MiniMax M3

Correctness?

Efficiency?

Robustness?

Correctness says whether it can do the work; efficiency whether it finishes inside the same budget everyone gets; robustness whether it invents or omits. A low score should be readable as one of these, not a mystery.

Task type × domain?

Strict pass over this model’s tasks (share answered exactly right · task count). Rows are task types, columns are the data the answer lives in.

Where the points went?

Every task not answered exactly right, bucketed by what went wrong. 47 tasks.

Task families?

Lowest-scoring tasks?

Public example tasks link to the full question and every model’s answer; hidden tasks show only the family, the score and why.

• 0.0
  S-CR08-NPool Buy/Sell Transaction-Count Imbalance — count of matchesview task

  Buy/sell transaction ratio = number of buy swaps divided by number of sell swaps over 24h (transaction COUNTS, not volume — v0.1 proxy for the PRD's buy/sell volume imbalance). Find DEX liquidity pools in the top-200-by-24h-volume universe of each supported network with buy/sell transaction ratio greater than 3, pool liquidity greater than $1M. Scope: only entities with a known value for buy sell txn ratio in the snapshot; entities lacking any of them are out of scope and must not be returned. Entity ids are pool ids of the form network:pooladdress exactly as listed in the snapshot. EVM pool addresses (0x…) are lowercase hex and case-insensitive; Solana and other base58 addresses are CASE-SENSITIVE and must be returned exactly as shown. How many entities match ALL of the conditions above? Answer with the integer count (0 if none).

  Expected 8, got 0 (1.00 off) — outside tolerance.
• 0.0
  S-EQ09-XMoving Average Cross (EMA20 above EMA50 today) — plus one more hidden conditionview task

  A cross means EMA20 > EMA50 today while EMA20 <= EMA50 on the previous session (EMAs on split-adjusted closes, SMA-seeded). Find US-listed common equities (instruments without a reported market cap — warrants, units, rights, preferred series, notes — are not in the universe) with market cap greater than $1B, 20-day average dollar volume greater than $20M, EMA20 crossing above EMA50 today, 60-day return greater than 0%. Scope: only entities with a known value for 20-day average dollar volume, EMA20 crossing above EMA50 today, 60-day return in the snapshot; entities lacking any of them are out of scope and must not be returned. Return every matching ticker; order does not matter. Rebuild policy: metrics that are not served as fields (fundamental ratios, growth rates, valuation, technical signals) must be recomputed by you from the raw facts and daily bars the tools serve, following the definitions in list_metrics.compute_yourself to the letter (period selection, latest-filing copy, quarter arithmetic); do the arithmetic with run_calculation over the fetched data. The reference answer was computed from the same facts by the same definitions.

  Gold has 1, the model returned 3: 0 correct, 3 false positives, 1 missed → F1 0%. Not a strict pass.
• 0.0
  S-CR06-MToken DEX Liquidity Threshold (pools) — median of a rebuilt metric over the matchesview task

  A token's aggregate DEX liquidity is the sum of pool liquidity over every pool in the universe (same network) whose base token is that token; every pool of a qualifying token matches. Find DEX liquidity pools in the top-200-by-24h-volume universe of each supported network with base token's aggregate DEX liquidity greater than $50M, network equal to base. Scope: only entities with a known value for token dex liquidity usd, network in the snapshot; entities lacking any of them are out of scope and must not be returned. Entity ids are pool ids of the form network:pooladdress exactly as listed in the snapshot. EVM pool addresses (0x…) are lowercase hex and case-insensitive; Solana and other base58 addresses are CASE-SENSITIVE and must be returned exactly as shown. Across ALL entities that match the conditions above, what is the median Base token DEX liquidity (USD) (Sum of pool_liquidity_usd over every fetched pool on the same network whose base token has the same contract address (the token's aggregate liquidity within the supported pool universe).)? Median = the middle value, or the mean of the two middle values for an even count. Answer in absolute USD. Clarification: in "network equal to base", base is the network id of the Base L2 chain (the value of the network field), not the pool's base token; the condition keeps only pools whose network field equals "base".

  Expected 84.78M, got 130.67M (54% off) — outside tolerance.
• 0.0
  F09Free Cash Flowview task

  What was AMD's free cash flow for fiscal year 2023, defined as operating cash flow minus capital expenditures (FCF definition v1)? Answer in absolute USD. Filing-version policy: every financial value is taken from the LATEST filing accepted by the SEC at or before the data cutoff that reports the requested period. If a later filing (before the cutoff) restates an earlier period, the restated value is the answer; the original filing's value is not. A fiscal-year value is taken only from an annual report (10-K / 20-F / 40-F, including amendments) — a quarterly report never supplies an annual value.

  Expected 1.12B, got 1.14B (2% off) — outside tolerance.
• 0.0
  S-CR19-NPerp Volatility Expansion (ATR% threshold) — count of matchesview task

  Perpetual technicals use COMPLETED UTC daily bars only (Hyperliquid 1d candles). "The session" below means the last closed UTC session — its date is the row attribute last_closed_bar — never the bar still forming at snapshot time. ATR% = 14-day Wilder ATR divided by the session's close. Open interest (USD) is taken at snapshot time, not at the session close. Find Hyperliquid perpetual markets with open interest greater than $20M, perp atr pct 14 greater than 5%. Scope: only entities with a known value for perp atr pct 14 in the snapshot; entities lacking any of them are out of scope and must not be returned. How many entities match ALL of the conditions above? Answer with the integer count (0 if none). Rebuild policy: metrics that are not served as fields (fundamental ratios, growth rates, valuation, technical signals) must be recomputed by you from the raw facts and daily bars the tools serve, following the definitions in list_metrics.compute_yourself to the letter (period selection, latest-filing copy, quarter arithmetic); do the arithmetic with run_calculation over the fetched data. The reference answer was computed from the same facts by the same definitions.

  Expected 30, got 29 (3% off) — outside tolerance.
• 0.0
  S-EQ12-XXOversold / Overbought (RSI) — plus two more hidden conditionshidden

  Gold has 8, the model returned 0: 0 correct, 8 missed → F1 0%. Not a strict pass.
• 0.0
  F04-RRevenue Growth — six companies rankedview task

  For each of these six companies — MSFT, META, JPM, GOOGL, PANW, NVDA — answer the same question: "By what percentage did the company's full-year revenue grow from fiscal year 2023 to fiscal year 2025? Answer as a decimal ratio (e.g. 0.082 for 8.2%)." Then rank the six tickers by that value, highest first (all six have the required filings). Return the six tickers in order; order is graded. Filing-version policy: every financial value is taken from the LATEST filing accepted by the SEC at or before the data cutoff that reports the requested period. If a later filing (before the cutoff) restates an earlier period, the restated value is the answer; the original filing's value is not. A fiscal-year value is taken only from an annual report (10-K / 20-F / 40-F, including amendments) — a quarterly report never supplies an annual value.

  Budget ran out and the answer was forced. Order graded by pairwise over the top 6: 0.00.
• 0.0
  S-EQ19-XX-NRevenue Growth Screen — plus two more hidden conditions — count of matchesview task

  Fundamental ratios are not served as fields: rebuild each one from the raw SEC facts (get_equity_financials) exactly as defined in list_metrics.compute_yourself, using the latest filing at or before the data cutoff; a company whose filings lack the needed facts is out of scope. Find US-listed common equities (instruments without a reported market cap — warrants, units, rights, preferred series, notes — are not in the universe) with market cap greater than $10B, 20-day average dollar volume greater than $300M, sector equal to Utilities, latest-quarter revenue growth (YoY) greater than 20%, turnover ratio greater than 0.002%, RSI(14) less than 70. Scope: only entities with a known value for 20-day average dollar volume, sector, latest-quarter revenue growth (YoY), turnover ratio, RSI(14) in the snapshot; entities lacking any of them are out of scope and must not be returned. How many entities match ALL of the conditions above? Answer with the integer count (0 if none). Filing-version policy: every financial value is taken from the LATEST filing accepted by the SEC at or before the data cutoff that reports the requested period. If a later filing (before the cutoff) restates an earlier period, the restated value is the answer; the original filing's value is not. A fiscal-year value is taken only from an annual report (10-K / 20-F / 40-F, including amendments) — a quarterly report never supplies an annual value. Rebuild policy: metrics that are not served as fields (fundamental ratios, growth rates, valuation, technical signals) must be recomputed by you from the raw facts and daily bars the tools serve, following the definitions in list_metrics.compute_yourself to the letter (period selection, latest-filing copy, quarter arithmetic); do the arithmetic with run_calculation over the fetched data. The reference answer was computed from the same facts by the same definitions.

  Expected 4, got 2 (50% off) — outside tolerance.