In 2026, this chapter uses flash tier to mean a model service built around a lower per-token bill and fast enough responses for high-volume work. A token is a chunk of text the model reads or produces. The category is not one architecture. Qwen3.8-Flash-Next, GLM-5.3-Flash, and DeepSeek-V4-Flash all try to reduce the cost of that token, but they save work in different places. Qwen's distinctive move is to keep part of its capacity in ordinary host random-access memory (RAM) instead of placing it permanently on the graphics processing unit (GPU).
That difference matters because an inference bill has more than one source. A model must perform arithmetic through activated parameters, read and write attention state for the growing conversation, and move data between memory tiers. Sparse models reduce the first cost by activating only selected experts. Long-context designs address the second and third costs in different ways. Comparing only total parameter counts would hide the engineering choice each system made.
Click a column label to keep that property highlighted. Start with “active/token,” then compare “uses ordinary RAM?” and “attention recipe.” The table is most useful as a map of mechanisms: it tells you which expense each design is trying to avoid before you look at the price column.
Compare three Flash designs
Click a column, then compare the highlighted values.
Each design reduces a different bottleneck. Qwen moves some memory off the accelerator. DeepSeek compresses the memory it must keep. GLM reduces attention work and cache growth. They pursue a similar price target through different designs.
For Qwen, the memory split is the n-gram table discussed in Chapter 2. Its 51B lookup parameters are addressed from recent tokens and prefetched, so they can add local-pattern capacity without sitting in GPU memory all the time. GLM's hybrid pattern uses a running state in three of every four layers, so those layers do not reread a growing key-value (KV) cache. DeepSeek's path compresses the cache that attention needs to keep. The diagram is a systems comparison: three attempts to make long context less dominated by memory traffic.
How should we read the benchmark bar?
The panel below compares Qwen's reported pre-training quality with its own predecessor. It does not establish a universal ranking, and it does not turn a benchmark score into a price quote. It does show the central tradeoff Qwen is asking readers to evaluate: the report says Flash-Next leads on eight of fourteen pre-training benchmarks, trails by at most 2.6 points on the rest, and uses roughly one third of the predecessor's active parameters to serve.
Pre-training quality vs its own predecessor
report Tab. 11, 14 benchmarks: leads on 8, trails by ≤2.6 pts on the rest (self-reported)
- Flash-Next wins: of 14 pre-training benchmarks
- predecessor wins: worst deficit: 2.6 points
The 89% as good at ~11% of the training cost phrasing is an editorial shorthand in this comparison, rather than a benchmark Qwen reports. The underlying reported result is eight wins, six losses with a worst 2.6-point deficit, and 1/3 the active parameters to serve. Independent flash-tier comparisons (Local AI Zone's 24-dimension analysis) group Flash-Next with GLM-5.3-Flash and DeepSeek-V4-Flash as the three architectures defining the category.
The deeper story is that the three models place their savings at different points in the serving path. Qwen says capacity can be split across the memory hierarchy. DeepSeek says the remembered state can be compressed. GLM says much of the context path can use linear state and sparse retrieval. Kimi, in the larger weight class named in this article, argues for scaling the attention path differently. The ideas can coexist, but they create different hardware needs and different failure modes when context grows.
When you compare flash-tier models for a real workload, start with the work rather than the headline. A prompt that repeats a large prefix stresses cache reuse. A long agent trace stresses retrieval and state growth. A deployment constrained by accelerator memory makes Qwen's host-RAM table more relevant. The published benchmark and price numbers are useful evidence, but the architecture tells you what must remain true for those numbers to carry over to your own request pattern.