Mellum2.1 Thinking vs Qwen3.8 Flash, head to head

Side A

Side B

Overall

Qwen3.8 Flash is clearly better. It wins 55 of the 60 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. Mellum2.1 Thinking spends 2.0× as many thinking tokens per test case.

Thinking tokensAnswer tokens

TaskAMellum2.1 ThinkingBQwen3.8 FlashMellum2.1 ThinkingQwen3.8 Flash

Overall

91m 56s

10.3K4K

3258.5s

Agents

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

92m 3s

10.8K4.2K

491m 25s

Structured Extraction

Turns logs, diffs and records into exact JSON

51m 36s

8.3K3K

3336.9s

Coding

Writes code and SQL that pass hidden tests

52m 11s

11.3K5.8K

471m 21s

Grounded Answers

Answers from the given documents with sources, never guesses

131m 24s

7.3K3.1K

2641.4s

Instruction Following

Applies every rule, including ones changed mid-conversation

283m 10s

17.3K5K

371m 4s

Business Logic

Computes billing, scheduling and pricing rules exactly

01m 39s

9K3.9K

1746.8s

Security & permissions

Ignores planted instructions, still does the allowed work

81m 40s

9.1K3.2K

5054.1s

Diagnosis & investigation

Chooses diagnostic checks, repairs supported causes and verifies recovery

01m 15s

6.8K3.3K

058.4s

What you'd run

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

SpecAMellum2.1 ThinkingBQwen3.8 Flash

Speed

90 tok/s

77 tok/s

Machine

Apple M5 Pro, 48 GB

Hosted

Memory at 32K

12.9 GB

—

Context

128K

Cost per run

$2.46 est.

$0.46

Cost per case

$0.016 est.

$0.003

Time per case

1m 56s

58.5s

Quantization

Q8_0

—

Harness

llama.cpp

openrouter

The race

No lead changes

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

Mellum2.1 ThinkingQwen3.8 Flash

Every test case

153 test cases

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

Agents09

Structured Extraction18

Coding012

Grounded Answers15

Instruction Following35

Business Logic05

Security & permissions011

Diagnosis & investigation00

5 won by Mellum2.1 Thinking55 won by Qwen3.8 Flash93 tied