Breaking

IFM Releases Six-Model K2 Horizon Family From 0.9B to 375B

September 3, 2026 at 10:09 EDT

Foundation Models
AI Agents
Open Source

The Institute of Foundation Models (IFM) has released K2 Horizon, a six-model family ranging from 0.9B to 375B parameters. The lineup is designed to cover deployment from constrained edge devices to enterprise reasoning, while releasing the weights, code, training data and methodologies behind the models. Launch details

SEP 3, 2026 | IFM

One family, six deployment tiers

K2 Horizon spans 0.9B to 375B parameters, from constrained devices to enterprise workloads.

At a glance

models in the family

0.9B-375B

parameter range

23B

flagship active per token

~3X

parallel-token speedup

Stored capacity vs active compute

The sparse flagship stores 375B parameters but activates 23B for each token.

375B stored

Flagship total

23B active

Per token

Same scale: 375B stored versus 23B active per token

One fleet, one deployment path

PROTOTYPE

Start small

Use a compact model for local work

→

ROUTE

Match task to cost

Direct work to the right model tier

→

SCALE

Keep the workflow

Move to larger deployments as needed

What changes on day one

MORE TO INSPECT

Transparent by design

Weights, code, training data, methodology and intermediate checkpoints are part of the release.

MORE TO VALIDATE

Practical runtime questions

Early testing raised requests for smaller quantized builds and noted that llama.cpp support depended on an upstream change.

Continue reading

The rest of this article is for AI News Blitz readers. Choose an option below to keep reading.