LFM2.5 24B A2B vs Mellum2.1 Thinking, head to head

Side A

Side B

Overall

Too close to call. Mellum2.1 Thinking wins 14 of the 25 test cases where they differ.

Learn to build evals like this in the AI Engineering Academy

Task by task

A score turns orange only when that model wins clearly more of the task's test cases. Under each score, the average time per test case on the machine it ran on. LFM2.5 24B A2B spends 1.3× as many thinking tokens per test case.

Thinking tokensAnswer tokens

TaskALFM2.5 24B A2BBMellum2.1 ThinkingLFM2.5 24B A2BMellum2.1 Thinking

Overall

71m 53s

9.2K10.3K

91m 56s

Agents

Runs multi-step jobs through tools, within a written policy

123m 49s

18.5K10.8K

92m 3s

Structured Extraction

Turns logs, diffs and records into exact JSON

1054.9s

4.4K8.3K

51m 36s

Coding

Writes code and SQL that pass hidden tests

11m 15s

6.1K11.3K

52m 11s

Grounded Answers

Answers from the given documents with sources, never guesses

557.8s

4.7K7.3K

131m 24s

Instruction Following

Applies every rule, including ones changed mid-conversation

162m 44s

13.4K17.3K

283m 10s

Business Logic

Computes billing, scheduling and pricing rules exactly

51m 2s

5K9K

01m 39s

Security & permissions

Ignores planted instructions, still does the allowed work

101m 52s

9K9.1K

81m 40s

Diagnosis & investigation

Chooses diagnostic checks, repairs supported causes and verifies recovery

04m 12s

19.9K6.8K

01m 15s

What you'd run

The build behind each score. Orange marks the faster, smaller, cheaper or longer-context side.

SpecALFM2.5 24B A2BBMellum2.1 Thinking

Speed

82 tok/s

90 tok/s

Machine

Apple M5 Pro, 48 GB

Memory at 32K

25.0 GB

12.9 GB

Context

128K

Cost per run

Not measured

$2.46 est.

Cost per case

Not measured

$0.016 est.

Time per case

1m 53s

1m 56s

Quantization

Q8_0

Harness

llama.cpp

The race

1 lead change

Each model's running score as the test cases go by. The dashed line is a perfect run.

LFM2.5 24B A2BMellum2.1 Thinking

Every test case

153 test cases

Each square is one test case, colored by the model that scored higher.

Agents11

Structured Extraction21

Coding22

Grounded Answers13

Instruction Following25

Business Logic10

Security & permissions22

Diagnosis & investigation00

11 won by LFM2.5 24B A2B14 won by Mellum2.1 Thinking128 tied