← Blog

GLM: Thirty-seven hours for 98 percent

GLM: Thirty-seven hours for 98 percent

A frontier-level local model: GLM 5.3 at maximum effort. A new kid on the block, MiMo v2.6 Flash, is carving out a solid place with the luxury of being very fast and serving in parallel batch across four nodes at insane speeds. Capable of serving sixteen concurrent inferences without flinching.

Ninety-eight

GLM 5.3 in its MAX profile, 8-bit weights with bf16 heads, scores 98.0 on my benchmark: 42 tasks across agentic work, code, planning, reasoning and writing, judged by Opus 4.8. Its cloud version on OpenRouter scores 98.3. The gap is noise. What closes it is the maximum-effort profile, not the 8-bit quantization on its own.

Planning and reasoning both land at 100. Twenty tasks, not a single point lost. Across the whole panel, 24 of the 42 tests sit at the maximum.

When more than half of a benchmark sits at the ceiling, the ceiling is measuring the grid as much as the model. My grids no longer separate GLM MAX from a perfect score. Separating it from itself will take harder ones.

Thirty-seven hours

The 98 has a price. 1.41 million completion tokens for 42 tasks, 33,650 per task on average. One code task, c06, took 119,160 tokens and three hours nine minutes on its own. At about ten tokens per second, the run adds up to 37 hours of generation, spread over four days.

Those tokens include the thinking. The harness exposes no separate reasoning count: reasoning and answer are fused in the completion. I cannot say how much of the 33,650 is thought and how much is answer. What I can say is that this volume is what buys the score.

Speed is not this model’s argument. GLM MAX is a background model: batch jobs, long reviews, work where quality comes before latency. It does not belong behind a real-time endpoint.

Sixteen at once

MiMo v2.6 Flash runs at 9 bits and fits on a 256 GB Mac. On a cluster of four nodes in parallel, that means sixteen concurrent requests handled, sixty in total, 500 tokens max each. Sixty out of sixty succeeded. Aggregate throughput 279.5 tokens per second. Per stream, 26.1 at the median, with every request between 25.3 and 26.5. Contention takes nothing from each stream; the batching adds throughput on top. First token in 0.74 seconds at the median, 2.81 at P99.

On the same benchmark it scores 86.0, twelve points below GLM MAX. It spends 282,578 tokens on the 42 tasks, 6,728 per task, under two hours of generation.

The number that matters as much as the 86 is the zero. Zero non-answers on 42: no collapse, no truncated generation, no reasoning stream handed back where code was expected.

A Qwen3.8 Flash-Next at Q8 scores slightly higher, 90.3. It does not fit in 256 GB.

The car wash

The twelve points widen on shaky logic. On a puzzle with a hidden constraint, MiMo picks the intuitive answer, and it is wrong: walking to the car wash, when it is the car that needs washing. GLM does not fall for it. The grid has a hard rule: miss the constraint and the first section scores zero.

GLM MAX: 47 out of 50. MiMo v2.6 Flash: 22.5.

This is the test that discriminates where global scores converge. Noticing what the question does not say is exactly the reasoning MiMo gives up, and it is the real price of its throughput.

Two jobs

Absolute quality, in batch, with time to spare: GLM 5.3 MAX. A local endpoint that has to answer sixteen people at once: MiMo v2.6 Flash.

Now, on my cluster, MiMo is running as a pair with GLM. A perfect duo, speed and accuracy.

Sophie, The Monocle Bear