LongCat-2.0 and DeepSeek V4-Flash: Open-Weight MoE Models Compared

Two Scaling Philosophies

LongCat-2.0 (1.6T total, ~48B active) represents the compute-heavy end of the spectrum: the publisher states the model completed full training and inference on approximately 50,000 domestic AI ASICs, with an architecture (ScMoE, LSA sparse attention, N-gram Embedding) designed for compute-constrained, non-NVIDIA hardware.

DeepSeek V4-Flash (284B total, ~13B active) represents the efficiency end: a compact MoE with 6 of 256 experts active per token, released as a post-training-only update to the V4-Flash preview — the publisher reports the architecture and parameter count were unchanged, with the Agent-capability gains (DeepSWE 7.3 to 54.4) coming from reinforcement learning on high-quality data. The model's Triton-based kernels decouple it from the NVIDIA CUDA ecosystem for inference deployment.

Pricing and Cache Economics

DeepSeek V4-Flash is roughly 2.1× cheaper on input ($0.14 vs $0.30 per M tokens) and 4.3× cheaper on output ($0.28 vs $1.20 per M tokens) at published API rates. The larger gap on output reflects the different cost structures of a 13B-active model versus a ~48B-active model.

Cache economics differ in structure rather than headline generosity: LongCat-2.0 does not bill cache hits at all, while DeepSeek V4-Flash applies a ~98% discount to cached input ($0.0028 vs $0.14 per M tokens) — an unusually deep discount versus the industry-standard 90%. For agentic workloads with heavily repeated prefixes, both models converge toward very low effective input cost; for one-shot workloads with little prefix reuse, DeepSeek's flat rates are lower.

Third-party analysis (Artificial Analysis) reports V4-Flash averaging ~$0.03 per task in weighted testing — an order of magnitude below most flagships. These figures are third-party, not independently verified by this site.

Licensing: Both MIT

Both models are released under the standard MIT License, permitting commercial use, modification, redistribution, and distillation without additional conditions. This contrasts with Kimi K3's modified MIT license and with proprietary closed models. For organizations that value license simplicity, both LongCat-2.0 and DeepSeek V4-Flash are equivalent on this dimension.

Deployment and Hardware Independence

Both models are designed to run outside the NVIDIA-only stack, but through different mechanisms. LongCat-2.0 documents GPU deployment via SGLang and vLLM plus a dedicated NPU inference path for domestic accelerators. DeepSeek V4-Flash ships Triton-based custom operators, which the publisher states enable deployment on domestic GPUs without vendor-specific adaptation and without CUDA dependency.

Both expose OpenAI-compatible and Anthropic-compatible API formats, allowing drop-in integration with the same agent tooling (Claude Code, OpenClaw, Codex-style harnesses). Both support self-hosted deployment with open weights under MIT.

Ecosystem Position

DeepSeek V4-Flash topped global token-volume leaderboards within a week of its July 31 release — third-party reports cite single days exceeding 8T tokens on one platform and its entry into the OpenRouter weekly top tier. LongCat-2.0 operates in the meituan-longcat ecosystem with adoption measured in the top tier of OpenRouter monthly call volume.

These are ecosystem observations from cited third-party sources, not a claim of product superiority by this site. Developers should evaluate both models against their own workloads, budgets, and hardware constraints.