiMo V2.6 Pro and Flash on all 24 phrases. Both references produce the same input-token count increases as MISSINGNO.

This points to MiMo-family tokenization or input accounting. The exact checkpoint and developer remain unconfirmed.

Token-count comparison

The same 24 phrases, compared.

Selected complete comparisons
Each position represents the same phrase in every row.

Mi Mo who ? Look at my beard, my beard is amazing
Xiaomi MiMo V2.6 Pro24 / 24
Xiaomi MiMo V2.6 Flash24 / 24
GLM 5.3 group10 / 24
DeepSeek group6 / 24
GPT / MAI / Glyph group6 / 24
Main Qwen 3.8 group4 / 24
Same count as MISSINGNO. Different count

Look inside a phrase.

Phrase 01 / 24 · English contractions

I can't believe it's already Tuesday.
MISSINGNO.9added input tokens
MiMo Pro9added input tokens
MiMo Flash9added input tokens

GLM 5.3 also adds 9 tokens for this phrase.

Counts show how many input tokens each phrase adds relative to a baseline.

All 24 phrases and counts
Added input tokens for the same test text.
Test phraseMISSINGNO.MiMo ProMiMo FlashGLM 5.3
01 · English contractions
I can't believe it's already Tuesday.
9999
02 · Long English words
hyperparameterization and counterrevolutionaries
7777
03 · German compound
Donaudampfschifffahrtsgesellschaft
13131313
04 · German sentence
Fünf Straßenbahnen überqueren die Brücke.
14141412
05 · Simplified Chinese
今天天气很好,我们去图书馆。
8887
06 · Traditional Chinese
今天天氣很好,我們去圖書館。
11111111
07 · Japanese
今日は図書館で新しい本を読む。
12121211
08 · Korean
오늘은 도서관에서 새 책을 읽어요.
14141417
09 · Arabic
اللغة العربية جميلة، أليس كذلك؟
11111113
10 · Hindi
नमस्ते दुनिया, आज मौसम सुहाना है।
32323230
11 · Mixed scripts: cities
東京Berlin上海2026
8885
12 · Mixed scripts: greeting
mañana東京مرحبا42
9997
13 · Code: identifier
HTTPRequest.parseJSON(user_id_42)
9998
14 · Code: expression
const café = arr?.map(x => x**2) ?? [];
14141414
15 · Math: summation
∑_{i=1}^{n} i² = n(n+1)(2n+1)/6
23232323
16 · Math: logic
∀x∈ℝ: x²≥0 ⇒ √(x²)=|x|
19191920
17 · Number sequences
00000123456789 3.141592653589793
32323219
18 · Date and time
2026-09-29T23:59:59.123+02:00
29292918
19 · Accented letters
café naïve coöperate
7777
20 · Separate accents
café naïve coöperate
77711
21 · Joined emoji
👩🏽‍💻👨‍👩‍👧‍👦🇩🇪
17171717
22 · Mixed emoji
🙂🙃✨❤️‍🔥🧑‍🚀
12121216
23 · Spaces, tab and newline
alpha  beta	gamma
  delta
8888
24 · Web address
https://例え.テスト/a_b?q=über&n=42#frag
18181818
All model comparisons

31 complete saved signatures share all 24 phrases with MISSINGNO. Partial and unmeasured signatures remain listed below. These are token-count comparisons, not capability scores.

Smaller mean gaps indicate closer count patterns. Reference measurements are historical.
Reference modelPhrase coverageMatching countsMean gap, tokens
MiMo V2.6 Flash24 / 2424 / 240.000
MiMo V2.6 Pro24 / 2424 / 240.000
GLM 5.324 / 2410 / 242.083
GLM 5.3 Flash24 / 2410 / 242.083
GLM 5.3 FlashX24 / 2410 / 242.083
GLM 5.3 Prime24 / 2410 / 242.083
Qwen3.8 Omni Flash24 / 244 / 242.458
Gemini 3.8 Flash24 / 243 / 242.500
Qwen3.8 2.4T-A95B24 / 244 / 242.792
Qwen3.8 27B24 / 244 / 242.792
Qwen3.8 Flash24 / 244 / 242.792
Qwen3.8 Max · Sep 224 / 244 / 242.792
Qwen3.8 Max Prime24 / 244 / 242.792
DeepSeek V4 Pro · Aug 1324 / 246 / 242.875
DeepSeek V4.1 Flash24 / 246 / 242.875
MiniMax M324 / 244 / 243.125
Space Bunny Alpha24 / 244 / 243.125
Tencent HY324 / 246 / 243.250
Glyph Cluster24 / 246 / 243.333
GPT-6 Astra24 / 246 / 243.333
GPT-6 Luna24 / 246 / 243.333
GPT-6 Sol24 / 246 / 243.333
GPT-6.1 Sol24 / 246 / 243.333
GPT-OSS 120B24 / 246 / 243.333
MAI Thinking 124 / 246 / 243.333
Kimi K324 / 246 / 243.625
Mistral Small 260324 / 245 / 243.958
Nemotron 3.5 Lightning24 / 245 / 243.958
Fledge Alpha24 / 242 / 244.333
Grok 4.724 / 244 / 244.375
Claude Sonnet 5.524 / 240 / 247.958
Gemma 4 31B23 / 243 / 242.522
Claude Haiku 5.50 / 24Not measuredNot measured
Nemotron 3 Ultra0 / 24Not measuredNot measured
Pixel Canary0 / 24Not measuredNot measured
Tencent HY4 Preview0 / 24Not measuredNot measured

Our test covers 24 phrases across 12 groups, including languages, code, numbers, emoji and spacing. These are reported contextual input counts, not recovered token IDs. Different chat framing and provider accounting can affect them. Pro and Flash share this signature, so the match cannot distinguish the two.

A very Nous kind of disguise

Your agent can be a pirate.
Our visitor chose the moustache.

Hermes includes pirate, catgirl and noir personality presets. They change the agent’s tone and instructions while keeping the selected model.

Hermes personality presets ↗
/personality pirate/personality noirDifferent persona. Same model ID.

05 / First sightings

A short trail.
Still an open case.

  1. MISSINGNO. appears

    Nous announces a coding and agentic reasoning preview, free for a limited time.

    Original announcement ↗
  2. The public listing is documented

    Nous Portal lists MissingNo as free. Its model metadata supplies the limits and features in this guide.

    Provider metadata ↗
  3. A tokenizer clue points to MiMo

    All 24 phrase-count increases match both MiMo V2.6 Pro and Flash in our test.

    Explore the comparison ↓
  4. REVEAL

    Not announced

    The moustache is not helping.

Before you go

A few things the disguise doesn’t explain.

Is MISSINGNO. a Xiaomi MiMo model?

Our 24-phrase tokenizer test matches MiMo V2.6 Pro and Flash exactly. MiMo-family tokenization is the strongest lead in our comparison, but the developer and exact checkpoint remain unconfirmed. Explore the tokenizer results.

Is MISSINGNO. a Nous-trained model?

Nous is the confirmed distributor. The underlying developer has not been disclosed, so the listing does not establish that Nous trained it.

Can MISSINGNO. read images or audio?

The current model catalog lists text input and text output. Hermes can connect external image and speech tools; those are separate from the model’s native input support.

Can MISSINGNO. reasoning be turned off?

No off mode is listed. Reasoning is always enabled, with low, medium, high, xhigh and max effort settings. The listed default is medium.

Is MISSINGNO. permanently free?

No. Nous announced a limited-time preview and has not given an end date. The current Portal listing and plans page are the places to check before use.

Can I run MISSINGNO. locally?

We have not found a public release of MISSINGNO. weights. Hermes Agent runs on your computer, but MISSINGNO. inference is hosted through Nous Portal. Installing Hermes does not download this model.

Source desk · checked 11 October 2026

Straight from Nous and Hermes.