InstructionsFollowing instructions

IFBench

Loose prompt grading · 300 tasks. Exact task sets and protocols remain separate.

Will it follow the precise instructions?

68.3%Recorded score · 205 / 300 tasks

GLM-5.3-Flash · NVFP4 (LibertAIDAI) · 2× GB10 (DGX Spark head + ASUS GX10 worker) · 256GB total

Inspect the measured run ↗How scoring works ↓

THE QUESTION IT ANSWERS

Will it follow the precise instructions?

Looks beyond a plausible answer to whether a response obeys specific, verifiable requirements.

01Prompt + constraints
02Produce a response
03Constraint checks

Interpret the score Our local headline uses loose prompt-level grading. Strict grading and published protocols differ.

SETUP PERFORMANCE

A capability profile, test by test

Each bar uses its benchmark’s scoring rule. Higher is better.

050100%

IFBenchTested by LocalsOnly

68.3%

Open any bar for the score, sample size, and evaluation settings. A missing measurement is never shown as zero.

What this test is

What this benchmark measures

Instruction-following constraints. Local receipt uses a different protocol from the vendor card.

The benchmark tests generalization to 58 unfamiliar precise constraints. Strict and loose criteria answer different questions and must be labeled separately.

IFBench paper ↗

How it grades

How a task earns its score

Instruction-following is reported as an aggregate. The local receipt uses a different protocol from the vendor card and has no per-task traces on this board.

Published rows are vendor or bench cites. They do not share the measured task grid.

SuiteIFBench

ReceiptNo per-task records

Headline / aggregateNo per-task records on this receipt

Follow one task · this measured recipe

A passed task and a failed one

The LocalsOnly measured receipt for this test has no per-task records to inspect. The aggregate result is available below.

Subject & category breakdown · this recipe

Results by subject or work category

This receipt does not record a subject or work-category breakdown. A suite-wide label cannot support a finer subject comparison. Outcome groups such as “passed” and “timeout” are not subjects.

Ranking

Results & published references

Place is assigned only among rows that share a comparable score on this board. Evaluation settings can differ across sources.

Tested by LocalsOnly

Results from the exact local setup shown below.

6 rows

#	Model / recipe	Reported score ↓	Runs on	Evidence
01	◎GLM-5.3-Flash · SGLangNVFP4 (LibertAIDAI) · SGLang	68.3% 205 / 300	2× GB10 (DGX Spark head + ASUS GX10 worker) · 256GB total	● Tested by LocalsOnlyLocalsOnly	↗
02	◎GLM-5.3-Flash · vLLMEXL3 TR3 4bpw · vLLM	67.7% 203 / 300	2× GB10 (DGX Spark head + ASUS GX10 worker) · 256GB total	● Tested by LocalsOnlyLocalsOnly	↗
03	◎GLM-5.3-Flash · TensorFoldEXL3 TR3 4bpw · TensorFold	66.3% 199 / 300	2× GB10 (DGX Spark head + ASUS GX10 worker) · 256GB total	● Tested by LocalsOnlyLocalsOnly	↗
04	✳Qwen 3.8 Flash-Next · llama.cppUD-IQ3_XXS · llama.cpp	66.0% 198 / 300	GB10 · 128GB	● Tested by LocalsOnlyLocalsOnly	↗
05	◎DeepSeek V4 Flash · llama.cppUD-IQ3_XXS · llama.cpp	65.3% 196 / 300	GB10 · 128GB	● Tested by LocalsOnlyLocalsOnly	↗
—	✳Qwen 3.8 27B · llama.cppUD-Q4_K_XL · llama.cpp	— 127 / 300	GB10 · 128GB	● Tested by LocalsOnlyLocalsOnly	↗

Published reference

Results reported by model developers and benchmark authors.

2 rows

#	Model / recipe	Reported score ↓	Runs on	Evidence
01	✳Qwen 3.8 27BBF16 reference · Reported evaluation	79.5% 79.5%	Published reference	↗ Published resultQwen	↗
02	◎GLM-5.3-Flash NVFP4Published reference · Reported evaluation	60.5% 60.5%	Published reference	↗ Published resultNVIDIA	↗