Tether's BotPsy: a 0.8B open tool-calling model that beats a 2B sibling

Tether AI Research, the AI arm of Tether Data, published BotPsy-0.8B on Hugging Face on 11 November 2026. It is a 0.8-billion-parameter model fine-tuned from Qwen3.5-0.8B for function calling, meant to run on a phone or laptop as a personal assistant that drives tools: calendar, email, files, device settings, web search, shopping.

The claim that matters for a team sizing an on-device or edge agent: Tether says the model beats Qwen3.5-2B, which is 2.5 times its size, on five of six tool-use benchmarks. Tether ran all of the tests itself.

What was released

• Weights. Safetensors in BF16, plus GGUF files for llama.cpp and Tether's QVAC SDK. The repository lists the files at about 1.52 GB (BF16), 0.81 GB (Q8) and 0.53 GB (Q4).
• Licence. The model card and the licence file say Apache 2.0. The post adds the phrase "for research and educational use"; the licence text itself carries no such restriction.
• Language. English is the recommended language. Tether reports 72 to 80 on the Multi-IF test in Chinese, French, Spanish, Italian and Portuguese, 57.6 in Russian and 24.5 in Hindi.

Tether's numbers

Tether compared five models on one shared harness: BotPsy, its Qwen3.5-0.8B base, Qwen3.5-2B, LFM2.5-1.2B and MiniCPM5-1B. Scores are percentages.

• BFCL v4 (function calling): 44.0, against 23.4 for the base and 37.7 for Qwen3.5-2B. On the multi-turn part it rises from 7.5 to 46.1.
• τ-bench (customer-service conversations under a policy; mean of retail, telecom and banking): 51.5, against 23.2 and 46.5. Telecom alone is 82.5.
• ToolSandbox: 64.6, against 30.5 and 41.2.
• Declining to call a tool when none fits (BFCL hallucination tests): 86.1, the highest of the five.
• Reasoning loops: runs that got stuck repeating themselves on IFBench fell from 22.9% to 1.4%, and runs that hit the 32,000-token limit from 28.3% to 0.2%.

Single tool calls are where it is strongest. Tether groups its test scenarios by application and reports 75 on simple ones (one turn, one call) and 39.5 on multi-turn or multi-tool ones. The base model scores 54 and 17.9.

Where Tether says it falls short

• Banking. On τ³-bench banking it scores 4.1, unchanged from its base. No model in the comparison passes 5.
• Multi-step research. BrowseComp-Plus is 4.5, the best of the five and still close to zero. Complex information-retrieval scenarios score 15.
• Knowledge. GPQA Diamond is 38.1 against 59.9 for Qwen3.5-2B. Tether says the model should call tools for facts, not answer from memory.
• Safety. In Tether's automated red-teaming the model gave unsafe responses 21% of the time, and 31% under prompt-injection attacks. The base model's figures are 36% and 46%.

The limitations section says the model is not intended for safety-critical or high-stakes automated decisions, or for applications involving money, personal information or health. That sits awkwardly beside the email, calendar and finance tasks the post lists as target domains.

How it was trained

Four stages, per the post: supervised fine-tuning on tool-use conversations; distillation from three Qwen3.5-4B teacher models, one each for tool use, reasoning and instruction following; then two rounds of reinforcement learning with rule-based rewards. Tether reports that the first stage raised tool use but pushed knowledge and instruction following below the base model, and that distillation recovered both. That is a useful warning for anyone fine-tuning a small model on tool traces alone.

What is missing

• Speed and memory on a device. The post argues that shorter reasoning "should reduce latency on-device" but gives no tokens per second, memory use or battery figure on any phone or laptop.
• Quantised accuracy. Tether served the BF16 weights with vLLM for its tests. There are no scores for the Q4 or Q8 files, which are the ones a phone would run.
• Independent results. Tether chose the benchmarks and the four comparison models, and none is larger than 2 billion parameters.
• Context length. Neither the post nor the model card states a supported context window.

If you are testing small models for a tool-calling step, the practical use is narrow: one-shot calls with a fixed tool list, measured on your own traces with the quantised file you would ship.