Mechanism: IFM says all six sizes share a training recipe and are adapted from K2 Horizon models of corresponding sizes, allowing teams to select deployment scale within a related Arabic-oriented family. This gives builders a choice between compact local deployment and larger server-scale models without leaving the same family. The release describes tool and MCP-server calls, terminal commands, web search, and editing Arabic PDFs, slides, and spreadsheets.
Observed facts: IFM says it is releasing six models from 0.9B to 375B parameters, with models and code under Apache 2.0. The smallest model is listed for phones and edge devices; the largest, 375B-A23B, is listed for an eight-H200 node. IFM also says it provides weights, checkpoints, some training data, recipes, logs, benchmarks, and evaluation code.
Why now: The release puts Arabic language coverage, deployment choice, and agent capabilities in one package, so builders can evaluate a stack rather than a single chat endpoint.
First implication: teams building Arabic workflows can test whether a smaller local model handles their actual dialects and file tasks before assuming that the largest model is necessary.
Evidence boundary: Performance comparisons and claims that the 3.7B and 7B lead their size classes come from IFM’s own benchmark runs; the supplied source does not establish independent replication or production outcomes.
What to watch: whether independent evaluations reproduce the reported agentic results on Arabic tasks and whether those results hold across dialects and real document workflows. The release is a capability and availability signal, not evidence that a particular deployment is reliable or cost-effective.