minimax-m3
MiniMax-M3 is a natively multimodal sparse Mixture-of-Experts foundation model (~428B total / ~22–23B active parameters) from MiniMax, built for frontier coding, long-horizon agentic work, and 1M-token context tasks. It combines Grouped-Query Attention with a blockwise MiniMax Sparse Attention scheme, accepts text/image/video input with text output, and runs at a measured ~97 tokens/sec. Additive exact-variant scores verified here include SWE-bench Verified 80.5%, SWE-bench Pro 59.0%, SWE-fficiency 34.8%, KernelBench Hard 28.8%, SAGE 50.6%, and SkillsBench 51.5%.
Empirical Evaluation Results
Architectural Profile & Capabilities
MiniMax-M3 (OpenRouter slug minimax/minimax-m3, Benchmark Heaven ID minimax-m3::default) is MiniMax's frontier multimodal foundation model, officially released and documented via the MiniMax launch page, NVIDIA NIM model card, NVIDIA Megatron-Bridge configs, and the MiniMaxAI/MiniMax-M3 Hugging Face repository. It is a sparse Mixture-of-Experts (MoE) transformer rather than a dense model or diffusion LM: reported totals are approximately 428B parameters with roughly 22–23B active parameters per token, a hidden size of 6,144, and a routing configuration of 128 routed experts with four routed experts plus a shared expert activated per token. Attention is Grouped-Query Attention (64 query heads, 4 key/value heads, 128-dim head size) extended with a blockwise 'MiniMax Sparse Attention' mechanism used to make the 1M-token context window tractable; sources do not establish MLA or hybrid linear attention for this variant. The model is natively multimodal (text, image, and video input; text output) and is explicitly positioned by MiniMax for long-horizon agentic work, coding, and tool-use harnesses. On exact-variant evidence, MiniMax's own launch page reports SWE-Bench Pro at 59.0%, Terminal-Bench 2.1 at 66.0%, SWE-fficiency at 34.8%, and KernelBench Hard at 28.8%. Independent third-party evaluators add SWE-bench Verified at 80.5% (Vals AI / Morph), Terminal Bench 2.1 at 53.56% on the Vals AI harness, SAGE at 50.6%, and SkillsBench at 51.5%. These are appended to the immutable Benchmark Heaven ground-truth baseline (which already records GPQA Diamond 92.9%, HLE (no tools) 39.0%, AA Coding Index 58.6%, TAU-2 Telecom 88.9%, Terminal-Bench 2.1 65.2%, Terminal-Bench 4.0 42.4%, and others). Note the deliberate distinction between benchmark variants: Terminal-Bench 2.1, Terminal-Bench Hard, and Terminal-Bench 4.0 are separate suites and must not be conflated, and SWE-bench Pro must not be merged with SWE-bench Verified. Discrepancies between MiniMax's own Terminal-Bench 2.1 figure (66.0%) and the Artificial Analysis/OpenRouter listing (65.2%) and the Vals AI harness (53.56%) reflect differing scaffolds, harness settings, timeouts, and evaluation snapshots rather than a single canonical number. Operationally, M3 sits at the high-throughput end of the reasoning-model class at a measured ~97 tokens/sec on Benchmark Heaven telemetry, with hybrid reasoning behavior (models 'think deeply' and hold context across 50+ turns without drift per practitioner reports). It is priced aggressively at $0.3/1M input and $1.2/1M output, roughly an order of magnitude cheaper per SWE-bench Pro point than GPT-5.5-class proprietary models, making it a strong cost-efficiency play for agentic coding and long-context document workflows. Its relative weaknesses in the ground-truth record are on very-long-horizon terminal autonomy (Long Horizon Terminal Bench 1.0 at 0.4), extreme reasoning (Critpt 3.7%, Chess Puzzles 14.0%), and frontier knowledge recall (AA-Omniscience Index 1.4), so teams should scope deployments toward coding, multimodal document understanding, and mid-horizon agentic tasks rather than deep scientific discovery or marathon autonomous terminal sessions.
~428B total parameters with ~22–23B active parameters per token; hidden size 6,144; 128 routed experts with 4 routed experts + 1 shared expert activated per forward pass. Attention is Grouped-Query Attention (GQA: 64 query heads, 4 KV heads, 128-dim head size) augmented with MiniMax's blockwise Sparse Attention to enable the 1M-token context window. Sources do not confirm MLA or hybrid linear attention for this variant. Native multimodal input (text, image, video) with text-only output; supports extended test-time reasoning and long-horizon tool use.
Recommended Workloads & Primary Use Cases
- •Long-horizon autonomous terminal work degrades sharply: Long Horizon Terminal Bench (1.0) scores just 0.4 despite strong single-session Terminal-Bench 2.1 results — keep sessions bounded and checkpoint aggressively.
- •Context quality erodes well before the 1M ceiling: Context Arena MRCR V2 (8-needle) is 53.1% and Mlcr is 17.2%, so retrieval-heavy prompts should be chunked with explicit re-grounding rather than relying on raw window size.
- •Frontier knowledge recall and hard science reasoning are weak relative to its coding strength (AA-Omniscience Index 1.4, Critpt 3.7%, SciCode 47.1%, Mystery Game Puzzles 7.0%) — do not treat it as a scientific-reasoning or open-domain authority model.
- •Benchmark variant ambiguity is real and must be managed: official Terminal-Bench 2.1 (66.0%), Artificial Analysis (65.2%), and Vals AI (53.56%) diverge due to scaffold/harness differences, and SWE-bench Pro (59.0%) is not SWE-bench Verified (80.5%) — pin your exact harness when reporting.
- •Hybrid reasoning behavior can inflate latency and cost via long thinking traces; first-token latency on large prompts and throughput during reasoning phases will be materially slower than the ~97 tps headline, and prompt/tool-call formatting should follow MiniMax's documented chat template rather than being inferred from OpenAI-style payloads.
Calls to minimax-m3 are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by MiniMax.
More Models from MiniMax
Compare other engines available in this laboratory.
MiniMax-M3.1-Flash is a low-cost, open-weight sparse MoE multimodal model (23B active / ~428B total params) with a 1,048,576-token context, built for high-throughput agentic, coding, computer-use, and cybersecurity workloads at $0.15/$0.60 per 1M tokens.
PinchBench: 86.2% (5th overall, within 1.2pts of Opus 4.6), SWE-Pro: 56.22%, GDPval-AA: 1495 Elo (highest open-source), MLE Bench Lite: 66.6%, Terminal Bench 2: 57.0% | Agentic coding, full-project delivery, complex engineering systems, autonomous debugging, production incidents, office productivity.
SWE-bench: 80.2% — rivals Claude Opus and GPT-5.2 | Coding, debugging, refactoring, full-stack development, front-end components, test coverage, technical implementation.