DeepSeek V4:
flagship, flash and the open 32B

What is new in V4

The generation landed in August 2026 and immediately set two boundaries for the family. The first is a million tokens of context across all four models, including the cheap flash: a long window used to be a privilege of the senior tiers. The second is that the open line no longer trails the closed one: deepseek-v4-660b-a37b is the same flagship MoE, released as weights.

For practice the third point matters most: deepseek-v4-32b is a dense model with native vision — exactly the "compact" class you can put on your own hardware instead of pushing documents through someone else's API. It, and the earlier small sizes, get their own write-up.

The flagship sets the bar, but decisions are made where a model can actually run: on your own machine, inside your own perimeter.

How to read the names

A name is assembled from four parts. Let us use real examples from the V4 generation.

• deepseek-v4 — the generation. Within one generation the training data, tokenizer and call format are shared;
• -660b-a37b — size and active part: 660 billion weights, about 95 billion active per step. This marks an MoE;
• -27b — a dense model: all 32 billion take part in every step, but behaviour and latency are more predictable;
• max and flash — positions in the line-up, not architecture: the flagship and the fast variant for load.

The quantisation suffix stands apart: Q4_K_M, Q5_K_S, AWQ, GPTQ. That is no longer a property of the model but a way of packing weights — and it is what decides whether the model fits in your memory.