Contents

BotPsy is a promising option for running narrowly scoped assistant workflows locally. According to Tether AI Research, this 800-million-parameter model outperforms the 2B comparison model on five of six tool-use benchmarks. It is worth trying for routine tool calls, but its search performance and safety scores suggest it is not ready for unrestricted autonomous operation.

Key Takeaways#

  • BotPsy specializes in function calling and reportedly outperforms its base model, Qwen3.5-0.8B, on all six agent benchmarks evaluated.
  • It outperforms Qwen3.5-2B-Think on five of those benchmarks, while falling well behind in general knowledge.
  • Local inference keeps model processing on your hardware, but where data is sent also depends on the connected tools.
  • The announcement lists safetensors and GGUF formats. Download availability, license terms, and performance on your device need to be checked.

What BotPsy Can Do#

Announced on November 11, 2026, BotPsy is a model further trained from Qwen3.5-0.8B to select tools and specify arguments. Intended tasks include calendar updates, document operations, customer support, and device control.

It helps to distinguish choosing an action from knowing an answer. BotPsy can use connected services to retrieve information or perform tasks, but it needs an application that provides tools, manages permissions, and executes calls. Model weights alone do not make a complete personal assistant.

Start with clearly bounded tasks, such as finding free calendar slots, retrieving order status, or drafting a message for review. It is far less suited to open-ended research.

What the Scores Show#

The research team reports the following results on a 0–100 scale. Higher scores indicate better performance.

| Benchmark | BotPsy 0.8B | Base Model 0.8B | Qwen3.5 2B | |---|---:|---:|---:| | 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 | | BrowseComp-Plus | 4.5 | 1.9 | 2.8 | | Claw-Eval | 46.5 | 39.3 | 48.9 |

These are developer-reported results against selected comparison models, not independent proof of broad superiority. The strongest practical evidence is the improvement in structured workflows. In BFCL's multi-turn evaluation, the score rises from 7.5 for the base model to 46.1 for BotPsy.

Absolute scores matter too. Even though it leads the comparison on BrowseComp-Plus, a score of 4.5 indicates poor performance on difficult research tasks. Similarly, the task-specific results show an overall score of 75 for simple scenarios, but only 39.5 for complex ones. One successful call does not necessarily mean an entire workflow can be reliably automated.

Training Method#

BotPsy is trained sequentially in four stages. Supervised fine-tuning teaches tool-use conversations, including examples where no tool should be called. Next, multi-teacher on-policy distillation uses three specialized Qwen3.5-4B teacher models to improve tool use, reasoning, and instruction following.

Two stages of reinforcement learning follow. The first focuses on tool calls, awarding partial credit for correct tool names, argument names, and argument values. The second broadens the scope to verifiable tasks, including math, code, structured output, and simulated business operations.

This training process addresses a specific tradeoff. Early tool specialization reduced knowledge and instruction-following scores, but later stages restored them. The final IFEval score reached 90.8, exceeding the base model's 59.7. This supports testing it with strict output requirements, but does not prove it can remain reliable over long conversations.

What to Check Before Deployment#

The announcement lists bf16, Q8, and Q4 versions and says safetensors and GGUF files will be provided for the QVAC SDK and llama.cpp. However, some wording describes the release in the future tense, so check whether the files you need are available. The actual license also needs checking: the text mentions Apache 2.0 while also describing research and educational use.

A small parameter count alone does not establish practical smartphone performance. The reported results include no data on device-specific latency, peak memory use, battery consumption, or quality loss from quantization. Fewer reasoning tokens suggest potential efficiency gains, but do not demonstrate measured speedups on hardware.

Safety is another deployment constraint. The team reports unsafe responses in 21% of automated red-team tests and 31% under prompt-injection attacks. Limit tool permissions, treat retrieved content as untrusted, and require confirmation before sending messages, making payments, or deleting files. Local inference also cannot prevent search or email tools from sending data externally.

Next Steps#

Before enabling access to real accounts, test 20 representative tasks in a sandbox on the target device, recording tool-argument accuracy, task completion, latency, and attempts at unauthorized actions.

Source: https://huggingface.co/blog/qvac/botpsy

Contents