Overview
This week saw a wave of domestic LLM releases under a clear theme: fully domestic compute and extreme cost-performance. On October 22, iFLYTEK launched Spark X2.6 at its 1024 Developer Festival—a 293B-A30B MoE with a 1-million-token context, claimed as China’s first model to achieve coding and agent breakthroughs trained entirely on domestic compute. Dynamic parallel training lifts efficiency ~4x and a four-tier KV Cache distributed inference ~4.5x.
On October 23, MiniMax open-sourced M3.1 Flash, processing over 100 trillion tokens within 72 hours of launch, cutting cost 70% and lifting inference speed ~3x; Morgan Stanley called its pricing highly competitive among Chinese Flash models. DeepSeek shipped V4.1-Flash-1017 under the MIT license, cutting average per-task cost ~60%, and open-sourced DeepInfer, an inference framework compatible with both Nvidia and Ascend.
The competitive focus has shifted from “parameters and benchmarks” to “domestic compute + extreme cost-performance + device-cloud synergy.”
Background & Interpretation
1. Technical Background and Evolution
LLM competition has entered engineering depth. Spark X2.6 tackles domestic accelerators’ memory-bandwidth limits with dynamic parallel training and four-tier KV Cache distributed inference, validating the feasibility of training/inferring large models on domestic chips. The open-source on-device X2.6-4B splits work “60% on-device, 30% device-cloud, 10% cloud,” balancing data security and token cost. MiniMax and DeepSeek use sparse attention, quantization-aware training and open weights to keep pushing token-unit and per-task costs down.
2. Core Drivers and Underlying Mechanism
“Fully domestic compute” is both a constraint and a differentiator. With high-end GPUs restricted, domestic accelerators’ bandwidth and interconnect gaps force algorithmic innovation (long-sequence parallelism, heterogeneous distributed inference). “Device-cloud synergy” keeps ~70% of tasks on-device, cutting cloud cost and privacy risk. “Extreme cost-performance” via open-source and dual-track interfaces reshapes industry pricing and expands enterprise adoption.
3. Market Structure and Industry Chain Impact
Huawei Ascend, Moore Threads, Nvidia, AMD and multiple inference clouds connected to MiniMax M3.1, signaling faster interoperability of the domestic-compute ecosystem. DeepSeek’s rumored ~$75B valuation and 2027 IPO show capital’s faith in the open-source route. Competition moves from “whose model is stronger” to “whose unit intelligence is cheaper and more deployable,” benefiting inference clouds, on-device chips and the application layer.
Implications & Outlook
Near term, domestic-compute LLMs will accelerate into government, finance and manufacturing (trustworthy-computing) scenarios. Medium term, device-cloud synergy and cost-performance shift AI from “centralized training” to “distributed inference,” raising edge/on-device chip demand. Long term, open-source and low pricing reshape the profit pool; the application layer and embodied AI gain a scale-deployment window as “brain” costs fall.
Takeaways for Industry Participants
Model vendors: build moats on domestic compute and cost-performance; prize device-cloud synergy and open-source operations. Compute and chip firms: adapt mainstream open frameworks (e.g., DeepInfer); close bandwidth and interconnect gaps. Enterprise users: evaluate on-device/cloud trade-offs for compliance and cost; prefer privately deployable options. Developers: exploit low token prices and open weights; watch the inference-cost curve.
This column compiles industry information and shares technical perspectives; it does not constitute investment advice.