{"id":4269,"date":"2026-09-25T17:59:44","date_gmt":"2026-09-25T08:59:44","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/model-ifm-k2-horizon-en\/"},"modified":"2026-11-12T00:08:04","modified_gmt":"2026-11-11T15:08:04","slug":"model-ifm-k2-horizon-en","status":"publish","type":"page","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/model-ifm-k2-horizon-en\/","title":{"rendered":"K2-Horizon Guide: VRAM Requirements, GGUF Builds"},"content":{"rendered":"<h2>About This Model<\/h2>\n<p><strong>K2-Horizon<\/strong> is a family of language models from IFM. What sets it apart from most open-weight models, which publish weights only, is that it aims to be <strong>fully open: pre-training and mid-training data, training code, and intermediate checkpoints are all being released.<\/strong> Every size supports a 524,288-token (512K) context, and the family emphasizes reasoning (thinking) mode and agentic use with tool calls.<\/p>\n<p>The family has three models of different sizes:<\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Model<\/th>\n<th>Architecture<\/th>\n<th style=\"text-align: right;\">Parameters active per token<\/th>\n<th>Role<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>K2-Horizon-375B-A23B<\/td>\n<td>MoE (375B total)<\/td>\n<td style=\"text-align: right;\">23B<\/td>\n<td>Flagship<\/td>\n<\/tr>\n<tr>\n<td>K2-Horizon-MoVA-36B-A4B<\/td>\n<td>MoE (36B total, with its own &#8220;MoVA&#8221; attention)<\/td>\n<td style=\"text-align: right;\">4B<\/td>\n<td>Lightweight sparse model<\/td>\n<\/tr>\n<tr>\n<td>K2-Horizon-32B (Stage 1)<\/td>\n<td>Dense 32B<\/td>\n<td style=\"text-align: right;\">32B<\/td>\n<td>Mid-training release (final version to come)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>The models this site has covered, with their memory requirements, are listed under &#8220;Our Coverage and Data&#8221; below.<\/p>\n<h2>What Makes It Stand Out<\/h2>\n<ul>\n<li><strong>The flagship 375B-A23B is positioned as strong on agentic tasks.<\/strong> In the publisher&#8217;s comparison it scores 70.2% on Terminal-Bench 2.1 (terminal use), 65.3% on Toolathlon Verified (tool use) and 67.7% on MCPMark (tool use via MCP). The publisher claims it matches or beats open MoE models up to 2.6\u00d7 its size and competes with closed frontier models.<\/li>\n<li><strong>MoVA-36B-A4B scores high for a model that runs only 4B parameters per token.<\/strong> It posts 58.6% on Terminal-Bench 2.1 and 26.8% on tau3-Banking (business tool use), and the publisher says it outscores ~30B dense models and MoE models up to 15\u00d7 its size. With so little compute per token, it should run faster than a dense model of similar size.<\/li>\n<li><strong>Honestly, the 32B (Stage 1) is still a work in progress.<\/strong> Even in the publisher&#8217;s own table, it trails similarly sized models from other developers on most benchmarks. It is released as an intermediate checkpoint so researchers can study how capabilities change during training.<\/li>\n<li><strong>The NVFP4 builds lose very little.<\/strong> The average across seven evaluations the publisher ran under identical settings:<\/li>\n<\/ul>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Model<\/th>\n<th style=\"text-align: right;\">Original (BF16)<\/th>\n<th style=\"text-align: right;\">NVFP4 build<\/th>\n<th style=\"text-align: right;\">Difference<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>K2-Horizon-375B-A23B<\/td>\n<td style=\"text-align: right;\">91.7<\/td>\n<td style=\"text-align: right;\">91.2<\/td>\n<td style=\"text-align: right;\">\u22120.5<\/td>\n<\/tr>\n<tr>\n<td>K2-Horizon-32B (Stage 1)<\/td>\n<td style=\"text-align: right;\">89.5<\/td>\n<td style=\"text-align: right;\">88.4<\/td>\n<td style=\"text-align: right;\">\u22121.1<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>Shrinking the weights to roughly 4 bits costs only about one point on average. Note that the NVFP4 builds were evaluated only on non-agentic tasks so far (agentic results are to follow).<\/p>\n<h2>Running It Locally<\/h2>\n<ul>\n<li><strong>The NVFP4 builds require an NVIDIA GPU from the Blackwell generation or later with native NVFP4 support.<\/strong> They will not run on RTX 40-series cards, nor on Hopper-generation data-center GPUs such as the H100 and H200.<\/li>\n<li><strong>The 375B-A23B is not realistic for personal hardware.<\/strong> For the original BF16 version, the publisher&#8217;s validated setup is a single node with eight H200s. If you want to try the family locally, the GGUF build of MoVA-36B-A4B is the most practical option.<\/li>\n<li>The publisher recommends always using <code>reasoning_effort=\"high\"<\/code> and enabling the dedicated <code>k2_horizon<\/code> parsers in vLLM and SGLang.<\/li>\n<li><strong>The license is Apache-2.0, which allows commercial use.<\/strong><\/li>\n<\/ul>\n<p><em>Sources: model cards for <a href=\"https:\/\/huggingface.co\/IFM\/K2-Horizon-375B-A23B-NVFP4\">IFM\/K2-Horizon-375B-A23B-NVFP4<\/a>, <a href=\"https:\/\/huggingface.co\/IFM\/K2-Horizon-32B-NVFP4\">IFM\/K2-Horizon-32B-NVFP4<\/a> and <a href=\"https:\/\/huggingface.co\/IFM\/K2-Horizon-MoVA-36B-A4B\">IFM\/K2-Horizon-MoVA-36B-A4B<\/a>, as of 2026-09-25. Benchmark figures are as published by IFM.<\/em><\/p>\n<h2>Our Coverage and Data<\/h2>\n<p>Everything Local Model Watch has published about the <strong>K2-Horizon<\/strong> family: 3 article(s) covering the base model and its fine-tunes, plus converted builds we tracked after publication. Memory requirements below are computed by this site from file sizes, not quoted from model cards. Part of our <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/models-en\/\">model family index<\/a>.<\/p>\n<h2>At a Glance<\/h2>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Item<\/th>\n<th>Value<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>Base model(s)<\/td>\n<td><a href=\"https:\/\/huggingface.co\/IFM\/K2-Horizon-32B-NVFP4\">IFM\/K2-Horizon-32B-NVFP4<\/a>, <a href=\"https:\/\/huggingface.co\/IFM\/K2-Horizon-375B-A23B-NVFP4\">IFM\/K2-Horizon-375B-A23B-NVFP4<\/a>, <a href=\"https:\/\/huggingface.co\/IFM\/K2-Horizon-MoVA-36B-A4B\">IFM\/K2-Horizon-MoVA-36B-A4B<\/a><\/td>\n<\/tr>\n<tr>\n<td>Publisher<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/publisher-ifm-en\/\">IFM (MBZUAI)<\/a><\/td>\n<\/tr>\n<tr>\n<td>License (model card)<\/td>\n<td>apache-2.0<\/td>\n<\/tr>\n<tr>\n<td>Smallest VRAM tier<\/td>\n<td>32GB<\/td>\n<\/tr>\n<tr>\n<td>Articles<\/td>\n<td>3<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<h2>Hardware Requirements<\/h2>\n<p><strong>Estimated requirements (calculated by Local Model Watch)<\/strong><\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Your VRAM<\/th>\n<th>Quantization<\/th>\n<th>File size<\/th>\n<th>Est. memory needed<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>32GB (RTX 5090, etc.)<\/td>\n<td>NVFP4<\/td>\n<td>21.7GB<\/td>\n<td>26.0GB<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><em>Memory estimates add a 20% runtime overhead (KV cache, etc.) to the actual size of the distributed files. Actual usage varies with context length, batch size and inference engine. These figures are computed by this site from file sizes, not published by the model&#8217;s authors. Compare with other models in our <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/vram-guide-en\/\">VRAM quick reference<\/a>. What the quantization names mean: <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/glossary-quantization-en\/\">glossary<\/a>.<\/em><\/p>\n<p><em>Figures are for IFM\/K2-Horizon-32B-NVFP4. Other sizes in this family are covered in the articles below.<\/em><\/p>\n<h2>Can You Run It Locally?<\/h2>\n<p><strong>Not usable in Ollama, LM Studio and llama.cpp yet \u2014 we have found no GGUF build.<\/strong><\/p>\n<p>The publisher ships safetensors only. However, llama.cpp&#8217;s registry does list this architecture, so <strong>conversion to GGUF is possible<\/strong> and the model will run once someone publishes a converted build. Today it can be run with transformers or vLLM, using the memory figures in the table above.<\/p>\n<p><strong>License \u2014 <code>apache-2.0<\/code> (Commercial use allowed):<\/strong> Permits commercial use, modification and redistribution. Redistribution requires including the license and stating changes; includes a patent grant.<\/p>\n<p><em>Compiled by this site&#8217;s code from the published formats, converted builds we have found, and each engine&#8217;s own model registry. &#8220;Not found&#8221; means we have not seen such a build, not that none exists. License summaries are not legal advice \u2014 check the publisher&#8217;s original terms before relying on them.<\/em><\/p>\n<p><em>This assessment is for IFM\/K2-Horizon-32B-NVFP4.<\/em><\/p>\n<h2>Quantized and Converted Variants<\/h2>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<p class=\"lmw-table-hint\" style=\"margin:0 0 4px;font-size:0.85em;opacity:0.7;\">\u2192 Scroll horizontally to see all columns<\/p>\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Added<\/th>\n<th>Publisher<\/th>\n<th>Format<\/th>\n<th>Repository<\/th>\n<th>Smallest VRAM tier (build, est. memory)<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>2026-10-08<\/td>\n<td>bartowski<\/td>\n<td>GGUF (imatrix)<\/td>\n<td><a href=\"https:\/\/huggingface.co\/bartowski\/K2-Horizon-MoVA-36B-A4B-GGUF\">bartowski\/K2-Horizon-MoVA-36B-A4B-GGUF<\/a><\/td>\n<td>IQ2_XXS 11.9GB (fits in 12GB VRAM)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>File sizes of each build:<\/p>\n<ul>\n<li>Available builds in bartowski\/K2-Horizon-MoVA-36B-A4B-GGUF: IQ2_XXS 9.9GB \/ IQ2_XS 10.9GB \/ IQ2_S 11.2GB \/ IQ2_M 12.2GB \/ Q2_K 12.6GB \/ IQ3_XXS 14.7GB \/ Q3_K_S 15.6GB \/ IQ3_XS 16.1GB \/ Q3_K_M 16.2GB \/ Q3_K_L 16.6GB \/ IQ3_M 16.7GB \/ IQ4_XS 19.0GB \/ IQ4_NL 20.1GB \/ Q4_0 20.1GB \/ Q4_K_S 20.5GB \/ Q4_K_M 21.1GB \/ Q4_K_L 21.6GB \/ Q4_1 22.1GB \/ Q5_K_S 24.2GB \/ Q5_K_M 24.9GB \/ Q6_K 29.9GB \/ Q6_K_L 30.2GB \/ Q8_0 37.1GB \/ BF16 69.8GB<\/li>\n<\/ul>\n<p><em>This section is appended automatically by Local Model Watch when a converted build of this model appears after publication. Memory figures are estimated from the size of the distributed files.<\/em><\/p>\n<h2>Popularity on Hugging Face<\/h2>\n<p><em>Weekly snapshots recorded by Local Model Watch after publication. &#8220;Downloads&#8221; is Hugging Face&#8217;s own trailing-30-day count, not a lifetime total.<\/em><\/p>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Week<\/th>\n<th>Likes<\/th>\n<th>Downloads (trailing 30 days)<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>2026-W40<\/td>\n<td>0<\/td>\n<td>13<\/td>\n<\/tr>\n<tr>\n<td>2026-W41<\/td>\n<td>0<\/td>\n<td>251<\/td>\n<\/tr>\n<tr>\n<td>2026-W42<\/td>\n<td>0<\/td>\n<td>296<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<h2>Articles (the family&#8217;s own models first, then newest)<\/h2>\n<div class=\"lmw-table-scroll\" tabindex=\"0\" style=\"overflow-x:auto;-webkit-overflow-scrolling:touch;max-width:100%;\">\n<table style=\"width:max-content;min-width:100%;border-collapse:collapse;\">\n<thead>\n<tr>\n<th>Published<\/th>\n<th>Model<\/th>\n<th>Type<\/th>\n<th>Article<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>2026-09-23<\/td>\n<td>IFM\/K2-Horizon-32B-NVFP4<\/td>\n<td>New Models<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/09\/23\/ifm-k2-horizon-32b-nvfp4-released\/\">K2-Horizon-32B-NVFP4: Our Test Answers, 32GB+ VRAM<\/a><\/td>\n<\/tr>\n<tr>\n<td>2026-09-23<\/td>\n<td>IFM\/K2-Horizon-375B-A23B-NVFP4<\/td>\n<td>New Models<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/09\/23\/ifm-k2-horizon-375b-a23b-nvfp4-released-2\/\">K2-Horizon-375B-A23B-NVFP4 Text Generation Model: ~257GB Memory<\/a><\/td>\n<\/tr>\n<tr>\n<td>2026-09-06<\/td>\n<td>IFM\/K2-Horizon-MoVA-36B-A4B-GGUF<\/td>\n<td>New Models<\/td>\n<td><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/2026\/09\/06\/k2-horizon-mova-36b-a4b-gguf\/\">Japanese article only<\/a><\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<h2>Repositories<\/h2>\n<ul>\n<li><a href=\"https:\/\/huggingface.co\/IFM\/K2-Horizon-32B-NVFP4\">IFM\/K2-Horizon-32B-NVFP4<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/IFM\/K2-Horizon-375B-A23B-NVFP4\">IFM\/K2-Horizon-375B-A23B-NVFP4<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/IFM\/K2-Horizon-MoVA-36B-A4B\">IFM\/K2-Horizon-MoVA-36B-A4B<\/a><\/li>\n<li><a href=\"https:\/\/huggingface.co\/IFM\/K2-Horizon-MoVA-36B-A4B-GGUF\">IFM\/K2-Horizon-MoVA-36B-A4B-GGUF<\/a><\/li>\n<\/ul>\n<p><em>Last updated 2026-11-12 (JST). The explanation at the top of this page was written with the help of AI from the primary sources it cites. The tables and lists under &#8220;Our Coverage and Data&#8221; are assembled by code from our article log.<\/em><\/p>\n","protected":false},"excerpt":{"rendered":"<p>About This Model K2-Horizon is a family of language models from IFM. What sets it apart from most open-weight models, which publish weights only, is that it aims to be fully open: pre-training and mid-training data, training code, and intermediate checkpoints are all being released. Every size supports a 524,288-token (512K) context, and the family [&hellip;]<\/p>\n","protected":false},"author":1,"featured_media":0,"parent":0,"menu_order":0,"comment_status":"closed","ping_status":"closed","template":"","meta":{"footnotes":""},"class_list":["post-4269","page","type-page","status-publish","hentry"],"lang":"en","translations":{"en":4269,"ja":4268},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/pages\/4269","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/pages"}],"about":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/types\/page"}],"author":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/users\/1"}],"replies":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/comments?post=4269"}],"version-history":[{"count":49,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/pages\/4269\/revisions"}],"predecessor-version":[{"id":15956,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/pages\/4269\/revisions\/15956"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=4269"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}