Turning GLM-5.3-Flash into a Jev-like model

The post claims on-par accuracy and speed against Jev and a wide margin over Laya, using GLM-5.3-Flash with vLLM and a prompt shaped so the first output token is the answer. A commenter says he got competitive accuracy with Qwen 3.8 and Gemma 4 and calls GLM 5.3 overkill.

Speed and cost. That setup costs several times more per decision than Jev, and the comment that isolates the difference says Jev's addition is that the reported probabilities track actual likelihoods, which stock LLMs do not give you even when they pick the right option.

One commenter asks whether the measurement reused the same long excerpt throughout and whether partial cache hits tainted it, noting any provider could hit those numbers that way. He calls 20k tokens per second on B200 ordinary for a model this size and 50k genuinely impressive.

One commenter reports it was badly wrong on his use cases, and that the team's reply was to boil the question down further, a spiral of how much to dumb a request down before it lands. Accuracy here depends heavily on how narrowly you write the question.

Craft the input prompt so the first output token is the answer, which turns the decision into a single forward pass with no generation loop. The write-up applies it to GLM-5.3-Flash on vLLM and benchmarks against Jev and Laya, finding on-par accuracy and speed with a wide margin over Laya, at several times the cost per decision and with vision input support.

One commenter tried it with a random book excerpt of about 23,000 words, roughly 30k input tokens, and got responses in 800ms and sometimes 500ms. He puts that in the 20,000 to 50,000 tokens per second prefill range and calls it impossible for normal LLMs, and another comment notes it also processes every question in parallel, which normal LLMs cannot do.

Several, none confirmed. One says the splash is from ultra-low cost and speed that may be architectural, and suggests a BERT-like bidirectional model rather than an autoregressive one is one way to get it. Another proposes running prefill in a fully sliding-window mode with block-boundary duplication so it is almost embarrassingly parallel. One reply notes he ran Cerebras as an experimental ultrafast Jev and it was faster.