SQE Leaderboard/MiniMax M2.5

MiniMax M2.5

QuantTrio/MiniMax-M2.5-AWQ

fourlineCommunity AWQ SWE-bench Pro 55.4% community2026-08-02
Methodology

SQE score

57

blended mean, 5 reps

Gate

3/5

all reps counted

Wilson 95% low

≥23.1%

lower confidence bound

Shots to green

1

median of 5 reps

Every published metric

SQE score
Headline blended mean percent across reps (gate 45 / pass 25 / repair 15 / style 10 / thrift 5).
57n=5
Gate
Reps that cleared known and held-out honesty checks, shown as k/n after tainted exclusions.
3/5n=5
Pass
Reps that cleared public verification (compile, known conformance, scenario threshold), as k/n. Distinct from gate, which also requires held-out honesty.
3/5n=5
Wilson 95%
Lower bound of the gate pass rate at 95% confidence.
≥23.1%n=5
Harness ±
Blend of all score components, mean ± spread.
57.5 ±92.7n=5
Shots to green
Feedback rounds needed before tests pass. Lower is better.
1n=5
Scenario pass
Share of graded scenarios passing in the green build.
91.7%n=5
Trap/spec rate
Share of trap parameter values correctly synthesized from the spec.
0%n=5
1-shot gap
How much the first shot trailed the final result.
0.0 ptsn=5
Effective n
Reps counted after tainted-run exclusion.
5/5n=5
Median tokens / rep
Prompt + completion tokens per rep, median over untainted reps. Lower is better.
1.83M tokn=5 untainted
Min VRAM
Catalog minimum cluster VRAM (GB) from syndicAI referenceGpuRequirements. Closed APIs and packs not in the catalog show as -.
171 GB
SWE-bench Pro
Catalog SWE-bench Pro score for the resolved syndicAI model, or the vendor-published score for closed-reference models. Inferred scores are labelled. Packs without a score show as -.
55.4%

Run configuration

Endpoint model
QuantTrio/MiniMax-M2.5-AWQ
SWE-bench Pro
55.4% · SWE-bench Pro · MiniMax (M2.5)
Reps
5
Max shots
3
Feedback tier
1
Context cap
200K
Token budget / run
4M
Turn cap / shot
120
Thinking budget
4.096K
Sandbox digest
quality-eval-sandbox@sha256:242ebd2f73c1c1ffa90c3ccb7913574f91e74a89dd1e4cf874f907a3129149c5

Review

Generated by claude-haiku-4-5

Overall Verdict

The model achieves 60% gate pass rate (3/5 runs) on the fourline Connect-Four task, with strong performance when successful but concerning inconsistency. Three runs reach full correctness with single-shot repair and 93–95% overall scores; two runs plateau without recovery, hitting logic bugs (gravity/win-detection failures) that don't self-correct despite additional turns. Passing runs compile cleanly, pass all 14 scenarios including held-out conformance, and exit efficiently. Failing runs incur 110–120 turns and 3.7–4M prompt tokens without convergence, suggesting the model struggles to debug certain state-machine semantics once initial implementation diverges from correct behavior.

The 4-bit quantization (AWQ) is not a credible root cause: passing runs demonstrate precise TypeScript and game-logic reasoning under identical constraint. Failures instead reflect agent-level issues—difficulty diagnosing why gravity-drop returns null instead of "Y", or why win-state transitions remain stuck in "playing"—suggesting the model needs stronger introspection or error-message parsing to break cycles.

Strengths

  • Reliable when correct: 3/5 runs compile, pass all 14 scenarios, and clear held-out conformance on first attempt.
  • Self-directed quality: All runs self-initiate typecheck and lint; passes respect functional linting constraints (const, readonly).
  • Token thrift in success cases: Passing runs use 0.40–0.71 token-thrift score despite high prompt volume (1.7–3.5M tokens), indicating reasonable code generation efficiency.
  • No overfitting: Trap specs and held-out scenarios unambiguously passed in green runs.

Failure Modes

  • Logic bugs don't self-repair: r2 and r5 fail on gravity-drop (returns null instead of piece color) and/or win-vertical/conformance-replay (state remains "playing" instead of transitioning to "won"). Despite 3–6 tool-call errors per run and extended deliberation (110 turns in r2), the model does not recover; repair curve plateaus.
  • Tool-call error clustering: Failing runs incur 5–6 tool errors vs. 2–4 in passing runs, suggesting repeated format misalignment or harness friction during repair loops.
  • Context burn without convergence: r2 consumes 3.76M prompt tokens and 693 wall-clock seconds across 3 shots without reaching green; token thrift drops to 0.37, indicating long non-productive turns.

Serving & Harness Notes

  • Turn counts: Passing runs settle in 66–74 turns; failing runs expand to 110–120 turns without escape condition, suggesting the harness should enforce stricter repair-attempt limits or require explicit strategy restart.
  • No reasoning-only stalls observed: All runs complete with status completed; no hung inference.
  • Compile always succeeds: Type-safety preserved even in failed logic runs, reducing risk of runtime crashes in deployment.

Per-rep results