• The bottom line: Wait. The tool-use gains are large, but the post gives no latency, memory or hardware figures for the on-device claim.
• BFCL v4 rises from 23.4 (Qwen3.5-0.8B-Think base) to 44.0, ahead of Qwen3.5-2B-Think at 37.7.
• The six-benchmark tool-use average is 45.0 against 38.8 for the 2B model, but the 2B model still wins on GPQA Diamond (59.9 vs 38.1) and MMLU-Pro (65.5 vs 48.9).
• Multi-step retrieval is the weak spot: BrowseComp-Plus is 4.5, and complex research scenarios score 15 versus 79 for single calls.
• The release links no paper or code, and the authors say the model is not for money, personal-data or health use.

The bottom line: Wait. BotPsy-0.8B, released by Tether AI Research, roughly doubles the tool-calling score of its Qwen3.5-0.8B base, but the announcement offers no latency, memory or device measurements. Those are the numbers an on-device team needs before shipping. The benchmark story is strong; the production story is, so far, unmeasured.

The one number that decides it

The deciding number is BFCL v4: 44.0 for BotPsy-0.8B against 23.4 for its Qwen3.5-0.8B-Think base and 37.7 for the larger Qwen3.5-2B-Think, according to the release. A 0.8B model beating a 2B sibling on function calling is a real result if it holds up. The blog reports BFCL v4 as an average over three runs, which makes the margin more credible than a single pass would.

Where it beats the bigger model, and where it does not

BotPsy-0.8B leads Qwen3.5-2B-Think on the six-benchmark tool-use average, 45.0 to 38.8, and on ToolSandbox, 64.6 to 41.2. It trails badly on knowledge and reasoning: GPQA Diamond is 38.1 against 59.9, and MMLU-Pro is 48.9 against 65.5. The authors say as much, describing a model meant to call tools for facts rather than store them.

How the gains were produced

The release describes four sequential stages: full-parameter supervised fine-tuning on tool-use conversations, on-policy distillation from three Qwen3.5-4B specialist teachers, and two rounds of DAPO reinforcement learning with rule-based rewards. The tool-call reward gives partial credit for matching tool names, argument names and values. The distillation and RL steps also shortened reasoning: IFBench looping fell from 22.9% to 1.4%.

The missing production numbers

The post reports no latency, memory, quantized-speed or hardware figures. It says shorter reasoning should reduce on-device latency, which is a projection, not a measurement. The token data supports the direction: on GPQA Diamond questions all models answered correctly, BotPsy used 3,690 reasoning tokens versus 11,563 for the base. Cost per call on an actual phone or laptop remains unpublished.

Failure modes the demo would hide

Multi-step retrieval is the clearest weakness. BrowseComp-Plus scores 4.5, and complex information-retrieval scenarios reach 15 versus 79 for single calls, per the authors. Instruction-following decays across turns, with Multi-IF turn 3 at 51.0, and Hindi scores 24.5 on Multi-IF. A single-call assistant may work; an agent chaining many lookups is where this model is likely to fail.

Safety and scope limits

The authors report a 21% unsafe-response rate in red-teaming, down from 36% for the base, and 31% under prompt injection, down from 46%. Roughly one in three injected prompts still succeeding is a high rate for an assistant with email and file access. The authors say the model is not intended for applications involving money, personal information or health, which covers much of the personal-assistant use case.

The bottom line: Wait

The bottom line is Wait. Prototype with the open weights if single-call tool use on constrained hardware is your use case, and measure latency, memory and your own tool schemas first. Skip it for multi-step research agents or anything touching money, personal data or health. Reassess when independent reproductions or a technical report with device measurements appear; the release links neither a paper nor code.