DeepSeek V4.1 Flash 1017 vs LFM2.5 24B A2B, head to head

Side A

Side B

Overall

DeepSeek V4.1 Flash 1017 is clearly better. It wins 88 of the 89 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 1.8× as many thinking tokens per test case.

Thinking tokensAnswer tokens

TaskADeepSeek V4.1 Flash 1017BLFM2.5 24B A2BDeepSeek V4.1 Flash 1017LFM2.5 24B A2B

Overall

5626.0s

5.3K9.2K

71m 53s

Agents

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

5826.7s

3.3K18.5K

123m 49s

Structured Extraction

Turns logs, diffs and records into exact JSON

8117.2s

4.2K4.4K

1054.9s

Coding

Writes code and SQL that pass hidden tests

4840.8s

9.4K6.1K

11m 15s

Grounded Answers

Answers from the given documents with sources, never guesses

7119.8s

4.4K4.7K

557.8s

Instruction Following

Applies every rule, including ones changed mid-conversation

7625.6s

5.5K13.4K

162m 44s

Business Logic

Computes billing, scheduling and pricing rules exactly

6726.6s

6.6K5K

51m 2s

Security & permissions

Ignores planted instructions, still does the allowed work

4826.6s

4.7K9K

101m 52s

Diagnosis & investigation

Chooses diagnostic checks, repairs supported causes and verifies recovery

020.2s

3.2K19.9K

04m 12s

What you'd run

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

SpecADeepSeek V4.1 Flash 1017BLFM2.5 24B A2B

Speed

243 tok/s

82 tok/s

Machine

Hosted

Apple M5 Pro, 48 GB

Memory at 32K

—

25.0 GB

Context

128K

Cost per run

$0.75

Not measured

Cost per case

$0.0049

Not measured

Time per case

26.0s

1m 53s

Quantization

—

Q8_0

Harness

openrouter

llama.cpp

The race

No lead changes

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

DeepSeek V4.1 Flash 1017LFM2.5 24B A2B

Every test case

153 test cases

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

Agents120

Structured Extraction150

Coding101

Grounded Answers140

Instruction Following150

Business Logic130

Security & permissions90

Diagnosis & investigation00

88 won by DeepSeek V4.1 Flash 10171 won by LFM2.5 24B A2B64 tied