DeepSeek V4: Fitting a Million-Token Context in Your Pocket — How Did They Do It?

Forum topic · 小凯 · 2026-04-27

Summary

DeepSeek V4 Pro introduces a 1 million-token context window while cutting KV cache memory from 83.9 GiB (V3.2 at 128K) to just 9.62 GiB. The 1.6T-parameter Mixture-of-Experts model activates only 49B parameters per forward pass and uses a CSA/HCA hybrid attention scheme—compressed KV streams, sparse attention over compressed tokens, and local sliding windows—plus FP4 indexing and FP8 caching. Released under the MIT license with a 58-page technical report covering mHC, MoE routing, and quantization details, V4 Pro ranks as the top open-weights model, behind only closed models like GPT-5.x and Claude Opus. It supports Huawei Ascend 950 + CANN on day one, signaling supply-chain diversification away from NVIDIA's CUDA ecosystem. The cheaper V4 Flash (284B total / 13B active) is priced at $0.14/$0.28 per million tokens (input/output). Caveats include verbose outputs and a reported 94% hallucination rate on AA-Omniscience, which measures whether models admit ignorance on out-of-knowledge questions. Source: easy-learn-ai commit d9b875d.

Key points

1M-token context window: DeepSeek V4 can ingest roughly the first two-thirds of a full novel and analyze it end to end—previously feasible only with expensive closed APIs.
Radical KV cache compression: KV cache was reduced from 83.9 GiB (V3.2 at 128K context) to 9.62 GiB at 1M tokens—about 8x longer context at roughly one-ninth the memory.
Efficient MoE architecture: V4 Pro has 1.6T total parameters but activates only 49B per forward pass.
CSA/HCA hybrid attention: Rather than full quadratic attention, distant context is compressed into summaries processed with sparse attention, while nearby tokens retain full-resolution local sliding-window attention. FP4 indexing + FP8 caching compress memory further.
Benchmark standing: V4 Pro holds the top position among open models, second only to closed models like GPT-5.x and Claude Opus.
True open source: MIT license (no commercial restrictions), with a 58-page technical report covering mHC (manifold-constrained hyper-connections), MoE routing, CSA/HCA math, and FP4/FP8 mixed-precision implementation.
Non-NVIDIA support: Day-one compatibility with Huawei Ascend 950 + CANN; pricing for V4 Pro is expected to drop once Ascend 950 supernodes scale in H2. Companion open-source releases include DeepEP V2 and TileKernels for communication and operator-parallel optimizations.
Aggressive pricing: V4 Flash (284B total / 13B active) costs $0.14/$0.28 per million tokens (input/output); in full reasoning mode it approaches mid-tier Pro performance.

Caveats discussed in the post

Verbosity: In one benchmark run, V4 Flash emitted 240M tokens—well above comparable models—so low unit pricing doesn't guarantee low total bills; users should explicitly request concise answers.
94% hallucination rate on AA-Omniscience: Artificial Analysis found V4 Pro Max tends to fabricate answers when asked about things outside its training data. This matters less for the model's core strength—answering questions grounded in user-provided documents within its 1M context—but factual claims from memory should be verified.

Takeaway

The post frames DeepSeek V4 as a turning point for open AI in 2026: the gap with closed frontier models is narrowing fast, and the combination of 1.6T parameters, 1M context, MIT licensing, full engineering disclosure, domestic-chip support, and near-free API pricing amounts to a complete, commercially viable, decentralized alternative. Remaining issues—hallucinations, verbosity, and real-world long-context UX—are acknowledged as open problems for the developer community to refine.

*Source: easy-learn-ai commit d9b875d*