DeepSeek V4 Flash Ships: 13B Active Parameters, 6x Agent Jump, and a Price Floor of $0.003/M Tokens
A 13B-active-parameter model now competes with far larger systems on agentic coding benchmarks while costing roughly 1/90th of Claude. The 6x DeepSWE gain from post-training alone suggests that architecture changes are not the only path to capability jumps, which has direct implications for how teams budget training versus inference spend.
DeepSeek quietly released the official version of its V4 Flash model, a 284B total-parameter MoE model that activates only 13B parameters. The architecture and base model are unchanged from the preview; the leap comes entirely from a new post-training regimen. On the DeepSWE benchmark, which measures autonomous, long-horizon coding tasks, the score vaulted from 7.3 to 54.4. Terminal Bench 2.1 hit 82.7, surpassing the earlier V4 Pro preview and landing just behind Opus 4.8.
The pricing remains aggressive: 0.02 yuan per million input tokens on a cache hit, with a 98% cache-hit discount that far exceeds the industry's typical 90%. Developers report cutting per-request costs by 85% after restructuring their caching strategy. The model supports a 1M-token context window, a 384K-token max output, and natively speaks both OpenAI and Anthropic API formats.
NetEase Youdao has already integrated the model across its entire AI product line, and community developers are pairing it with Claude Code for terminal-native coding agents. DeepSeek says the same post-training approach will be applied to V4 Pro, whose official release is still pending.
The 6x DeepSWE jump without any architecture change challenges the assumption that capability gains require scaling parameters or redesigning the model. Post-training alone can unlock agentic behaviors that were latent in the base model.
DeepSeek's 98% cache-hit discount is an operational moat as much as a pricing one. It rewards developers who design cache-friendly workflows, effectively locking them into the platform through cost optimization rather than API compatibility alone.
Publishing a 13B-active-parameter model that rivals systems with far higher active counts resets expectations for what 'small' models can do in agentic tasks, which may pressure other labs to justify their larger, more expensive inference footprints.