Ant's Falcon TST 2.0 tops GIFT-Eval as global banks adopt the model
Ant International's Falcon TST 2.0 claims the top spot on GIFT-Eval, the most-watched benchmark for time-series foundation models, and it got there by chasing a harder prize than leaderboard points: forecasts that hold up in live financial markets. The Singapore-based payments arm of Ant Group says version 2.0 of its Time-Series Transformer cut mean absolute scaled error to 0.666 across the benchmark's 28 datasets, ahead of models from several global tech giants, and that forecast accuracy has held above 93% in production use. The company built the model for its own problem — predicting cash flows and foreign-exchange exposure across borders — and Barclays, Citi, Deutsche Bank and Standard Chartered have since put it to work on exactly those desks. Ant's roadmap pushes beyond finance next, into e-commerce supply chains and airline operations.
The interesting part is the order of operations. Most time-series foundation models — Amazon's Chronos-2, Google's TimesFM 2.5, Salesforce's Moirai 2.0, IBM's FlowState — shipped as public benchmarks first and hunted for business cases after. Falcon ran the loop in reverse: it spent 2025 proving itself inside real money-management workflows at Barclays' FX hedging platform and Standard Chartered's liquidity engine before opening weights on Hugging Face, then spent 2026 rebuilding for scale — an encoder-only Falcon-2.0 for fast single-series prediction with quantile outputs, and a separate Falcon-X for modeling how exchange rates, rates and volatility move together.
That path matters because finance is where general benchmarks go to be humbled. Independent studies over the past year have found off-the-shelf time-series models often barely beat a random walk on market data, and that gains evaporate when unrelated variables get mixed into multivariate inputs. Ant is effectively betting that domain-specific validation — not parameter count or a single SOTA snapshot — is what converts a forecasting model into infrastructure. One caveat: the QbitAI piece carrying the announcement was supplied by Ant International itself, so treat the headline numbers as the company's own accounting until independent replications land.
A small team hand-writing GPU assembly for AMD chips just out-benchmarked stock vLLM by more than 2x — and open-sourced every line.
Netra Runtime's open-source kernel stack pushed Qwen3.6-35B-A3B-FP8 to 81,331 output tokens per second aggregate across eight AMD Instinct MI350X GPUs — roughly 2.16 times what stock vLLM delivers on the same hardware. A single MI350X hit 11,161 tokens per second, and the team says raw compute on AMD silicon actually exceeds Nvidia's B200; the gap has always been software. So they went below ROCm's comfort zone, writing kernels in raw GCN assembly for the MoE routing, FP8 attention and recurrent paths where generic libraries leave performance on the table.
Why this matters: AMD's MI350X packs 288 GB of HBM3E per card — enough to hold the entire 35B-A3B model, KV cache and serving state on one GPU — yet almost nobody runs inference on it because CUDA tooling remains years ahead. Netra's results are self-reported benchmarks rather than third-party audited numbers, so apply the usual discount, but the code is MIT-licensed and reproducible, which is more than most vendor benchmarks can say. If hand-written assembly can close even part of the CUDA–ROCm gap, hyperscalers starved for Nvidia supply gain a credible second source — and we covered the flip side earlier this week: Qwen3.8-27B ships a fully NVFP4 build that keeps its benchmark scores at a third of the size, showing the open-model world squeezing efficiency from both the hardware and the weights at once.
China's state-backed chip champion YMTC told investors during IPO preparations that it intends to become the world's largest NAND flash supplier by the end of 2027, per Financial Times sources — a $5 billion listing that doubles as a declaration of war on Samsung and SK Hynix.
YMTC's parent filed this week to raise about $4.9 billion (RMB 35 billion) on Shanghai's STAR Market, which would make it China's biggest chip IPO and one of the largest tech listings globally this year. The ambition inside the roadshow is bigger than the raise: becoming the top NAND supplier within roughly 16 months. That would require leapfrogging the Korean giants that have dominated flash memory for decades — no small claim for a company still operating under US sanctions, though Beijing's massive AI-driven demand for storage gives it a home market big enough to matter.
For AI watchers, the memory angle is the story. Every data-center GPU deployment leans on NAND for the storage tiers feeding model training and inference, and the same US export-control regime that gates advanced logic chips has pushed China to pour state capital into domestic memory. An IPO of this size funds exactly that capacity expansion. We've tracked the pattern all month — edge-AI firm ModelBest also filed for an A-share IPO at ¥20B — and YMTC is the heaviest-weighted entry yet.
What to watch: whether YMTC's STAR Market pricing holds up post-listing — a strong debut hands Beijing a template for funding more sanctioned-sector champions.
Which of these three moves reshapes its field fastest — bank-grade forecasting, AMD's software catch-up, or China's memory push? Tell us in the comments.
Sources: Netra Kernel (GitHub)
