Tether BotPsy-0.8B: A 529 MB Tool-Calling Model for Phones
BotPsy-0.8B is an open tool-calling model from Tether's QVAC research team, released on 11 November 2026 under Apache 2.0. It is fine-tuned from Qwen3.5-0.8B, ships as GGUF files from 529 MB to 1.52 GB, and beats the 2.5 times larger Qwen3.5-2B on several tool-use tests in Tether's own evaluation. It is English-first and not meant for sensitive tasks.
On this page
What Tether released and how it was trained
Tether's AI research group published BotPsy-0.8B on 11 November 2026 through its QVAC project, describing it as an agent model small enough to run on a phone or laptop. The job it is built for is narrow: turn an everyday request such as replying to an email or adding a calendar event into a correct function call, rather than hold a long open-ended conversation.
The model is a fine-tune of Alibaba's Qwen3.5-0.8B, a text-only model that mixes Gated DeltaNet layers with gated attention. Both the base and the fine-tune are licensed under Apache 2.0. Training ran in four stages: supervised fine-tuning, distillation from three teacher models that were themselves produced by running reinforcement learning on Qwen3.5-4B, then two rounds of reinforcement learning on the small model.
One thing is not open. The model card says the training code and the training data are not released, and part of that data is an unpublished synthetic set generated with larger Qwen models. You get weights you can use commercially, but you cannot reproduce the training run.
How the benchmark scores compare with larger models
All of the numbers below come from Tether's own evaluation, so treat them as vendor results until someone else repeats them. Across six tool-use benchmarks, BotPsy averages 45.0, against 24.7 for its base model and 38.8 for Qwen3.5-2B, which has 2.5 times as many parameters. It leads the 2B model on BFCL v4 (44.0 against 37.7), ACEBench (58.7 against 55.4), tau-bench (51.5 against 46.5), and ToolSandbox (64.6 against 41.2).
The largest gap is in multi-turn function calling, where BotPsy scores 46.1, the 2B model 23.8, and the untuned base model 7.5. Tether also reports 86.1 on a test of whether the model declines when no tool fits the request, the highest of the models it tested. For an assistant that can act on your device, refusing to invent a call matters as much as getting the right one.
The larger model still wins where general knowledge counts. Qwen3.5-2B scores 59.9 on GPQA Diamond against 38.1 for BotPsy, 65.5 against 48.9 on MMLU-Pro, and edges ahead on Claw-Eval. Open-ended web research remains out of reach at this size: BotPsy manages 4.5 on BrowseComp-Plus.
File sizes, quantization loss, and how to run it
A separate repository carries three GGUF builds for llama.cpp and the QVAC SDK: full BF16 at 1.52 GB, Q8_0 at 812 MB, and an imatrix Q4_K_M at 529 MB. Tether recommends the Q8_0 file. Averaged over ten benchmarks, BF16 scores 50.4, Q8_0 scores 50.7, and Q4_K_M scores 48.9, so the 8-bit file loses nothing measurable and the 4-bit file gives up about 1.5 points for a third less storage.
Two settings are easy to miss. The chat template defaults to non-thinking mode, while the published scores were measured with thinking enabled, so you need to turn it on. Tether also recommends a temperature of 0.3 and leaving at least 2,048 output tokens because the model reasons before it answers.
Tether has not published speed or memory figures for any phone. The GGUF benchmarks were run on NVIDIA H100 GPUs, so real handset performance is still something you have to measure yourself.
Limits to weigh before building on it
Tether is direct about the boundaries. BotPsy is recommended for English, with lower scores in Russian and much lower in Hindi. It is meant for low to medium difficulty tasks, and the announcement says it should not be used for safety-critical decisions or for applications involving money, personal information, or health. For apps with strict safety needs, the authors suggest adding a separate input and output filter.
That caveat cuts against the obvious use. An on-device assistant that reads your messages and calendar is handling personal information by definition, so a developer has to supply the confirmation steps and permission limits that the model does not. The privacy gain is still real: with a 529 MB to 812 MB file, the tool-routing step of an assistant can stay on the device instead of going to a cloud API.