MiniMax M3 (NovitaAI)

A technical analysis of MiniMax-M3 describes it as a Mixture-of-Experts model with approximately 428 billion total parameters and 23 billion active per token. Its MiniMax Sparse Attention mechanism uses an index branch to rank key-value blocks independently for each GQA group, then a main branch that computes exact attention only over selected blocks. The article further says MiniMax co-designed GPU kernels for H800 hardware using exp-free Top-k selection to improve tensor-core utilization with sparse memory access. These architectural and hardware details provide implementation context for M3, while the source’s separate discussion of gateway access is promotional and outside the model-focused scope.

This launch tracker describes MiniMax M3 as a sparse Mixture-of-Experts foundation model that takes text, image, and video as input and produces text, positioned for long-horizon agent tasks, whole-repository coding, and long-video understanding. The launch framing claims M3 is the first open model pairing frontier coding and agentic performance, a one-million-token context window, and native multimodality including computer operation. The tracker notes that listed context and maximum output figures come from an OpenRouter listing rather than primary documentation, and that MiniMax declined to disclose parameter count or compute infrastructure at launch. It also points out that M3 was MiniMax's first major product release since the company began preparing for a STAR Market IPO.

MiniMax’s official page describes M3 as a native multimodal model with coding and agentic capabilities, built on MiniMax Sparse Attention. The API supports a context window of up to 1 million tokens, with at least 512,000 tokens usable, positioning the model for long-running agents, coding sessions, and video understanding. MiniMax reports that M3 scored 83.5 on BrowseComp, ahead of Opus 4.7 at 79.3. It also documents autonomous research demonstrations, including reproducing an ICLR 2025 distinguished paper over nearly 12 hours, optimizing an FP8 matrix-multiplication kernel with 147 benchmark submissions, and completing a post-training workflow without human intervention.

BenchLM's catalog entry for MiniMax M3, updated November 7, 2026, reports a capability score of 49.3 out of 100, roughly 107 tokens per second throughput, a 19.73-second first-token latency, and a 1M-token context window. Pricing is listed at $0.30 per million input tokens and $1.20 per million output tokens, with $0.60 for cached input and a blended rate of $0.75. The ranking breakdown shows MiniMax M3 at rank 6 of 157 for instruction following with a 93.8 score, rank 73 of 131 for agentic work, rank 76 of 157 for coding, rank 48 of 59 for multimodal, and rank 61 of 178 for knowledge. Coding is its weakest eligible category, reasoning and multilingual were not measured, and the page notes 28 published benchmark rows with some tracked slots empty.

MiniMax officially released M3 on June 1, 2026, describing it as the first open-weight model combining frontier coding, a 1M-token context window, and native multimodality. The architecture uses MiniMax Sparse Attention (MSA), reintroduced after being dropped in the previous M2 generation, and the model is reachable via MiniMax Code, token plans, and the API with open weights and a technical report promised shortly after launch. Launch benchmarks cited include 59.0% on SWE-Bench Pro, 66.0% on Terminal Bench 2.1, 34.8% SWE-fficiency, 28.8% KernelBench Hard, and 74.2% MCP Atlas. The write-up notes MSA is what allows context to scale to one million tokens, and it separates confirmed launch facts from figures developers should independently validate.

MiniMax M3, built by MiniMax, launched on Fireworks as a day-0 open-weight release combining long-horizon agentic coding with native multimodality. The post describes a single model that handles image and video understanding alongside autonomous coding runs lasting hours, aimed at long-horizon agents, multimodal pipelines, and full-repository coding workflows. Fireworks reports supporting up to 500K tokens of context at launch while working with the MiniMax team toward the full 1M-token window. It also cites Artificial Analysis's third-party intelligence index, saying M3 surpasses other open-source models in overall intelligence and exceeds some closed-source models. Creator attribution stays with MiniMax; Fireworks is the serving provider.

This technical reference compiles M3's architecture from the MiniMax model card and API release notes: a mixture-of-experts vision-language model with 428B total parameters and 23B active, GQA attention, three dense layers plus 57 MoE layers, and a 1M context window. It accepts text, image, and video input and emits text, published under an image-text-to-text pipeline tag. Attention specifics include MiniMax Sparse Attention with partial RoPE at rotary factor 0.5, SwiGLU-OAI feed-forward blocks, and 128 routed experts plus one shared expert with four active per token. The page notes MiniMax API release notes date the launch to June 1, 2026, while the Hugging Face repository was created June 2, 2026, though it warns this InferenceX deployment is unofficial and its data may be modified.

MiniMax describes M3 as an open-weight model combining three capabilities in one release: frontier coding and agentic work, a 1M-token context window, and native multimodality built through multimodal training from step zero. The model runs on the proprietary MiniMax Sparse Attention (MSA) architecture, and the API supports up to 1M tokens of context with a guaranteed minimum of 512K tokens. On benchmarks, the page reports a BrowseComp score of 83.5 for M3 against 79.3 for Opus 4.7, highlighting autonomous browsing and retrieval ability. It also details a long-horizon test in which M3 ran nearly 12 hours to replicate an ICLR 2025 Outstanding Paper, producing 18 commits and 23 experimental figures. The company positions M3 as usable for AI coding assistants, automated workflows, and long-video understanding.