Tether releases BotPsy-0.8B open weights for on-device tool calling

What Tether announced

Tether AI Research announced BotPsy on November 11, 2026: a 0.8B-class model for function calling and agent workflows, built on Qwen3.5-0.8B. Builders can find source-format weights at qvac/BotPsy-0.8B and GGUF exports at qvac/BotPsy-0.8B-GGUFs. Both repositories are public and ungated in the Hugging Face API records we retrieved on November 12. The source repository’s LICENSE contains Apache 2.0 terms.

This is a downloadable model release, not a newly priced managed API or a complete personal-assistant application. Tether positions it for low- to medium-difficulty tasks on local devices. The release changes the choice for teams that want a small tool-calling model they can serve themselves rather than sending every request to a frontier cloud model.

There is a wording mismatch worth making explicit. The announcement’s deployment section says the weights “will be released,” but the linked model cards and public API listings already enumerate model.safetensors and three GGUF files. Our availability statement rests on those current artifact listings. We did not download or execute the binaries, so this is not a claim that every runtime or device works.

What is available now

The GGUF card lists botpsy-0.8b-bf16.gguf at about 1.52 GB, botpsy-0.8b-q8_0.gguf at about 0.81 GB, and botpsy-0.8b-q4_k_m-imat.gguf at about 0.53 GB. Q8_0 is the vendor’s recommended compromise. Q4_K_M uses importance-matrix calibration and trades some reported accuracy for a smaller file. These are approximate vendor-listed download sizes, not measured RAM budgets.

The source card describes text-only use of the Qwen3.5 architecture and recommends English. Do not read the underlying architecture’s multimodal tags as evidence that BotPsy has been evaluated as a vision assistant. The card documents Transformers, vLLM, llama.cpp and the QVAC SDK, with Transformers 5.19.0 and QVAC SDK 0.23.0 named as tested versions by the vendor.

The weights are labeled Apache 2.0, while the intended-use section describes research and educational purposes and warns against sensitive applications. Those are different kinds of statements: a license grant does not establish product fitness or safety. Training code and training data are not included in this release, although the card identifies training datasets and describes supervised fine-tuning, multi-teacher on-policy distillation and two reinforcement-learning stages.

The benchmark gain is specific, not universal

Tether reports BFCL v4 rising from 23.4 for Qwen3.5-0.8B-Think to 44.0 for BotPsy, ACEBench from 30.1 to 58.7, and its reported τ-bench aggregate from 23.2 to 51.5. The latter combines retail, telecom and banking evaluations and excludes airline. All five comparison models were evaluated in thinking mode; the source BF16 checkpoint was served with vLLM. These are vendor-reported results, not independent AccessAllGPT tests.

The limits matter as much as the gains. BrowseComp-Plus is 4.5, BFCL web search is 14.0 and BFCL memory is 16.3. The larger Qwen3.5-2B-Think still leads on Claw-Eval, at 48.9 versus 46.5. A model that improves bounded tool use is not automatically a capable open-ended research agent or a replacement for a larger knowledge model.

The quantized card adds useful detail: Q8_0 averages 50.7 across its ten reported benchmarks versus 50.4 for BF16 GGUF, while Q4_K_M averages 48.9. But ToolSandbox was not run on the GGUFs, some run counts differ, and external-API evaluations occurred on different days from the vLLM reference. These aggregates are not a controlled estimate of a universal quantization penalty.

The GGUF evaluations used NVIDIA H100 GPUs with llama.cpp, not a phone. They therefore do not prove a specific handset’s response time, battery consumption or usable context size. Downloading a half-gigabyte file is a promising starting point for an edge experiment, not evidence that a full agent application fits in that much memory.

Two integration details to get right

First, enable thinking explicitly. BotPsy was trained and evaluated in thinking mode, but its chat template defaults to non-thinking. The cards specify enable_thinking=True for Transformers and corresponding template settings for servers. The reported sampling configuration uses temperature 0.3, top_p 0.95, top_k 20 and presence_penalty 1.5. Match the relevant settings before comparing local behavior with the published results.

Second, use the correct tool parser. The vLLM example specifies --language-model-only, --enable-auto-tool-choice, --tool-call-parser qwen3_xml and --reasoning-parser qwen3. The llama.cpp example uses --jinja and explicit thinking template kwargs. These are vendor instructions we reviewed, not commands we executed. Parsing a proposed function call is separate from authorizing or running it.

Local inference also does not make an entire workflow private. A local model can still call a remote search service or send information through application tools. Decide which tools exist, where their arguments go and which actions require confirmation. Tether’s own model card says the model is not intended for safety-critical or high-stakes automated decisions, or applications involving money, personal information, health and other sensitive topics.

What builders should do next

Start with the vendor-recommended Q8_0 file and a sandbox containing only harmless tools: read a synthetic schedule, change a mock display setting or sort sample documents. Test missing arguments, requests that match no tool and multi-turn corrections, not just successful single calls. Keep execution deterministic and require confirmation before any real-world side effect.

Record the exact artifact revision, inference engine, thinking setting, tool parser, peak memory and task completion on the device you intend to ship. Compare against the base model and your existing managed route on the same tasks. Only a measured local result can tell you whether the smaller download yields an acceptable user experience.

Our assessment: BotPsy is worth a narrow local-agent evaluation because usable artifact formats and a concrete tool-use specialization are present. Keep a larger model or explicit fallback for open-ended research, and do not treat the published benchmark gains as permission to automate sensitive work. We have not tested BotPsy hands-on.

Disclosures

AccessAllGPT received no payment, credits or private model access for this report. Tether did not review or endorse it. AccessAllGPT Research is operated by NeuralArc, is independent, and is not affiliated with Tether or Hugging Face. Publication-wide relationships appear on the disclosures page.