minimax-m3.1-flash
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.
Empirical Evaluation Results
Architectural Profile & Capabilities
MiniMax-M3.1-Flash (minimax/minimax-m3.1-flash, released 19 October 2026) is the efficiency-tier successor to MiniMax-M3, tuned explicitly for agentic execution, coding, computer use, and cybersecurity while cutting online inference cost by roughly 70% and improving throughput roughly 3x versus M3. It is an open-weight sparse mixture-of-experts model with 23B active parameters against a ~428B total parameter pool, natively multimodal on the input side (text + image; some third-party listings also claim video, which should be verified against the official API spec before production use) with text-only output. The 1,048,576-token context window is one of its headline features and is corroborated by a strong AA-LCR score of 79.7% and a LOCA-Bench 256K score of 60.3. Positioning within the October 2026 frontier landscape: M3.1-Flash is not a frontier-reasoning competitor to OpenAI GPT-6 Astra, Anthropic Claude Opus 5, Google Gemini 3.8 Pro, or DeepSeek V4 Pro. Its Artificial Analysis Intelligence Index of 42.7 places it firmly in the mid-tier 'cheap agent workhorse' band — roughly the price/performance niche occupied by DeepSeek V4.1 Flash, Gemini 3.8 Flash, and Qwen 3.8. Its differentiation is not raw reasoning ceiling but cost-normalized agentic execution: DeepSWE v1.1 at 73.7%, OSWorld V2 at 61.7%, Automation-Bench v1.0.6 at 62.8, and BankerToolBench at 76.2 (reported as tying Claude Opus 5 on that narrow business-tool benchmark) at an input price an order of magnitude below flagship rates. Conversely, Terminal-Bench 4.0 at 25.8% and CritPt at 22.3% show that hard multi-step CLI/terminal orchestration and frontier physics reasoning remain clear weaknesses — AA's own comparison table places it roughly 8 points behind the 33.8% figure of at least one sibling configuration on Terminal-Bench 4.0. Architecture transparency is partial. The MoE structure, active/total parameter split, context length, and modality support are documented. However, expert count, top-k routing, attention variant (MLA vs. GQA vs. hybrid linear attention), and layer-level feed-forward design are NOT officially published in indexed sources; one secondary report attributes part of the speed gain to 'sparse attention' but this is insufficient to assert a specific attention mechanism. Treat attention internals as undisclosed rather than inferring them from throughput. Reasoning/tool-calling is supported and the model exposes a reasoning_effort control (medium is used in many published evaluation configs), meaning latency and token burn scale materially with reasoning budget. Operationally, the model is best treated as a high-concurrency agent substrate: cheap enough to run long tool loops, multimodal for screenshot-driven computer-use, and long-context for repo- or document-scale grounding. It should not be treated as an unattended high-stakes decision maker — independent coverage flags a notably high hallucination rate on knowledge-sensitive prompts (reflected in AA-Omniscience accuracy of 56.2%), and its agentic wins are concentrated in structured, verifiable tool environments rather than open-ended terminal debugging.
23B active parameters out of ~428B total parameters, giving roughly a 5.4% activation ratio per token — a very sparse configuration that explains the low serving cost ($0.15/1M input, $0.60/1M output, $0.015/1M cached input) relative to the total weight footprint. Context window is 1,048,576 tokens. Input modalities: text and image (video input is claimed by some aggregators but should be treated as unverified against the official API spec); output is text only. The exact number of experts, top-k routing configuration, and attention design (MLA, GQA, or hybrid linear/'sparse attention' as one secondary report suggests) are NOT published in official indexed material and must be recorded as undisclosed. Test-time reasoning is supported via an exposed reasoning_effort parameter (low/medium/high); published AA-derived evaluations commonly run at medium, which means reasoning-token overhead is variable and latency-sensitive workloads must pin effort explicitly. Reported relative to its predecessor M3: ~3x faster inference and ~70% lower usage cost.
Recommended Workloads & Primary Use Cases
- •Terminal/CLI orchestration is a weak spot: Terminal-Bench 4.0 lands at only 25.8% (AA run), well below sibling configurations and frontier agents — do not deploy as an autonomous shell-debugging agent without a stronger supervisor model.
- •High hallucination/knowledge-fidelity risk: AA-Omniscience Accuracy is 56.2% and independent coverage explicitly flags a high hallucination rate. Never use it as an unattended source of truth for factual claims; gate all knowledge-dependent output behind retrieval or a verifier.
- •Reasoning budget is a hidden latency and cost multiplier: reasoning_effort (commonly 'medium' in published evals) can swing output tokens several-fold, so TTFT and cost per task are not well predicted by the measured ~81 tps output rate alone. Pin effort per workload and benchmark on your own traces.
- •Context length ≠ reliable recall at depth: the 1M window is supported, but AA-LCR 79.7% and LOCA-Bench 256K 60.3% indicate real degradation well before the limit. Treat ~256K as the practical working depth and use retrieval/compaction beyond that; long prefill also inflates TTFT substantially.
- •Architecture and normalization opacity: attention variant, expert count, and routing are undisclosed, and it is an open-weight release — self-hosted quantization will diverge from the API's benchmark behavior. Also verify image/video modality support and any tool-call schema constraints (parallel tool calls, strict JSON mode) against the current official API spec, since aggregator listings disagree.
Calls to minimax-m3.1-flash 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 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%.
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.