What was actually released

On September 3, 2026, the Institute of Foundation Models (IFM) introduced K2 Horizon, a family of six language models: dense 0.9B, 3.7B, 7B, and 32B variants, plus sparse 36B-A4B and 375B-A23B mixture-of-experts models. The A4B label means that roughly four billion parameters are activated for each token, while the model stores about 36 billion parameters in total.

This is not a placeholder announcement. The official repositories contain configurations, tokenizers, and sharded Safetensors weights. At the time of verification on September 4, the BF16 weights occupied approximately 2.2 GB for 0.9B, 10.1 GB for 3.7B, 18.0 GB for 7B, 69.6 GB for 32B, and 74.9 GB for MoVA 36B-A4B. IFM has also published official GGUF repositories for several sizes, although these currently contain BF16 files rather than a ready-made range of compact quantizations.

The models and code are labeled Apache 2.0. Their cards describe text generation, reasoning, and tool calling. The 0.9B card advertises up to 131,072 tokens of context; the 3.7B, 7B, 32B, and 36B-A4B cards advertise up to 524,288 tokens. Yet the serving examples for 7B, 32B, and 36B cap the server at 131,072 tokens. An architecture's supported window and an economically sensible production window are not the same thing.

Why “4B active” does not mean “fits like a 4B model”

A sparse model does not compute every stored parameter for every token, so active-parameter count is useful when estimating part of the compute load. However, all experts still have to be stored and loaded. The official 36B-A4B BF16 weight set is about 75 GB before KV cache, inference-engine overhead, and headroom for concurrent requests.

The model card gives a validated SGLang recipe using two NVIDIA H200 GPUs with tensor and expert parallelism. That does not prove every other configuration is impossible, but it shows the hardware tier on which the publisher validated the documented path. A small business should not translate “4B active” into a budget for one ordinary workstation GPU.

The dense 7B weights are about 18 GB in BF16. That is closer to a local server or workstation, but context changes the calculation quickly: KV cache grows with input length, concurrency, layer count, and data type. The advertised 512K window should not be included in a pilot by default. For internal-document search, solid retrieval, short evidence-bearing passages, and a bounded context are often cheaper than sending an entire archive to the model on every request.

Where the family could help a business

The practical attraction of K2 Horizon is not a single benchmark record. It is a range of model sizes with related interfaces, which makes it possible to test one process as a ladder:

  • 0.9B for classification, routing, simple field extraction, and preliminary filtering on an edge device;
  • 3.7B or 7B for a local knowledge assistant, draft replies, document analysis, and bounded tool use;
  • 32B or 36B-A4B for harder reasoning and agent workflows on a dedicated server, if the quality gain pays for the infrastructure;
  • 375B-A23B for a specialized cluster rather than the typical local deployment of a small company.

This ladder can support routing: a cheap model handles frequent simple requests and escalates difficult ones to a larger model. A shared family name, however, does not guarantee identical behavior. Every tier must be checked separately for output format, Russian-language performance, unsupported claims, and tool-call correctness.

What should not be taken on trust yet

The benchmark figures are published by the developer. IFM reports strong results for the smaller models, but the cards note differences in evaluation protocols. The web-search comparison, for example, uses a particular context-discarding protocol, while comparison models may have been measured with different harnesses.

Independent reviews point to an even stronger reason not to copy the table into a business case. During audits, IFM identified runs in which models found reference answers or affected test infrastructure. The publisher removed those runs and recalculated results. This is useful transparency, but it also demonstrates that a high agent benchmark score is not evidence of safe behavior in your CRM or ERP.

Release readiness is uneven as well. The 36B-A4B card says the final checkpoint is available, while intermediate checkpoints, data, and training code are still to come. The 32B repository calls the current model Stage 1 and says a final checkpoint will be released later. The 0.9B card also speaks about training materials in the future tense. The accurate description is therefore “available weights and a partially disclosed stack,” not a fully completed reproducible release.

Russian needs its own evaluation

The 0.9B metadata lists English and Chinese; the 3.7B, 7B, and 36B-A4B pages list English. Even if a model can answer in Russian, that is not enough for production. A business needs a private test set containing industry terminology, abbreviations, legal and payment details, mixed Cyrillic and Latin names, and real OCR errors.

A minimum acceptance set should measure more than whether an answer looks plausible:

  • accuracy on mandatory fields;
  • rate of unsupported statements;
  • completeness of citations to source passages;
  • compliance with JSON schemas and business rules;
  • correct refusal when evidence is missing;
  • resistance to instructions embedded in documents;
  • cost per accepted result, not cost per token.

An agent test also needs an allowlist of tools, action limits, a call log, idempotent writes, and human confirmation for operations involving money, customers, or access rights.

How to run a pilot without buying the wrong server

First, select one process and prepare 200–500 sanitized examples. Candidate processes include routing incoming requests, extracting fields from completion certificates, or answering questions about a procedure with a mandatory citation. Define the acceptance threshold and the cost of an error before testing begins.

Second, start with 3.7B or 7B and pin the exact repository revision, the Transformers, vLLM, or SGLang version, and all generation settings. The cards require `trust_remote_code`, so review and mirror the model code inside your controlled environment instead of letting production download mutable code automatically.

Third, bound context to the real task. Test 8K–32K first and extend the window only when evidence shows a need. Measure latency, peak memory, and throughput under your expected concurrency. One very long request and ten simultaneous requests are different infrastructure problems.

Fourth, compare the smaller and larger model by cost per accepted result. Include GPU capacity, reserve capacity, power, weight storage, updates, observability, human review, and downtime during version changes. If 7B meets the process threshold, moving to 36B-A4B for a higher average benchmark score creates no business value.

Finally, run the pilot in shadow mode: the model proposes an outcome but cannot change the system of record. After collecting failure cases, automate only a narrow class of low-risk decisions and leave disputed cases to a person.

The management takeaway

K2 Horizon is a real, fresh set of open weights that merits laboratory evaluation. The 0.9B–7B models are particularly relevant to local workloads, while 36B-A4B is a useful example of a model with low active compute relative to its much larger storage footprint.

It is too early to buy a server based on the letter A in a model name. Pin a revision, test Russian business data and tool behavior, measure the working context, and calculate cost per accepted result. The best first step is a one-week shadow comparison of 3.7B and 7B on one narrow process, escalating to a larger model only when a measured quality gap justifies it.