MiniMax M3.1 Flash

Original Article ↗ Hacker News Discussion ↗

MiniMax’s new M3.1 Flash large language model is being scrutinized against rivals like DeepSeek V4.1 Flash, GLM 5.3 Flash, and frontier models such as Opus 5.5 and GPT-5.6 Sol, with many users finding it slower, more verbose, and less reliable in real coding workflows despite strong benchmark claims. Commenters highlight evidence of “benchmaxxing” and cherry‑picked comparisons, arguing that official scores no longer track real-world usefulness and that models like DeepSeek Flash and Qwen 3.8 Flash Next currently offer a better price‑performance and latency sweet spot. The exchange broadens into concerns about benchmark flaws, opaque marketing, and the economics of API pricing versus heavily subsidized subscriptions, underscoring growing skepticism toward incremental LLM launch hype.

Model architecture & “Flash” positioning

  • FAQ cites MiniMax M3.1 Flash as ~428B total parameters with ~23B active (MoE), roughly comparable in size/structure to DeepSeek V4.1 Flash.
  • “Flash” is used loosely across vendors; here it does not mean “small.” Ideas floated: single experts fitting cost‑efficient GPUs, MoE sparsity for cheaper inference, and a marketing label as non‑1T “regular” models become common.

Comparisons to DeepSeek, GLM, Qwen, others

  • Many commenters say it underperforms DeepSeek V4.1 Flash (especially 1017) and GLM 5.3 Flash in practical coding and SWE use.
  • Stealth deployment as “Space Bunny” in OpenCode led many users to switch back to Qwen 3.8 Flash Next or DeepSeek V4 Flash.
  • Some third‑party comparisons (e.g., aibenchy) show it clearly behind GPT‑5.6 Sol and Opus 5, contradicting MiniMax’s own charts.

Benchmarks, “benchmaxxing,” and real‑world behavior

  • Strong skepticism that MiniMax’s benchmark results reflect typical use; several call the model “benchmaxxed.”
  • Deepswe/TerminalBench scores are seen as inconsistent with users’ experiences; deepswe itself is called “flawed” and saturated.
  • General worry that many labs now overfit to benchmarks (Goodhart’s law), making scores less meaningful.

Cost, pricing, and economics

  • M3.1 Flash halves input and cache prices vs M3, making a “70% cheaper for agentic tasks” claim plausible.
  • Still slightly worse on cache cost than DeepSeek V4.1, which has unusually cheap, highly optimized KV caching.
  • Debate over API vs subscription economics: some argue heavily subsidized subscriptions (e.g., frontier labs) beat cheap APIs; others say quotas and time windows make cheap APIs like DeepSeek more usable day‑to‑day.

Speed, local usage, and caching

  • Multiple reports that MiniMax models are slow in practice (high CoT verbosity, long TTFT), versus very fast DeepSeek Flash and Qwen 3.8 Flash Next.
  • For interactive/agentic coding, decode speed and prefill performance are seen as critical; 13 tok/s is viewed as barely acceptable or unusable depending on workflow.
  • DeepSeek’s extremely compressed KV cache (and long TTL) is highlighted as a key differentiator on cost and UX.

Use‑case‑specific performance

  • Some data‑analytics benchmarks show M3.1 Flash tying Opus 5.5 and Argon in accuracy but being much slower due to excessive token usage.
  • Cybersecurity and reverse‑engineering tests reportedly favor GLM 5.3 Flash and other models; M3.1 Flash “far underperforms” there in at least one platform’s evaluations.
  • One niche adversarial‑programming benchmark finds it roughly on par with DeepSeek V4 Pro.

Trust, safety, and privacy

  • Several commenters say they’ve “lost trust” in MiniMax due to perceived benchmark inflation.
  • Broader concerns about “nerfed” frontier models over time, though others point to tracking projects suggesting performance is mostly stable.
  • Strong unease about being monitored via cloud subscriptions; skepticism toward “zero data retention” claims.
  • Discussion of secure enclave providers (e.g., TEE‑based services) and worries about in‑enclave scanning and potential data exfiltration patterns.
  • One comment notes prior DHS blacklisting of a MiniMax model and wonders if this one will be targeted too.

Meta and announcement fatigue

  • The minimax.io domain appears to have been shadowbanned on HN, causing many duplicate submissions.
  • Many express fatigue with near‑daily “cheaper, better, faster” model announcements with carefully cherry‑picked benchmarks and moving Pareto frontiers.
  • Some still see value: each new “flash” or mid‑tier model can shift the cost‑per‑task frontier, especially for non‑frontier use like local or cost‑sensitive workflows.