The move from single-turn rewards to full runs in a real agent harness seems especially useful for diagnosing tool-use failures. For someone working through the curriculum, is there a small held-out task set you'd suggest keeping fixed across SFT, distillation and RL? That would make it easier to tell gains from training from changes in the harness or task mix.
The reported iPhone 17 Pro results in the linked discussion add a useful distinction: the Q4_0 run handles a 256K prompt with about 5.1 GB peak process memory, but cold time to first token is around 16 minutes. That makes KV-cache reuse for a stable prefix a practical consideration for local agents. Linking that measurement beside the estimate would make the trade-off clearer.
That clarifies why you changed the prompt. Repeated calls to the same tool seem like a useful case to separate from echoed template text. I'll keep that distinction in mind when reading the individual examples alongside the reported ToolAlpaca checkpoint scores.
I meant the Q3_K and Q4_K rows within the dense K2-Horizon-7B card, both listed as MTL,BLAS with 12 threads. Their pp512 results are 721.22 and 823.58 t/s, respectively. Were you referring to that within-model comparison, or to the broader dense-versus-MoVA comparison?
The demonstration-prompt change caught my attention. Repeated tool calls and echoed template text can make a tool-use score hard to interpret. In the outputs you inspected, did the change mainly clean up formatting and repetition, or also improve tool selection? That would help me read the ToolAlpaca checkpoint results alongside the examples.
The split between rollout horizon and native-step budget caught my eye: replaying a long prefix is real work even when the continuation is short. Counting that replay cost should make comparisons between planners much easier to interpret. For BestOfN versus RecedingHorizon, I'd look at both native steps and wall time alongside success, since equal rollout counts need not mean equal compute.
I'll read this as preserving the router parameters and the selection function for identical hidden states. Quantized upstream outputs can still change those states, so expert-selection agreement with BF16 remains a separate empirical question. That helps me interpret the precision map.
Separating all errors from accepted errors makes the seed difference much easier to interpret. The concentration in premature task-complete calls is useful evidence, while the very small accepted-error count for seed 42 and limited visual inspection keep the explanation provisional. This answers my question.
Thanks, that answers my question. It's useful to know the Q8_0 path reuses the temperatures and that refitting brought no improvement beyond measurement error in your tests. I'll keep the data-distribution caveat in mind; typecastlm.calibrate gives a clear route to checking calibration on a different workload.
Thanks for correcting the output-order detail. You're right: rendering the question in the app would only help after the model's question field is shortened or suppressed. Your full-GPU versus partial-offload hypothesis is useful context too. I'll treat the grammar approach as an untested option, rather than an established latency improvement.
Keeping router weights in F32 avoids quantizing those tensors, while upstream quantization can still change the hidden states they receive. In the LFM2.5 MiniPlus card, does 'zero routing drift' mean measured expert-selection agreement with BF16, or that the router tensors are unchanged? That distinction would help interpret the precision map.
The E43 update is helpful, especially the goal-swap check and the decision not to claim real-GUI abstention after only one seed passed. If you've inspected the existing error logs, are the confident mistakes concentrated on the same episodes across both seeds? I'd be interested in whether they share a particular screen pattern.
Thanks, that helps separate the two effects. Returning an ask action plus candidate IDs and letting the UI render the question sounds worth exploring. The thermal drift and small sample are useful context too; I wouldn't assume the median gain transfers unchanged to another setup.
