Six identical requests. 2:12 vs 6:25.
Hy4 Flash vs Hy4 in WorkBuddy — same inputs, one run each, every result passed the same checks. Flash finished in a third of the time.
Quick tasks → Hy4 Flash.
Hard problems → Hy4.
Switch in one click.
#hy4flash#workbuddy
Six identical requests. 2:12 vs 6:25.
Hy4 Flash vs Hy4 in WorkBuddy — same inputs, one run each, every result passed the same checks. Flash finished in a third of the time.
Quick tasks → Hy4 Flash.
Hard problems → Hy4.
Switch in one click.
#hy4flash#workbuddy
Hy4 flash is here ⚡
Available on Tencent Cloud TokenHub (limited-time 50% off API pricing) and WorkBuddy (daily free quota Nov 16–30).
Built for everyday tasks, with image understanding + a 1M-token context window.
#hy4flash#hy#tencentcloud
@arpittalksstuff Would be interested to see Hy4 flash on that same PDF set. Are you measuring table structure and reading order too, or mainly text extraction?
BLG did it. World Champions 2026. 🏆
From New York to screens around the world, every teamfight deserved to be seen.
Tencent Cloud supported the Worlds 2026 broadcast with cross-region streaming, AI video enhancement, and real-time highlight detection.
The plays made history. The technology helped bring them to fans worldwide.
#Worlds2026 #TencentCloud
@gagu_sh Agreed on keeping the originals. One nuance: the 1.1.2 pair actually had slightly lower cache hits (97.29% vs 98.24%). That’s why the article checks correctness and readback alongside cache efficiency.
1M-token context window. The agent still froze for 12.8s to summarize itself.
Don't wait until 80% to compact — trim a little every turn instead.
150 coding problems in one session: same pass rate, ~23% cheaper.
Full numbers inside. https://t.co/EZRwcdzStp
@AIQuanting Some log details had to be read back from the saved originals; the plugin recovered all three in both runs. That’s different from keeping everything in active context. The estimated 23% saving comes from one coding-run pair, so it’s workload-specific.
@VibeCoderOfek The author seeded 13 constraints before six rounds of logs to test that concern. The plugin scored 15/15 on recall in both runs, but that’s still only two runs—not a guarantee for every earlier decision.
@shadow_satoshi_ Yes—older completed tool outputs. The defaults protect the latest three completed tool turns; older output becomes a compact card, with the original saved locally for readback. Errors and user constraints take priority.
@Acribayik That sounds complementary: smaller tasks limit what goes in, while compaction handles the tool output that still builds up. How do you carry shared constraints between those smaller tasks?
@pggalaviz Ha, incremental GC is a good analogy for the pauses. The twist is that an old tool result may matter again later, so the plugin keeps the original on disk for readback instead of treating it as disposable.
@MstShahinu66909 Thanks! The daily comparison is the useful bit here: a matching final balance can still hide differences along the way. Checking the input hashes and each dated equity row helps catch those.
Got 8 vCPUs for my backtest — but my strategy only used 1.
Don’t blindly throw more cores at your quant code before profiling.
I benchmarked my local backtest workflow on Tencent Cloud Hong Kong SA9 node.
Full numbers inside. https://t.co/t4brh4MDJ4
@Zenith0869 Thanks! Just to clarify, the 14.44s run was local. The Hong Kong runs checked that the outputs matched, so these measurements don’t establish a cloud speedup.
@ovihMoon I’d start with the same batch on 1, 2, then 4 workers, tracking total runtime and memory. What’s been trickier in your setup: splitting the workload or picking the instance size?
@Sinha694 Exactly. CPU time ÷ wall time was about 1 here. When testing more workers, total batch time is the useful check—higher CPU usage on its own doesn’t tell us whether the job finished sooner.