Be part of the occasion trusted by enterprise leaders for practically twenty years. VB Rework brings collectively the folks constructing actual enterprise AI technique. Study extra
Chinese language AI startup MiniMax, maybe greatest recognized within the West for its hit lifelike AI video mannequin Hailuo, has launched its newest giant language mannequin, MiniMax-M1 — and in nice information for enterprises and builders, it’s utterly open supply beneath an Apache 2.0 license, that means companies can take it and use it for business functions and modify it to their liking with out restriction or cost.
M1 is an open-weight providing that units new requirements in long-context reasoning, agentic device use, and environment friendly compute efficiency. It’s out there at the moment on the AI code sharing group Hugging Face and Microsoft’s rival code sharing group GitHub, the primary launch of what the corporate dubbed as “MiniMaxWeek” from its social account on X — with additional product bulletins anticipated.
MiniMax-M1 distinguishes itself with a context window of 1 million enter tokens and as much as 80,000 tokens in output, positioning it as some of the expansive fashions out there for long-context reasoning duties.
The “context window” in giant language fashions (LLMs) refers back to the most variety of tokens the mannequin can course of at one time — together with each enter and output. Tokens are the fundamental models of textual content, which can embrace whole phrases, elements of phrases, punctuation marks, or code symbols. These tokens are transformed into numerical vectors that the mannequin makes use of to symbolize and manipulate that means via its parameters (weights and biases). They’re, in essence, the LLM’s native language.
For comparability, OpenAI’s GPT-4o has a context window of solely 128,000 tokens — sufficient to alternate a couple of novel’s price of knowledge between the person and the mannequin in a single forwards and backwards interplay. At 1 million tokens, MiniMax-M1 may alternate a small assortment or e book collection’ price of knowledge. Google Gemini 2.5 Professional gives a token context higher restrict of 1 million, as nicely, with a reported 2 million window within the works.
However M1 has one other trick up its sleeve: it’s been skilled utilizing reinforcement studying in an revolutionary, resourceful, extremely environment friendly approach. The mannequin is skilled utilizing a hybrid Combination-of-Consultants (MoE) structure with a lightning consideration mechanism designed to cut back inference prices.
In response to the technical report, MiniMax-M1 consumes solely 25% of the floating level operations (FLOPs) required by DeepSeek R1 at a technology size of 100,000 tokens.
Structure and variants
The mannequin is available in two variants—MiniMax-M1-40k and MiniMax-M1-80k—referring to their “considering budgets” or output lengths.
The structure is constructed on the corporate’s earlier MiniMax-Textual content-01 basis and contains 456 billion parameters, with 45.9 billion activated per token.
A standout function of the discharge is the mannequin’s coaching price. MiniMax studies that the M1 mannequin was skilled utilizing large-scale reinforcement studying (RL) at an effectivity not often seen on this area, with a complete price of $534,700.
This effectivity is credited to a customized RL algorithm referred to as CISPO, which clips significance sampling weights moderately than token updates, and to the hybrid consideration design that helps streamline scaling.
That’s an astonishingly “low-cost” quantity for a frontier LLM, as DeepSeek skilled its hit R1 reasoning mannequin at a reported price of $5-$6 million, whereas the coaching price of OpenAIs’ GPT-4 — a greater than two-year-old mannequin now — was mentioned to exceed $100 million. This price comes from each the value of graphics processing models (GPUs), the massively parallel computing {hardware} primarily manufactured by corporations like Nvidia, which may price $20,000–$30,000 or extra per module, and from the vitality required to run these chips repeatedly in large-scale information facilities.
Benchmark efficiency
MiniMax-M1 has been evaluated throughout a collection of established benchmarks that take a look at superior reasoning, software program engineering, and tool-use capabilities.
On AIME 2024, a arithmetic competitors benchmark, the M1-80k mannequin scores 86.0% accuracy. It additionally delivers sturdy efficiency in coding and long-context duties, attaining:
- 65.0% on LiveCodeBench
- 56.0% on SWE-bench Verified
- 62.8% on TAU-bench
- 73.4% on OpenAI MRCR (4-needle model)
These outcomes place MiniMax-M1 forward of different open-weight opponents equivalent to DeepSeek-R1 and Qwen3-235B-A22B on a number of advanced duties.
Whereas closed-weight fashions like OpenAI’s o3 and Gemini 2.5 Professional nonetheless high some benchmarks, MiniMax-M1 narrows the efficiency hole significantly whereas remaining freely accessible beneath an Apache-2.0 license.
For deployment, MiniMax recommends vLLM because the serving backend, citing its optimization for giant mannequin workloads, reminiscence effectivity, and batch request dealing with. The corporate additionally gives deployment choices utilizing the Transformers library.
MiniMax-M1 contains structured perform calling capabilities and is packaged with a chatbot API that includes on-line search, video and picture technology, speech synthesis, and voice cloning instruments. These options purpose to help broader agentic conduct in real-world functions.
Implications for technical decision-makers and enterprise patrons
MiniMax-M1’s open entry, long-context capabilities, and compute effectivity handle a number of recurring challenges for technical professionals answerable for managing AI techniques at scale.
For engineering leads answerable for the total lifecycle of LLMs — equivalent to optimizing mannequin efficiency and deploying beneath tight timelines — MiniMax-M1 gives a decrease operational price profile whereas supporting superior reasoning duties. Its lengthy context window may considerably scale back preprocessing efforts for enterprise paperwork or log information that span tens or lots of of hundreds of tokens.
For these managing AI orchestration pipelines, the power to fine-tune and deploy MiniMax-M1 utilizing established instruments like vLLM or Transformers helps simpler integration into present infrastructure. The hybrid-attention structure might assist simplify scaling methods, and the mannequin’s aggressive efficiency on multi-step reasoning and software program engineering benchmarks gives a high-capability base for inner copilots or agent-based techniques.
From a knowledge platform perspective, groups answerable for sustaining environment friendly, scalable infrastructure can profit from M1’s help for structured perform calling and its compatibility with automated pipelines. Its open-source nature permits groups to tailor efficiency to their stack with out vendor lock-in.
Safety leads might also discover worth in evaluating M1’s potential for safe, on-premises deployment of a high-capability mannequin that doesn’t depend on transmitting delicate information to third-party endpoints.
Taken collectively, MiniMax-M1 presents a versatile choice for organizations seeking to experiment with or scale up superior AI capabilities whereas managing prices, staying inside operational limits, and avoiding proprietary constraints.
The discharge alerts MiniMax’s continued deal with sensible, scalable AI fashions. By combining open entry with superior structure and compute effectivity, MiniMax-M1 might function a foundational mannequin for builders constructing next-generation functions that require each reasoning depth and long-range enter understanding.
We’ll be monitoring MiniMax’s different releases all through the week. Keep tuned!
