TECHValidatedⓘEvidence tiers: Early signal = 1 uncorroborated request — treat as radar, not a validated opportunity · Corroborated = 2-3 requests · Validated = 4+ requests from 2+ different people. Every count appears exactly once, in the fact bar below — each traces to a real source post.
"I'm just looking for real performance numbers: * tok/s at 256K / 262K * stack used: llama.cpp / vLLM * ROCm version / patches, if relevant I’ve seen \~10 tok/s at 256K for AWQ/FP8, but I’m trying to know whether Q6 + MTP changes that"
— a builder, verbatim · sign in free to trace the source ↗
What they tried & dropped
Quoted from the posts below — each quote is checked word-for-word against the stored post.
nomic 1.5 — "results weren't all that satisfactory"verbatimsign in to trace ↗
llama.cpp — "stick with llama.cpp for running it or not"verbatimsign in to trace ↗
Llama.cpp Tensor Parallelism RPC — "scaling down to about 17 tps by 100k context"verbatimsign in to trace ↗
15 rivals shipping (fact bar)how verified → market is crowded; differentiate on the unmet requirements above, don't compete on breadth.
22 builder posts asking · 22 different people (fact bar)see posts → demand is independently reported, not one person repeating themselves.
12 new tracked posts (all types) in the last 30 days vs 14 the month beforesee the curve → demand is cooling.
22 demand signals vs 15 shipping rivals → demand and supply are balanced — differentiation decides.
0 willingness-to-pay quotes yet → no one has named a price — validate pricing before building.
last new signal 1 day ago → conversation is live now.
Signal history
62 tracked posts (22 asks · 40 product/news)first post Apr 2026▲ 12 in the last 30 days
Cumulative tracked posts (asks + products + news) by original post date · GapMine tracked communities — not search volume · every step traces to a real post in the evidence record (sign in to read them).
Evidence pool
Top 10 of the 62 tracked posts, picked as evidence — each insight links back to a real one.