> ## Content Index > Fetch the complete content index at: https://www.notatechguy.com/llms.txt > Use this file to discover other available public pages before exploring further. # BotPsy 0.8B agent model beats Qwen 2B on five benchmarks - URL: https://www.notatechguy.com/botpsy-0-8b-agent-model-beats-qwen-2b-on-five-benchmarks/ - Published: 2026-11-12T04:56:02.000Z - Updated: 2026-11-12T04:56:03.000Z - Description: Tether AI's BotPsy doubles its base model's tool-use scores and beats Qwen3.5-2B on five of six benchmarks, with no independent verification. - Author: Marcello Babbili - Tags: Technology & AI, AI Models, Google Tether AI Research released BotPsy, a 0.8-billion-parameter model that beats a rival 2.5 times its size on five of six agent benchmarks [S¹](https://huggingface.co/blog/qvac/botpsy?ref=notatechguy.com). Every score is self-reported in a Hugging Face blog post dated 11 November 2026, and nobody outside the company has verified any of them [S¹](https://huggingface.co/blog/qvac/botpsy?ref=notatechguy.com). **My read:** The BFCL v4 multi-turn jump from 7.5 to 46.1 is the number I would watch, because multi-turn tool use, where a model must call a function, read the result, and decide what to do next, is exactly where small models usually collapse. But every comparison here is Tether's own measurement against models Tether chose to test. Until a third party reproduces even one score, the "state-of-the-art" claim is the developer's assertion, not an established fact. BotPsy is built on Qwen3.5-0.8B and fine-tuned for function calling, the mechanism that lets a language model trigger external tools rather than just generate text [S¹](https://huggingface.co/blog/qvac/botpsy?ref=notatechguy.com). Tether AI Research describes it as a personal assistant for low-to-medium difficulty tasks across domains including web search, travel bookings, commerce, device control, calendar management, and email [S¹](https://huggingface.co/blog/qvac/botpsy?ref=notatechguy.com). The model ships in safetensors and GGUF formats, including bf16, Q8, and Q4 quantisations, making it small enough to run on a phone or laptop without a cloud connection [S¹](https://huggingface.co/blog/qvac/botpsy?ref=notatechguy.com). The headline result is that BotPsy roughly doubles the agentic capability of its base model. On BFCL v4, a standard tool-use benchmark, it scores 44.0, up from 23.4 for Qwen3.5-0.8B-Think [S¹](https://huggingface.co/blog/qvac/botpsy?ref=notatechguy.com). ACEBench rises from 30.1 to 58.7, and τ-bench, which tests multi-step agent tasks in retail, telecom, and banking scenarios, jumps from 23.2 to 51.5 [S¹](https://huggingface.co/blog/qvac/botpsy?ref=notatechguy.com). ### Where the 0.8B model wins and where it doesn't BotPsy also outperforms Qwen3.5-2B-Think, a model with 2.5 times the parameters, on five of the six benchmarks Tether tested [S¹](https://huggingface.co/blog/qvac/botpsy?ref=notatechguy.com). On ToolSandbox the margin is widest: 64.6 versus 41.2 [S¹](https://huggingface.co/blog/qvac/botpsy?ref=notatechguy.com). On BFCL v4 it leads 44.0 to 37.7, and on τ-bench, 51.5 to 46.5 [S¹](https://huggingface.co/blog/qvac/botpsy?ref=notatechguy.com). ![BotPsy scores across six agent benchmarks](https://storage.ghost.io/c/6e/89/6e896869-22ef-4281-a213-b4c462c17cff/content/images/2026/11/chart_6428ca82b09aeb815014.png) The one exception is Claw-Eval, where Qwen3.5-2B-Think scores 48.9 against BotPsy's 46.5 [S¹](https://huggingface.co/blog/qvac/botpsy?ref=notatechguy.com). On BrowseComp-Plus, a web-browsing benchmark, both models score below 5, with BotPsy at 4.5 and Qwen at 2.8 [S¹](https://huggingface.co/blog/qvac/botpsy?ref=notatechguy.com). A score that low means neither model can reliably browse the web and answer questions, regardless of who ranks first. The multi-turn results are more striking. On BFCL v4's multi-turn tasks, where the model must carry a conversation through several tool calls, BotPsy scores 46.1, up from 7.5 for its base model and nearly double Qwen3.5-2B-Think's 23.8 [S¹](https://huggingface.co/blog/qvac/botpsy?ref=notatechguy.com). On τ²-bench telecom, it scores 82.5, ahead of Qwen3.5-2B-Think at 69.3 [S¹](https://huggingface.co/blog/qvac/botpsy?ref=notatechguy.com). Tether AI Research is explicit about the limits. Because BotPsy is a small model, it still makes mistakes and is not intended for high-stakes use [S¹](https://huggingface.co/blog/qvac/botpsy?ref=notatechguy.com). The company scopes its "state-of-the-art" claim to the specific models it tested in the sub-1.2-billion-parameter class, which included Qwen3.5-0.8B-Think, LFM2.5-1.2B-Think, and MiniCPM5-1B-Think [S¹](https://huggingface.co/blog/qvac/botpsy?ref=notatechguy.com). The claim does not extend to all larger models, and the benchmark results have not been independently verified [S¹](https://huggingface.co/blog/qvac/botpsy?ref=notatechguy.com). For developers building on-device assistants, the appeal is straightforward: a model under 1 billion parameters that can call tools, run multi-step workflows, and fit in a phone's memory. The on-device trend is gathering pace, as we saw when [an on-device speech model passed 200,000 downloads](https://www.notatechguy.com/phonon-tts-200-000-downloads-for-on-device-speech-model/) earlier this month, and when [an AI tutor chose an 8B model over bigger rivals](https://www.notatechguy.com/coursechat-ai-tutor-picks-8b-model-over-bigger-rivals/) in October. BotPsy pushes that frontier further down the parameter count. A developer considering BotPsy should run one test themselves before committing. Pick a real workflow, say, a calendar booking that requires checking availability, finding a free slot, and sending an invite, and see whether the 0.8B model completes it end to end. If it fails on a task this simple, the benchmark scores will not save the deployment. The model weights and GGUF files are available on Hugging Face now [S¹](https://huggingface.co/blog/qvac/botpsy?ref=notatechguy.com). The fine-tuning code lives in Tether AI Research's GitHub repository, qvac-fabric-llm-finetune, which has 46 stars and was created in November 2025 [P³](https://github.com/tether-ai-research/qvac-fabric-llm-finetune?ref=notatechguy.com). --- *Sources: [S1 — BotPsy: On-Device Personal Assistant with a Sub-Billion-Parameter Agen](https://huggingface.co/blog/qvac/botpsy?ref=notatechguy.com) · [P2 — open-jarvis/OpenJarvis](https://github.com/open-jarvis/OpenJarvis?ref=notatechguy.com) · [P3 — tether-ai-research/qvac-fabric-llm-finetune](https://github.com/tether-ai-research/qvac-fabric-llm-finetune?ref=notatechguy.com) · [P4 — Qwen3 - a Qwen Collection](https://huggingface.co/collections/Qwen/qwen3?ref=notatechguy.com)* ## Related reading - [Phonon TTS: 200,000 downloads for on-device speech model](https://www.notatechguy.com/phonon-tts-200-000-downloads-for-on-device-speech-model/) — our technology desk, 2026-11-06 - [Google DeepMind announces Genie 4 world model](https://www.notatechguy.com/google-deepmind-announces-genie-4-world-model/) — our technology desk, 2026-10-27 - [CourseChat AI tutor picks 8B model over bigger rivals](https://www.notatechguy.com/coursechat-ai-tutor-picks-8b-model-over-bigger-rivals/) — our technology desk, 2026-10-05 --- *Written from 4 sourced items, 4 of them primary.*