BotPsy 0.8B: Tether's On-Device Tool Calling Model Guide

Last updated: November 12, 2026

BotPsy-0.8B is a free, open-weight model from Tether AI Research that turns requests into tool calls, such as adding a calendar event or sending a message. It is built on Qwen3.5-0.8B, runs on phones and laptops, and leads every model of 1.2B parameters or fewer that Tether tested on six agent benchmarks.

Contents

What BotPsy is and what is new
Specs and download files
How to run BotPsy with tool calling
Benchmarks compared with Qwen3.5
Workflows for power users
Limitations
Frequently asked questions

This guide is based on the official BotPsy-0.8B model card, the BotPsy GGUF model card, Tether’s technical write-up on Hugging Face and the QVAC launch post.

What BotPsy is and what is new

Tether AI Research released BotPsy on November 11, 2026, under its QVAC brand. The launch post describes it as a model “built to take actions for you,” like replying to an email or adding a trip to a calendar, while keeping that personal data on the device.

The model works through tool calls. An app gives BotPsy a list of functions, such as get_email, send_email or create_event. BotPsy picks the right one, fills in the arguments, reads the result and continues until the task is done.

What is new compared with its Qwen3.5-0.8B base is the training. Tether post-trained it in four stages: supervised fine-tuning, multi-teacher on-policy distillation from three specialist Qwen3.5-4B teachers, and two rounds of reinforcement learning with DAPO. According to the write-up, this roughly doubles the base model’s scores on key tool-use tests.

Specs and download files

Item	BotPsy-0.8B
Developer	Tether AI Research (QVAC)
Parameters	0.8B
Base model	Qwen3.5-0.8B (hybrid Gated DeltaNet and gated attention), used text only
Main use	Function calling and multi-step agent tasks of low to medium difficulty
Reasoning	Trained and evaluated in thinking mode (`enable_thinking=True`)
Recommended language	English
Weights	BF16 safetensors in `qvac/BotPsy-0.8B`
GGUF files	BF16 (about 1.52 GB), Q8_0 (about 0.81 GB), Q4_K_M with imatrix (about 0.53 GB)
License	Apache 2.0
Not included	Training code and training data

The GGUF card recommends the Q8_0 file, which scored like BF16 at about half the size. The base model’s multi-token prediction layers are not included in the GGUF files.

How to run BotPsy with tool calling

BotPsy keeps the Qwen3.5 architecture, so llama.cpp, vLLM and Hugging Face Transformers run it without custom code. These steps follow the official model cards:

Download the recommended file: hf download qvac/BotPsy-0.8B-GGUFs botpsy-0.8b-q8_0.gguf --local-dir .
Start an OpenAI-compatible server with thinking on: llama-server -m botpsy-0.8b-q8_0.gguf --jinja -c 32768 --temp 0.3 --top-p 0.95 --top-k 20 --presence-penalty 1.5 --chat-template-kwargs '{"enable_thinking": true}' --port 8080
Send a request to /v1/chat/completions with a tools array. The reply carries the reasoning in reasoning_content and the call in tool_calls.
For GPU serving, run vllm serve qvac/BotPsy-0.8B --language-model-only --enable-auto-tool-choice --tool-call-parser qwen3_xml --reasoning-parser qwen3 and pass chat_template_kwargs with enable_thinking set to true.
In Transformers (tested with 5.19.0), pass enable_thinking=True to apply_chat_template. Install flash-linear-attention and causal-conv1d for the fast kernels.
For a JavaScript app, install @qvac/sdk (tested with 0.23.0), load the GGUF file with tools: true, and keep the token limit at 2,048 or more because the model thinks before it answers.

The most important setting is thinking mode. The chat template defaults to non-thinking, but Tether trained and evaluated the model with thinking on. For hardware sizing, see our LM Studio system requirements answer, and for recent runtime changes, our llama.cpp guide and vLLM 0.32 guide.

Benchmarks compared with Qwen3.5

Tether ran all five compared models in thinking mode with one shared evaluation harness, serving BotPsy from BF16 weights with vLLM. These are the vendor’s own results.

Benchmark	BotPsy-0.8B	Qwen3.5-0.8B-Think (base)	Qwen3.5-2B-Think
BFCL v4	44.0	23.4	37.7
ACEBench	58.7	30.1	55.4
τ-bench	51.5	23.2	46.5
ToolSandbox	64.6	30.5	41.2
Claw-Eval	46.5	39.3	48.9
IFEval	90.8	59.7	80.0
GPQA Diamond	38.1	33.2	59.9
MMLU-Pro	48.9	41.6	65.5

BotPsy beats the 2.5-times-larger Qwen3.5-2B-Think on five of six agent tests, but not on Claw-Eval. Knowledge scores remain well below the 2B model, which is why Tether says BotPsy is designed to call a search tool for facts rather than answer from memory.

Two other results matter for agents. On the BFCL v4 hallucination test, which checks that a model declines when no tool fits, BotPsy scores 86.1, the highest of the models tested. On GPQA Diamond questions that every model answered correctly, it used 3,690 reasoning tokens per correct answer against 11,563 for its base.

Workflows for power users

Private personal assistant. Connect calendar, email and messaging functions so requests are handled on the device instead of a cloud service.
Cheap tool router. Put BotPsy in front of local functions for simple single-step requests, where Tether reports scores between 69 and 79 out of 100 in each app group with enough test cases.
Error recovery tests. Tether’s launch post shows BotPsy turning Wi-Fi back on and clearing a full inbox before resending a failed message. Build similar failure cases into your own tests.
Quantization checks. Compare Q8_0 and Q4_K_M on your own tool schemas; Tether found Q4_K_M lost little on tool use but more on knowledge.

For other recent small and local models, see our Alpie Seed 9B guide and our Bielik 11B v3.1 guide.

Limitations

All benchmark numbers come from Tether, not independent testing.
Multi-step tasks still score far lower than single calls, and deep web research is weak: BotPsy scores 4.5 on BrowseComp-Plus.
It trails Qwen3.5-2B-Think on BFCL web search (14.0 versus 19.8) and memory (16.3 versus 26.1).
Instruction following is much weaker in Hindi (24.5) and lower in Russian (57.6) than in English.
In red-team tests it still answered unsafely 21% of the time, and 31% under prompt injection. Tether recommends a separate input and output filter for apps with strict safety needs.
Tether says it is not intended for safety-critical decisions or for applications involving money, personal information or health.

Frequently asked questions

What is BotPsy?

BotPsy-0.8B is an open-weight language model from Tether AI Research, published under the QVAC name on November 11, 2026. It is built on Qwen3.5-0.8B and trained to turn a request into the right tool calls, over several steps when needed. It is small enough to run on phones and laptops.

Is BotPsy free?

Yes. The weights are free to download on Hugging Face under the Apache 2.0 license. Tether describes the model as developed for research and educational uses and says it is not intended for high-stakes decisions or applications involving money, health or personal information.

Which BotPsy GGUF file should I use?

Tether recommends botpsy-0.8b-q8_0.gguf, about 0.81 GB, which scored about the same as BF16 in its tests. The Q4_K_M file is about 0.53 GB and averaged about 1.5 points lower. The unquantized BF16 GGUF is about 1.52 GB.

Does BotPsy need an internet connection?

No. The model runs on the device, so an app built on it can work offline. A tool that the model calls, such as a web search or booking service, may still need a connection.

Is there a BotPsy app?

No. Tether says BotPsy is a model for developers who build it into their own assistants and agents. You run it through llama.cpp, vLLM, Hugging Face Transformers or the QVAC SDK.

BotPsy 0.8B: Power-User Guide to Tool Calling, GGUF Setup and Benchmarks