Is GRM-3.2-Sky good for coding?

In our benchmarks, GRM-3.2-Sky scores 78.5/100 for coding. It runs at about 98.0 tok/s, so if you want more speed, Qwen3.6-35B-A3B-Q4_K_XL is faster (~156.9 tok/s) and still scores well for coding (82.2/100).

Is GRM-3.2-Sky good for agentic (tool-using) tasks?

In our benchmarks, GRM-3.2-Sky scores 89.3/100 for agentic workflows. It runs at about 98.0 tok/s, so if you want more speed, swift-1.5-qwen3.8-flash-next-iq3_s is faster (~259.0 tok/s) and still scores well for agentic workflows (88.9/100).

How fast is GRM-3.2-Sky for local inference?

Across 4 community benchmark runs, GRM-3.2-Sky reaches up to 99.5 tok/s and averages 98.0 tok/s, with the fastest results on Tesla V100-PCIE-32GB.

How much memory does GRM-3.2-Sky need?

The leanest observed configuration used about 46.6 GB of memory.

Which tools have been used to run GRM-3.2-Sky?

Benchmarks were submitted using llama.cpp. Results are community-contributed and updated as new runs arrive.

Model benchmarks

GRM-3.2-Sky local LLM performance

As of November 2026, GRM-3.2-Sky runs at up to 99.5 tok/s for local inference (best of 4 community benchmark runs across 1 GPU).

llama.cpp

ShareReddit X

Model size

Unknown

Peak speed

99.5 tok/s

Average speed

98.0 tok/s

Avg PP

471.8 tok/s

Min memory

46.6 GB

Max context

32,768 tokens

Avg output / run

14,202 tokens

Avg runtime / run

2m 33s

Avg quality

84.6

Benchmark runs

GPUs tested

Quality by task

Average LLM-judged quality (0–100) with the run-to-run spread shown as a P5–P95 band, overall and for each benchmark task, across all 4 runs. The low and high columns show how much the judge’s score varies between runs, and need at least two runs to display.

Task	P5 (low)	Avg	P95 (high)
Overall	83.2	84.6	86.0
Agent Workflow	86.5	89.3	92.4
Code Generation	75.6	78.5	80.8
Role Play & Narrative	83.0	87.9	90.4
Research & Analysis	79.7	82.6	84.5

Performance by hardware and tool

Every hardware/tool/quantization combination GRM-3.2-Sky has been benchmarked on, ranked by peak token generation speed. Last updated November 2026.

Hardware	Tool	Quant	Best tok/s	Avg tok/s	Memory	Context	Quality	Runs
Tesla V100-PCIE-32GB	llama.cpp	—	99.5 tok/s	98.0 tok/s	46.6 GB	32,768 tokens	84.6	4

Benchmark runs

All 4 GRM-3.2-Sky runs submitted so far — expand one for its hardware, quality breakdown and per-scenario detail.

Open run page →

Frequently asked questions

Is GRM-3.2-Sky good for coding?: In our benchmarks, GRM-3.2-Sky scores 78.5/100 for coding. It runs at about 98.0 tok/s, so if you want more speed, Qwen3.6-35B-A3B-Q4_K_XL is faster (~156.9 tok/s) and still scores well for coding (82.2/100).
Is GRM-3.2-Sky good for agentic (tool-using) tasks?: In our benchmarks, GRM-3.2-Sky scores 89.3/100 for agentic workflows. It runs at about 98.0 tok/s, so if you want more speed, swift-1.5-qwen3.8-flash-next-iq3_s is faster (~259.0 tok/s) and still scores well for agentic workflows (88.9/100).
How fast is GRM-3.2-Sky for local inference?: Across 4 community benchmark runs, GRM-3.2-Sky reaches up to 99.5 tok/s and averages 98.0 tok/s, with the fastest results on Tesla V100-PCIE-32GB.
How much memory does GRM-3.2-Sky need?: The leanest observed configuration used about 46.6 GB of memory.
Which tools have been used to run GRM-3.2-Sky?: Benchmarks were submitted using llama.cpp. Results are community-contributed and updated as new runs arrive.