FLAT is Flash Linear Attention on TPUs. On one TPU v6e, forward + backward runs up to 3× faster than FLA on an H100 for GLA and 1.5× faster for Mamba-2.
Key points
Benchmark: 3× speedup over FLA on H100 for GLA on a TPU v6e.
Benchmark: 1.5× speedup over FLA on H100 for Mamba-2 on a TPU v6e.
Sources
DuoDiscover benchmarks show up to 2.68× TPU kernel speedup
DuoDiscover is announced with performance results, achieving up to 2.68× faster than production TPU kernels. It also ranks #1 on NVIDIA's SOL-ExecBench and shows gains on vLLM and FreeToken.
Key points
1 on NVIDIA's SOL-ExecBench
1.92× faster on vLLM
1.44× faster on FreeToken
Beats the best human and AI-designed CPU prefetchers
Sources
DeepSeek V4.1 Flash 425x cheaper than Opus 5.5 at 200 TPS/user
DeepSeek V4.1 Flash on @vllm_project is claimed to be 425× cheaper than Opus 5.5 and runs at 200 TPS per user. It is reported to cost about 900 M tokens per dollar, assuming the same cache hit rate and token mix. A quoted tweet adds that a GB300 NVL72 inference optimization gives 7.2× better performance per TCO than other AI chips.
Key points
Cost: 425× cheaper than Opus 5.5, ~900 M tokens per $
Throughput: 200 TPS per user on @vllm_project
Optimization claim: GB300 NVL72 gives 7.2× better perf per TCO than other AI chips
Sources
DuoDiscover full-stack code optimization agent benchmark results
DuoDiscover is a full-stack code optimization agent that co-discovers optimization targets and a cheap proxy evaluator. It achieved #1 on NVIDIA's SOL-ExecBench and reported speedups of 1.92× on vLLM, 1.44× on FreeToken, and 2.68× over production TPU kernels. The post also claims it beats the best human and AI-designed CPU prefetchers.
Key points
Benchmark: 1 on NVIDIA's SOL-ExecBench
Performance: 2.68× over production TPU kernels
Sources
DuoDiscover AI agents ship PRs, beat hardware baselines
DuoDiscover AI agents ship real PRs and outperform human-expert hardware baselines. They merged optimizations into main branches and achieved a 2.68× TPU kernel speedup.
Key points
Real optimizations merged into main branches (e.g., FreeToken #619)
First superhuman CPU prefetcher: +3.93% | +6.31% overall, outperforming AlphaEvolve
Sources
Dimuon optimizer reaches lower loss faster than AdamW
Dimuon is a new optimizer built for LoRA fine-tuning. In SFT tests it often reaches lower loss in fewer steps than AdamW. The guide is available for tinkerers.
Key points
Optimizer: Dimuon, designed for LoRA fine-tuning.
Claim: In SFT tests it often reaches lower loss in fewer steps than AdamW.