Beam 1.0 501B A23B vs Underdog Saluki 27B, head to head

Side A

Side B

Overall

Beam 1.0 501B A23B is clearly better. It wins 51 of the 70 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. Underdog Saluki 27B spends 1.7× as many thinking tokens per test case.

Thinking tokensAnswer tokens

TaskABeam 1.0 501B A23BBUnderdog Saluki 27BBeam 1.0 501B A23BUnderdog Saluki 27B

Overall

5717.2s

4K6.9K

335m 15s

Agents

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

5521.8s

4.8K11.8K

409m 2s

Structured Extraction

Turns logs, diffs and records into exact JSON

819.98s

3.4K4.4K

433m 17s

Coding

Writes code and SQL that pass hidden tests

4819.9s

4.5K9.2K

166m 54s

Grounded Answers

Answers from the given documents with sources, never guesses

6911.6s

2.7K4.5K

433m 22s

Instruction Following

Applies every rule, including ones changed mid-conversation

7820.9s

5.6K8K

466m 5s

Business Logic

Computes billing, scheduling and pricing rules exactly

5714.0s

4.2K5.2K

333m 55s

Security & permissions

Ignores planted instructions, still does the allowed work

5120.6s

3.2K4.9K

433m 47s

Diagnosis & investigation

Chooses diagnostic checks, repairs supported causes and verifies recovery

1724.3s

3.1K8.7K

06m 39s

What you'd run

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

SpecABeam 1.0 501B A23BBUnderdog Saluki 27B

Speed

274 tok/s

23 tok/s

Machine

Hosted

Apple M5 Pro, 48 GB

Memory at 32K

—

10.2 GB

Context

256K

Cost per run

$0.83

Not measured

Cost per case

$0.0054

Not measured

Time per case

17.2s

5m 15s

Quantization

FP8

IQ2_XS

Harness

openrouter

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.

Beam 1.0 501B A23BUnderdog Saluki 27B

Every test case

153 test cases

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

Agents64

Structured Extraction113

Coding104

Grounded Answers60

Instruction Following72

Business Logic61

Security & permissions45

Diagnosis & investigation10

51 won by Beam 1.0 501B A23B19 won by Underdog Saluki 27B83 tied