Imagine what this company will be able to do with actual funding and a more evolves Huawei stack. There's no doubt that Talent density in China is going to take a bite out of closed AI.
This on the same day that Alexander Wang pushing meta to acquire Manus AI and its Singapore washing has caught up with it.
Because of Deep Seek China is probably 6 months to a year ahead and open source AI now. Nonwithstanding what reflection AI,Tthinking Machines are doing behind closed doors with hundreds of times the capital.
Worth adding to the "American Training, Chinese Inference" section: on the day V4 released, FlagOS published benchmarks for V4-Flash running on 8 Chinese chip architectures simultaneously. Cambricon MLU, Biren, Moorethreads, Hygon DCU, Kunlunxin, others. All within hours of the public release.
The framing is accurate for the training stack. But on inference the picture is broader than just Huawei Ascend. The OpenMind community had multi-vendor inference support before V4 dropped. Which means the Ascend 950 volume question matters more for training cost competitiveness than for inference availability.
Good framing. Worth adding: on V4-Flash launch day, FlagOS had it adapted to 8 domestic chips (Ascend, Hygon, Kunlun, others). Day-0 multi-chip adaptation is new -- two years ago it took weeks. The domestic silicon stack now tracks DeepSeek releases in real time.
On developer economics: InfoQ documented ~90% cost reductions switching from frontier Western models to V4 (~3.48/M vs 30/M). The 'coding is the way forward' thesis maps to what Chinese dev forums show -- the focus is agentic coding pipelines, not chat.
Good article. Also, both poems are not bad at all.
FWIW I think kimi has the best creative writing
Right??
Interesting, thanks.
Imagine what this company will be able to do with actual funding and a more evolves Huawei stack. There's no doubt that Talent density in China is going to take a bite out of closed AI.
This on the same day that Alexander Wang pushing meta to acquire Manus AI and its Singapore washing has caught up with it.
Because of Deep Seek China is probably 6 months to a year ahead and open source AI now. Nonwithstanding what reflection AI,Tthinking Machines are doing behind closed doors with hundreds of times the capital.
Worth adding to the "American Training, Chinese Inference" section: on the day V4 released, FlagOS published benchmarks for V4-Flash running on 8 Chinese chip architectures simultaneously. Cambricon MLU, Biren, Moorethreads, Hygon DCU, Kunlunxin, others. All within hours of the public release.
The framing is accurate for the training stack. But on inference the picture is broader than just Huawei Ascend. The OpenMind community had multi-vendor inference support before V4 dropped. Which means the Ascend 950 volume question matters more for training cost competitiveness than for inference availability.
Good framing. Worth adding: on V4-Flash launch day, FlagOS had it adapted to 8 domestic chips (Ascend, Hygon, Kunlun, others). Day-0 multi-chip adaptation is new -- two years ago it took weeks. The domestic silicon stack now tracks DeepSeek releases in real time.
On developer economics: InfoQ documented ~90% cost reductions switching from frontier Western models to V4 (~3.48/M vs 30/M). The 'coding is the way forward' thesis maps to what Chinese dev forums show -- the focus is agentic coding pipelines, not chat.
Does DeepSeek V4 like to play Dungeons and Dragons?