Warning: this is an incomplete benchmarking run.
I ran the v5 prompt on Ornith 1.5 35B-A3B NVFP4, tensor-split across two RTX 5060 Ti 16 GB cards. The dashboard is the recorded sample.
Representative output
Run 1 (Ornith 1.5 35B-A3B NVFP4) interleaves reasoning through the code — the lower-quant fold reads more like a design note than a finished module. Trimmed to its start_link public API:
def start_link(), do: start_link([])
def start_link(opts) when is_list(opts) do
default_ttl = parse_default_ttl(Keyword.get(opts, :default_ttl))
name = Keyword.get(opts, :name) || unique_name()
sweep_ms = Keyword.get(opts, :sweep_interval_ms, default_sweep_interval())
ensure_registered!(name)
GenServer.start_link(__MODULE__, {default_ttl, name, sweep_ms}, name: name)
end
def start_link(default_ttl, opts) when is_integer(default_ttl) and default_ttl > 0 and is_list(opts) do
start_link(Keyword.put_new(opts, :default_ttl, default_ttl))
end
I have not mined Grok Build sessions for this weight yet, so there is no field evidence on this page. The needle is a single absolute depth (8192 tokens of a 262144 window). Later runs should plant it at 10 %, 20 %, 30 % of the slot window instead.