N
NEX AGI / nex-agi/nex-n2.6

Nex-N2.6

Nex AGI multimodal model based on GLM-5.3-Flash, with about 321 billion parameters and downloadable FP8 weights. It targets coding, computer use and long tool-based tasks. Public weights and OpenRouter access date to November 10, 2026. It accepts text and images and returns text and function calls.

01 / SUMMARY

The essential

ORIENTATIONMultimodal agent with FP8 weights

Primary use stated or inferred cautiously from official documentation.

CONTEXT262,144 tokens on OpenRouter; weight configuration with 1,048,576 positions

Maximum input capacity when published by the source.

EXITThe endpoint declares 235,929 tokens; the limit depends on the provider

Documented maximum generation limit.

ACCESSOpenRouter · Hugging Face

Verified availability channels.

MODEL TYPEMultimodal · Code · Text and reasoning

Primary functional family and verified specialties.

LICENSEApache 2.0

Declared terms for API access, model weights, or self-hosting.

02 / TECHNICAL SHEET

Limits and integration

API IDnex-agi/nex-n2.6
Model typeMultimodal · Code · Text and reasoning
Access modelAPI and open weights
LicenseApache 2.0
DeploymentAPI or self-hosted deployment
Release10 NOV 2026
Knowledge cutoffNot published
EntranceText · Image
ExitText
Context window262,144 tokens on OpenRouter; weight configuration with 1,048,576 positions
maximum outputThe endpoint declares 235,929 tokens; the limit depends on the provider
ReasoningNot published
Published toolsFeatures
Structured OutingsNot published
Batch processingNot published
Prompt cacheNot published
Fine-tuningNot published
Verified platformsOpenRouter · Hugging Face · SGLang
BEST SUITED FOR

Coding and visual automation agents that alternate planning, tools and result checks. The template provides low, high and max reasoning effort, defaulting to max. The client executes functions and returns results; calls do not by themselves grant system access. Self-hosting requires a GLM-5.3-Flash and FP8-compatible runtime, preserving the checkpoint’s tokenizer, processor and chat template.

WORTH MONITORING

The configuration declares 1,048,576 positions, but OpenRouter limits context to 262,144 tokens. Its endpoint API gives a 235,929-token maximum output while the commercial page says 262,144: apply the selected endpoint’s limit and remaining context, not the highest figure. FP8 is a format of the same model, not another version. Autonomous computer use requires bounded permissions, isolation and checks on irreversible actions.

03 / CAPABILITIES

What it can do

01

Orientation

Coding and visual automation agents that alternate planning, tools and result checks. The template provides low, high and max reasoning effort, defaulting to max. The client executes functions and returns results; calls do not by themselves grant system access. Self-hosting requires a GLM-5.3-Flash and FP8-compatible runtime, preserving the checkpoint’s tokenizer, processor and chat template.

02

Context

262,144 tokens on OpenRouter; weight configuration with 1,048,576 positions · The endpoint declares 235,929 tokens; the limit depends on the provider

03

Tools and integration

Features

04

Access

OpenRouter · Hugging Face

Input modalities
TextImage
04 / PRICES

Documented cost

ConceptWorthUnit/condition
Standard input0.12 USD / 1M tokensStandard API rate
Cached input0.024 USD / 1M tokensReading reused prefixes
Cache write or storageNot publishedThe condition varies by provider
Standard outputUSD 0.40 / 1M tokensMay include reasoning tokens
Batch inputNot publishedAsynchronous processing
Batch outputNot publishedAsynchronous processing

Nex AGI endpoint prices on OpenRouter checked on November 22: USD 0.12 input, USD 0.024 cached input and USD 0.40 output per million tokens. These are that service’s prices, not self-hosting or universal distributor prices. The weights’ Apache 2.0 license does not remove infrastructure costs. Separate batch rates have not been verified.

i Prices change and may depend on level, region or context length. Check the source before making a decision.

05 / EVALUATIONS

How to read the results

“

Nex AGI reports 83.0 on OSWorld, 84.1 on Terminal-Bench 2.1 and 45.7 on AutomationBench. These are model-level results, not a separate evaluation of this FP8 export or Inferama measurements. Reasoning repetition falls from 16.0% to 1.6% in an internal issue-replay test; 100% valid tool selection is not a universal success guarantee either. Comparisons require the same harness, tools and budget.

BenchmarkResultMetricSource
OSWorld83,0Published scoreView source ↗
Terminal-Bench 2.184,1Published scoreView source ↗
AutomationBench45,7Published scoreView source ↗

Inferama only highlights a “best result” when the metric, test set, configuration, and date allow for an equivalent comparison. The supplier's figures are presented as claims from its own source.

06 / VERSIONS

Chronology

nex-agi/nex-n2.6

Version added to Inferama's verified catalog.

07 / SOURCES

Traceability