{"id":15808,"date":"2026-11-17T11:04:04","date_gmt":"2026-11-17T02:04:04","guid":{"rendered":"https:\/\/localmodelwatch.tsuchitsuchi.com\/?p=15808"},"modified":"2026-11-17T11:13:36","modified_gmt":"2026-11-17T02:13:36","slug":"botpsy-0-8b-agentic-model","status":"publish","type":"post","link":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/botpsy-0-8b-agentic-model\/","title":{"rendered":"BotPsy: 0.8B On-Device Agentic Model by Tether AI Research"},"content":{"rendered":"<p><!-- lmw:facts --><\/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>Publisher<\/td>\n<td>Hugging Face Blog<\/td>\n<\/tr>\n<tr>\n<td>Published<\/td>\n<td>2026-11-11<\/td>\n<\/tr>\n<tr>\n<td>Source type<\/td>\n<td>Primary source (the publisher itself)<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p><em>Values determined by this site&#8217;s code when the information was collected. Dates are JST.<\/em><\/p>\n<p><!-- \/lmw:facts --><\/p>\n<h2>Overview<\/h2>\n<p>Tether AI Research has announced &#8220;BotPsy&#8221;, a 0.8B parameter model specialized in function calling and agent processing. Based on Qwen3.5-0.8B, it is designed for on-device operation. According to the blog, the weights are scheduled to be provided in safetensors and GGUF (bf16, Q8, Q4). Since the blog uses the phrasing &#8220;will be released&#8221;, please check the repository for the current availability status.<\/p>\n<p>It is targeted at simple personal assistant tasks that use tools, such as web search, travel, payments, device control, files, calendar, and email. The publishers themselves state that it is intended for low-to-medium difficulty tasks and is not suited for high-risk use cases.<\/p>\n<h2>Specifications<\/h2>\n<ul>\n<li>Number of Parameters: 0.8B<\/li>\n<li>Base Model: Qwen3.5-0.8B<\/li>\n<li>Training Procedure: A 4-stage post-training consisting of SFT, Multi-teacher On-Policy Distillation (MOPD), and two rounds of Reinforcement Learning (RL)<\/li>\n<li>Distillation Teachers: Three Qwen3.5-4B models trained via RL for tool use, general reasoning, and instruction following, respectively<\/li>\n<\/ul>\n<h2>Performance<\/h2>\n<p>The following figures are all self-reported measurements by the publishing blog and have not been independently verified by third parties. Table 1 (tool use and agent benchmarks, 0-100, higher is better) lists the 5 columns as presented.<\/p>\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>Benchmark<\/th>\n<th>BotPsy-0.8B<\/th>\n<th>Qwen3.5-0.8B-Think (Base)<\/th>\n<th>Qwen3.5-2B-Think<\/th>\n<th>LFM2.5-1.2B-Think<\/th>\n<th>MiniCPM5-1B-Think<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>BFCL v4<\/td>\n<td>44.0<\/td>\n<td>23.4<\/td>\n<td>37.7<\/td>\n<td>28.1<\/td>\n<td>22.8<\/td>\n<\/tr>\n<tr>\n<td>ACEBench<\/td>\n<td>58.7<\/td>\n<td>30.1<\/td>\n<td>55.4<\/td>\n<td>39.4<\/td>\n<td>24.5<\/td>\n<\/tr>\n<tr>\n<td>\u03c4-bench<\/td>\n<td>51.5<\/td>\n<td>23.2<\/td>\n<td>46.5<\/td>\n<td>6.5<\/td>\n<td>28.8<\/td>\n<\/tr>\n<tr>\n<td>Claw-Eval<\/td>\n<td>46.5<\/td>\n<td>39.3<\/td>\n<td>48.9<\/td>\n<td>31.6<\/td>\n<td>31.0<\/td>\n<\/tr>\n<tr>\n<td>BrowseComp-Plus<\/td>\n<td>4.5<\/td>\n<td>1.9<\/td>\n<td>2.8<\/td>\n<td>0.2<\/td>\n<td>0.8<\/td>\n<\/tr>\n<tr>\n<td>ToolSandbox<\/td>\n<td>64.6<\/td>\n<td>30.5<\/td>\n<td>41.2<\/td>\n<td>45.7<\/td>\n<td>28.1<\/td>\n<\/tr>\n<tr>\n<td>Average<\/td>\n<td>45.0<\/td>\n<td>24.7<\/td>\n<td>38.8<\/td>\n<td>25.3<\/td>\n<td>22.7<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<p>The knowledge and instruction-following tables (Table 2 and 3) are narrowed down to three columns: BotPsy, base, and Qwen3.5-2B-Think.<\/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>Benchmark<\/th>\n<th>BotPsy-0.8B<\/th>\n<th>Qwen3.5-0.8B-Think<\/th>\n<th>Qwen3.5-2B-Think<\/th>\n<\/tr>\n<\/thead>\n<tbody>\n<tr>\n<td>GPQA Diamond<\/td>\n<td>38.1<\/td>\n<td>33.2<\/td>\n<td>59.9<\/td>\n<\/tr>\n<tr>\n<td>MMLU-Pro<\/td>\n<td>48.9<\/td>\n<td>41.6<\/td>\n<td>65.5<\/td>\n<\/tr>\n<tr>\n<td>IFEval<\/td>\n<td>90.8<\/td>\n<td>59.7<\/td>\n<td>80.0<\/td>\n<\/tr>\n<tr>\n<td>IFBench<\/td>\n<td>55.9<\/td>\n<td>22.0<\/td>\n<td>40.9<\/td>\n<\/tr>\n<tr>\n<td>Multi-IF<\/td>\n<td>67.2<\/td>\n<td>43.2<\/td>\n<td>66.9<\/td>\n<\/tr>\n<\/tbody>\n<\/table>\n<\/div>\n<h3>What the Tables Tell Us<\/h3>\n<p>BotPsy achieves the highest score in 5 out of 6 agent-related benchmarks. On BFCL v4 (accuracy in function selection and argument assembly), it scores 44.0, outperforming the 2.5x larger Qwen3.5-2B-Think at 37.7. It also leads on \u03c4-bench (balancing policy-compliant dialogue and tool calling) at 51.5 versus 46.5. It shows nearly double the gains from the base model across many items.<\/p>\n<p>On the other hand, Claw-Eval is the exception at 46.5, falling short of Qwen3.5-2B-Think&#8217;s 48.9. BrowseComp-Plus is also the highest value, but at 4.5 the absolute value is low, and the blog itself notes open-ended search as a weakness.<\/p>\n<p>In terms of knowledge, GPQA Diamond and MMLU-Pro improved over the base model. However, they fall well short of Qwen3.5-2B-Think&#8217;s 59.9 and 65.5. The blog also explains that the model is designed to look up facts via tools and is not a knowledge model. IFEval is the highest among the comparison targets at 90.8, demonstrating a strong capability in following formal instructions. Multi-IF is at a comparable level, though according to the blog, low-resource languages such as Hindi (24.5) and Russian (57.6) remain weak.<\/p>\n<p>The blog additionally reports that the harmful response rate in red-teaming dropped to 21% (compared to 36% for the base).<\/p>\n<h2>Strengths and Use Cases<\/h2>\n<p>BotPsy is aimed at acting as a &#8220;personal assistant&#8221; that invokes tools to perform tasks. The domains cited by the blog include web search and research, travel and lifestyle, commerce and finance, business and customer support, device and vehicle operation, files and documents, calendars, and email and messaging.<\/p>\n<ul>\n<li><strong>Single-turn tool invocation<\/strong>: According to the blog, simple scenarios (1 turn, 1 invocation) score 70\u201379 across all application groups with sufficient samples. Overall, it is reported to score 75, surpassing the base model&#8217;s 54 and Qwen3.5-2B-Think&#8217;s 70.<\/li>\n<li><strong>Multi-turn dialogue<\/strong>: It is stated to reach 82.5 on \u03c4\u00b2-bench telecom (compared to 69.3 for Qwen3.5-2B-Think), and improved from 7.5 to 46.1 on BFCL v4 multi-turn.<\/li>\n<li><strong>Determining when not to call tools<\/strong>: It achieves 86.1 on the BFCL v4 hallucination test, which measures the ability to decline when no appropriate tool is available, the highest among tested models.<\/li>\n<li><strong>Instruction following<\/strong>: IFEval and IFBench show significant improvements.<\/li>\n<li><strong>Safety<\/strong>: The harmful response rate under prompt injection is reported to have dropped from 46% for the base model to 31%.<\/li>\n<\/ul>\n<p>Weaknesses are also explicitly noted by the publishers.<\/p>\n<ul>\n<li><strong>Multi-step information gathering<\/strong>: Complex Information Retrieval &amp; Research scenarios remain at 15 points (whereas single-turn is 79 points).<\/li>\n<li><strong>Web search and memory<\/strong>: BFCL web search stands at 14.0 and memory at 16.3, trailing Qwen3.5-2B-Think&#8217;s 19.8 and 26.1.<\/li>\n<li><strong>Knowledge<\/strong>: Knowledge capacity primarily depends on parameter count. The model is designed to look up facts via tools and should not be used as a knowledge model.<\/li>\n<li><strong>Multilingual<\/strong>: Hindi is low at 24.5 in Multi-IF.<\/li>\n<li><strong>Complex tasks<\/strong>: Complex scenarios overall score 39.5, remaining lower than single-turn tasks.<\/li>\n<\/ul>\n<p>The publishers note that due to its small size, errors can occur, and high-risk use cases are not envisioned. Please consider it suitable for low-to-medium difficulty tasks.<\/p>\n<p>Inference efficiency is also highlighted as a feature. Comparing problems where all models answered correctly, GPQA Diamond requires approximately one-third of the reasoning tokens of the base model (3,690 versus 11,563). The rate of falling into repetition loops on IFBench is reported to have dropped from 22.9% to 1.4%. The blog states that reduced latency on-device can be expected. However, these are measurements by the publishers.<\/p>\n<h2>How to Get It<\/h2>\n<p>According to the blog, BotPsy is scheduled to be released as an open-weight model under the Apache 2.0 license. However, the blog notes &#8220;research and educational use&#8221;, and the permissibility of commercial use cannot be clearly determined from the documentation. Please verify the license terms in the repository.<\/p>\n<ul>\n<li><strong>Distribution Format<\/strong>: Hugging Face safetensors and GGUF. GGUF quantization levels are available in three types: bf16, Q8, and Q4.<\/li>\n<li><strong>Supported Engines<\/strong>: GGUF is described as being for the QVAC SDK and llama.cpp.<\/li>\n<li><strong>Where to Get It<\/strong>: Links to the model (safetensors) and GGUF are provided within the blog. However, since the repository name is not written in the documentation, download commands are omitted. Please navigate via the blog post.<\/li>\n<\/ul>\n<p>The blog states &#8220;will be released&#8221;. Since confirmation of the release timing could not be obtained, please check the actual download availability in the repository first.<\/p>\n<p><!-- lmw:related --><\/p>\n<h2>Related Articles<\/h2>\n<ul>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/mistral-large-4-released-2\/\">Mistral Large 4 Released: Multimodal MoE Model<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/anylanguagemodel-1-0-released\/\">AnyLanguageModel 1.0 Released for Apple Platforms<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/meta-muse-spark-open-2\/\">Meta Releases Muse Spark Open: Multimodal MoE Model<\/a><\/li>\n<li><a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/2026\/11\/17\/liquidai-lfm2-5-24b-a2b-2\/\">LiquidAI Releases LFM2.5-24B-A2B for On-Device Agents<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:related --><\/p>\n<p><!-- lmw:next-steps --><\/p>\n<h2>What to Read Next<\/h2>\n<ul>\n<li><strong>How to read BFCL, \u03c4-bench, GPQA Diamond<\/strong> \u2192 <a href=\"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/glossary-benchmarks-en\/\">Benchmark glossary<\/a><\/li>\n<\/ul>\n<p><!-- \/lmw:next-steps --><\/p>\n<h2>Sources<\/h2>\n<ul>\n<li><a href=\"https:\/\/huggingface.co\/blog\/qvac\/botpsy\">BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agentic Language Model (Hugging Face Blog)<\/a><\/li>\n<\/ul>\n","protected":false},"excerpt":{"rendered":"<p>Explore BotPsy, a 0.8B parameter agentic model by Tether AI Research specialized for function calling and on-device personal assistant tasks.<\/p>\n","protected":false},"author":1,"featured_media":15807,"comment_status":"closed","ping_status":"closed","sticky":false,"template":"","format":"standard","meta":{"footnotes":""},"categories":[310],"tags":[4879,163,1907,165,996,4881,1547,2795],"class_list":["post-15808","post","type-post","status-publish","format-standard","has-post-thumbnail","hentry","category-new-models","tag-botpsy-en","tag-gguf-en","tag-hugging-face-blog-en","tag-moe-en","tag-qwen3-5-en","tag-tether-ai-research-en","tag-verified","tag--en"],"lang":"en","translations":{"en":15808,"ja":15806},"_links":{"self":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/15808","targetHints":{"allow":["GET"]}}],"collection":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts"}],"about":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/types\/post"}],"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=15808"}],"version-history":[{"count":2,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/15808\/revisions"}],"predecessor-version":[{"id":18090,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/posts\/15808\/revisions\/18090"}],"wp:featuredmedia":[{"embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media\/15807"}],"wp:attachment":[{"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/media?parent=15808"}],"wp:term":[{"taxonomy":"category","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/categories?post=15808"},{"taxonomy":"post_tag","embeddable":true,"href":"https:\/\/localmodelwatch.tsuchitsuchi.com\/en\/wp-json\/wp\/v2\/tags?post=15808"}],"curies":[{"name":"wp","href":"https:\/\/api.w.org\/{rel}","templated":true}]}}