Tether Releases BotPsy, a 0.8B On-Device Agent Model: Vendor Reports Doubled Tool-Calling Scores, Apache 2.0 Weights Labeled for Research and Education
According to a Hugging Face blog post published November 11, 2026, Tether AI Research released BotPsy, a 0.8B model post-trained from Qwen3.5-0.8B and aimed at function calling and agentic workflows for on-device deployment. The vendor reports it leads on five of six tool-use benchmarks, including over Qwen3.5-2B-Think, which has about 2.5 times the parameters, and says the weights will be open under Apache 2.0, restricted to research and educational use.

What changed
According to a Hugging Face blog post published November 11, 2026, Tether AI Research released BotPsy, a 0.8B-parameter model post-trained from Qwen3.5-0.8B. It targets function calling and agentic workflows for on-device deployment, with scenarios covering web search and research, travel and daily life, business and finance, device and vehicle control, files and documents, calendars, email and messaging. The vendor reports roughly doubling the base model's agentic ability: BFCL v4 from 23.4 to 44.0, ACEBench from 30.1 to 58.7, and τ-bench from 23.2 to 51.5. It scores higher than Qwen3.5-2B-Think, which has 2.5 times the parameters, on BFCL v4, ACEBench, τ-bench, ToolSandbox and BrowseComp-Plus, and trails only on Claw-Eval (46.5 versus 48.9). On multi-turn performance, BFCL v4 multi-turn rises from 7.5 to 46.1, and τ²-bench telecom scores 82.5; the BFCL hallucination test scores 86.1, meaning the model more often declines to call when no suitable tool exists. The vendor also lists weak spots: open-ended retrieval BrowseComp-Plus scores only 4.5, BFCL web search 14.0 and memory 16.3, all tested models score below 5 on τ³-banking, and complex multi-step information retrieval scenarios score only 15. Training is a four-stage post-training process: supervised fine-tuning, multi-teacher on-policy distillation (three Qwen3.5-4B expert teachers specializing in tool use, general reasoning and instruction following, respectively), and two rounds of DAPO reinforcement learning driven mainly by rule and verifier rewards, where the tool-calling reward gives partial credit for tool names, parameter names and parameter values. The publisher says it will offer three precisions, bf16, Q8 and Q4, released as Hugging Face safetensors and as GGUF for use with the QVAC SDK and llama.cpp, with weights open under the Apache 2.0 license for research and educational use. All benchmarks above are vendor-reported; external API benchmarks were run on November 3, 2026, and all models were evaluated in thinking mode.
What it means for developers
From an integration standpoint, the noteworthy point is less that “a 0.8B model can do more work” than that tool calling itself is treated as the core post-training objective for a small model. According to the post, the first reinforcement-learning round of the four stages centers on tool calling; the reward first checks reasoning and call format, then gives partial credit for tool names, parameter names and parameter values rather than a binary correct-or-wrong reward. This pushes errors from the format layer to the parameter layer: when the model emits a structurally valid call with a deviating parameter value, your tool layer absorbs the problem, so parameter validation, idempotency and rollback design need to be in place beforehand. The deployment path is relatively familiar. The publisher provides two repository types, safetensors and GGUF; the GGUF targets the QVAC SDK and llama.cpp, meaning it can plug into an existing local inference stack rather than binding you to a proprietary runtime; the three precision tiers also leave room to trade off memory, latency and format stability. The capability profile directly shapes product boundaries. The publisher states explicitly that the model is designed to “call a tool when it doesn't know” and that general knowledge still scales mainly with parameter count: it leads its size class on GPQA Diamond at 38.1 and MMLU-Pro at 48.9, but falls well short of the 2.5x Qwen3.5-2B-Think (59.9, 65.5), so retrieval and tool-return quality matter more to final answers than the model's own knowledge. Multi-turn gains are the largest, but Multi-IF drops to 51.0 by the third round of accumulated instructions, and low-resource languages (Hindi 24.5, Russian 57.6) clearly trail English, Chinese, French, Spanish, Italian and Portuguese, which sit in the 72–80 range. This means state management, instruction restatement and context compression on-device must be handled by the application side; you cannot expect the model to stay stable on its own across long sessions. On safety, red-teaming reports 21% unsafe responses (36% for the base model) and 31% under prompt injection (46% for the base). The remaining share means an external policy layer is still needed when reading untrusted tool output.
Integration and evaluation
The following are conditional recommendations and analytical inferences, not statements from the publisher. 1. Do traffic splitting first, not wholesale replacement. The post shows single calls scoring generally 70–79 across application groups, complex scenarios overall only 39.5, and multi-step information retrieval just 15. On that basis you can draw a line: single or parallel calls with clear tool schemas and parameters that upstream can fill go to the on-device model; requests needing multi-round exploration or cross-source verification still go to the cloud. The trigger is that single calls make up a high enough share of your historical sessions and parameter validation can be externalized outside the tool layer; if not, the model's parameter mismatches will hit business systems directly. Evaluate with two metrics sampled from your own traffic: exact-match rate for function names and parameter values, and the share of refusals when the given tool set cannot satisfy the request, the latter corresponding to the 86.1 on the BFCL hallucination test, though that is a vendor-configured number and must be self-tested. 2. Start with Q8, then try Q4. The publisher offers bf16, Q8 and Q4 precisions, and the choice depends on the device's VRAM or memory budget. A conditional approach: use Q8 to establish a format-validity baseline when memory is not tight, and drop to Q4 only when memory is constrained and the decline in Q4's format-validity rate is acceptable. What you need to measure yourself is the tool-call format-validity rate on a fixed prompt set, the proportion of repeated reasoning loops, and per-call latency. The post reports loop rate dropping from 22.9% to 1.4% and runs hitting the 32k limit dropping from 28.3% to 0.2%; these are measurements on its released weights, and whether they hold after quantization needs verification. 3. Keep safety judgment outside the model. The post says the red-team unsafe-response rate fell from 36% for the base model to 21%, and under prompt injection from 46% to 31%. 31% means a substantial share of injection samples may still yield unsafe output, which is not enough to hand safety responsibility to the model for agents that need to read untrusted content such as web pages and email. This applies when you let it call tools that change state (payments, sending messages, deleting files); the practice is to add allowlists and human confirmation for those tools. Measurable metrics are the share of write operations that pass through an external policy gate and the share of injection test cases the gate blocks. 4. Route by language and turn count. The post gives Multi-IF scores of 72–80 for English, Chinese, French, Spanish, Italian and Portuguese, versus 24.5 for Hindi and 57.6 for Russian, with instruction following dropping to 51.0 by the third round. If your users include these low-resource languages, or a single session needs three or more rounds of accumulated instructions, the precondition is not to send that traffic to the model before doing instruction restatement and compression at the prompt layer. Measure per-language instruction-following pass rates and average conversation turns.
Limits and open questions
All the benchmarks above are vendor-reported, completed under the vendor's own setup, with comparison models also run by the same team; there are no independent reproductions at this time. Among the evaluation conditions, all models ran in thinking mode, and benchmarks using external APIs were executed on November 3, 2026; reasoning tokens are counted by each model's own tokenizer, so token counts across models are not quite the same yardstick, and the vendor's claim of “1.4–3.1x fewer tokens” should be read under that convention. The “best among ≤1.2B” conclusion is also limited to the comparison set it selected. License and use need to be considered separately. The post says the weights are open under Apache 2.0, while also stating “for research and educational use”; the intended-use section likewise limits applicable scenarios to research and education and states it is not suitable for safety-critical or high-risk automated decision-making, or scenarios involving money, personal information or health. Weight availability does not mean all these uses are authorized; deployers should check the license file and intended-use statement in the Hugging Face repository themselves rather than inferring the scope of rights from the availability of the weights. One wording point also deserves attention: the body describes the weight release in the future tense in several places while also giving safetensors and GGUF repositories, and this brief cannot verify whether those files are currently downloadable or match the article's description. On capability boundaries, open-ended retrieval is a clear weakness: BrowseComp-Plus is self-reported at 4.5, BFCL web search 14.0 and memory 16.3; on τ³-banking all compared models score below 5, so that scenario is not currently usable. The model is trained and tested mainly in English, and English is the recommended language. KimiSeek's current model catalog has no entry corresponding to BotPsy-0.8B, and this piece's information on capabilities and distribution all comes from the publisher; the post also does not mention an officially hosted API, and the weight distribution channels follow the vendor's statements on the Hugging Face, QVAC SDK and llama.cpp side. If hosted services or third-party integrations appear later, their availability, terms and billing will be described by the vendor or the relevant service provider.
Illustrations from the original report







