The technical thesis: DeepSeek V4-Flash production release is the same 284B-MoE architecture as the April preview. Zero architectural changes. The 6.5× DeepSWE improvement came entirely from post-training. This article explains how — and what it means for model evaluation going forward.
1. Architecture: What’s under the hood
Mixture of Experts, 284B total, 13B active
V4-Flash uses a sparse MoE architecture. The full model contains 284 billion parameters distributed across multiple expert sub-networks. Each forward pass activates only 13 billion — roughly 4.6% of the total. This is the fundamental tradeoff: train a massive model, but inference at the cost of a much smaller one.
The expert routing is learned during pre-training. Each token is assigned to a small subset of experts based on the input representation. The router is trained jointly with the experts, so specialization emerges naturally rather than being pre-specified.
CSA + HCA: Hybrid Sparse Attention
The attention mechanism combines two sparse patterns:
-
CSA (Chunked Sparse Attention): The sequence is divided into fixed-size chunks. Each token attends to all tokens within its own chunk plus a sparse selection of tokens from previous chunks. This provides local coherence without quadratic scaling.
-
HCA (Hierarchical Compressed Attention): Long-range dependencies are handled through a compressed representation. Key-value pairs from distant chunks are aggregated into a smaller set of summary vectors, which every token can attend to. This is what enables the 1M-token context window without the memory explosion of full attention.
The combination is what makes the 1M-token context window practical: at 1M tokens, DeepSeek reports that KV cache memory is only 10% of what V3.2 required, and inference compute is 27% of V3.2’s.
1M-token native context
The context window is 1 million tokens natively — not through position interpolation or RoPE extension tricks. This means the attention patterns were trained at this length from the start, which matters for long-document retrieval accuracy. DeepSeek claims 97% retrieval accuracy on million-token documents.
2. The benchmark story: What changed and what didn’t
The pre-training didn’t change
DeepSeek explicitly stated: “模型结构、尺寸和预览版完全一致，仅重新进行了后训练” — the model structure and size are completely identical to the preview; only post-training was redone.
This means the pre-training corpus, the expert routing, the attention patterns, and the base knowledge are all from the original April checkpoint. The weights that encode factual knowledge, language understanding, and code syntax are unchanged.
What post-training targeted
The post-training phase for V4-Flash production release focused on three areas:
-
Agent task completion. Multi-step autonomous workflows where the model plans, executes, observes results, and adjusts. This is what DeepSWE and Terminal Bench measure.
-
Tool calling fidelity. Correct parameter formatting, proper tool selection, and error recovery across chained API calls. This is what Toolathlon measures.
-
Long-context instruction following. Maintaining task coherence across 100K+ token contexts. This is the retrieval accuracy and long-document benchmark improvements.
The numbers, with methodology notes
| Benchmark | Flash production release | Flash Preview | Pro Preview | Notes |
|---|---|---|---|---|
| Terminal Bench 2.1 | 82.7 | — | 67.9 (TB 2.0) | Different test set versions; not directly comparable |
| DeepSWE | 54.4 | 7.3 | — | Same test set; 6.5× improvement is real |
| Cybergym | 76.7 | — | — | Security automation; new benchmark for this release |
| Toolathlon Verified | 70.3 | — | — | Tool calling; verified subset |
| NL2Repo | 54.2 | — | — | Repository-level code understanding |
| DSBench-FullStack | 68.7 | — | — | Internal DeepSeek benchmark |
| DSBench-Hard | 59.6 | — | — | Internal DeepSeek benchmark |
The DeepSWE comparison is the cleanest signal: same test set, same architecture, 6.5× improvement. The Terminal Bench comparison uses different test set versions (2.1 vs 2.0), so the 82.7 vs 67.9 gap should be interpreted directionally, not as a precise magnitude.
3. Why this matters for model evaluation
The parameter-count heuristic is dead
A 13B-active model (V4-Flash) now outperforms a 744B model (GLM-5.2) on agent tasks. A 13B-active model is within 2.3 points of a model estimated at 300-400B active (Claude Opus 4.6). The correlation between parameter count and capability was never perfect, but it’s now broken in a way that can’t be ignored.
Post-training is the new frontier
If pre-training determines what a model knows, post-training determines what it can do with that knowledge. The V4-Flash result suggests that for agent tasks specifically, post-training quality dominates pre-training scale. A well-trained 13B model beats a poorly-trained 1.6T model on the tasks that matter for autonomous coding.
Benchmark inflation is real and accelerating
V4-Flash production release scores 82.7 on Terminal Bench 2.1. Three months ago, 67.9 was impressive. The ceiling is moving fast, and models that were “state of the art” in April are now below the new baseline. Teams need continuous evaluation pipelines, not point-in-time comparisons.
4. The Harness: Infrastructure, not just a model
DeepSeek’s Agent Harness framework was formally named in this release. The team, led by former Jane Street quant Cui Tianyi, was assembled in March 2026. The Harness provides:
- Long-horizon task planning with iterative self-checking
- Automatic retry on tool-call failures
- State management across multi-step workflows
- Integration with the Responses API format
This is significant because it means DeepSeek isn’t just shipping a model — they’re shipping the agent runtime. The model + Harness combination is the product, not just the model weights.
5. What to test
If you’re evaluating V4-Flash for coding tasks:
- Run your own DeepSWE-equivalent. Define 10-20 bounded coding tasks with acceptance criteria. Don’t trust the benchmark; trust your workload.
- Measure tool-call chain length. If your agent workflow chains 5+ tool calls, test specifically for context drift and parameter corruption at the tail end.
- Profile the 1M-token context. The 97% retrieval accuracy claim is encouraging, but verify with your actual document types and query patterns.
- Compare cost per accepted task. V4-Flash is 1/10 the price of GPT-4o per token. If it takes 2× more retries, it’s still 5× cheaper. Measure the actual economics.
Want the wider view? The head-to-head comparison pits V4-Flash against GPT-5.6, Claude, Kimi K3, and GLM-5.2, and the post-training analysis explains why training methodology now matters more than parameter count. For the complete launch picture including what Pro changes, start from the full launch briefing.
You’ve read the architecture. Now run the model.
MonkeyCode already supports V4-Flash-0731. Same-day integration, 30M free tokens daily, no hardware to provision.