MiniMax Weekly Intelligence Report
Week of 15–21 March 2026 | Published: 21 March 2026
Executive Summary
MiniMax entered this week as one of the most closely watched AI companies in global markets — and proceeded to justify that attention on every front. Three converging storylines define the company’s seven days: the landmark release of M2.7, the world’s first publicly deployed “self-evolving” large language model; a record-breaking stock surge of up to 29% driven by the OpenClaw agentic AI boom and an endorsement from Nvidia CEO Jensen Huang; and the ongoing digestion of FY2025 earnings that confirmed 159% revenue growth and gross margin expansion from 12% to 25%.
The M2.7 release is the defining event — technically, strategically, and narratively. MiniMax claims its new model handled 30–50% of its own reinforcement learning development workflow, running over 100 autonomous improvement cycles with no human intervention. The company also open-sourced OpenRoom, an interactive GUI-based agent experience framework. Together, these moves position MiniMax as the first Chinese AI lab — and one of the first globally — to publicly deploy a model with verifiable recursive self-improvement in a production environment.
The week’s developments are not without caveats: benchmark scores are self-reported, the self-evolution claims await independent verification, and regulatory headwinds around OpenClaw-integrated AI agents are beginning to materialise in Beijing. Nonetheless, MiniMax’s trajectory — from post-IPO momentum to frontier model developer — is accelerating at a pace few industry observers predicted at this stage of the company’s public market lifecycle.
Story 1: M2.7 — The Self-Evolving Model
Product Launch
On 18 March 2026, MiniMax released M2.7, a new proprietary LLM designed to power AI agents and serve as the backend for third-party harnesses and tools such as Claude Code, Kilo Code, and OpenClaw. The launch marks a qualitative shift in the M-series roadmap: where M2.5 (February 2026) was positioned as a coding and productivity leader, M2.7 is positioned as an autonomous engineering partner — a model that does not merely respond to instructions, but participates in building and iterating the systems that improve it.
M2.7 is MiniMax’s first model deeply embedded in its own evolutionary cycle. The company tasked an internal version of M2.7 to build a research agent harness that interacts and collaborates with different research project groups, supporting data pipelines, training environments, infrastructure, cross-team collaboration, and persistent memory.
MiniMax released M2.7 on March 18, a proprietary model the company says actively participated in its own reinforcement learning process. Per the announcement, the model built dozens of complex skills within its RL harness, updated its own memory, and optimised its training pipeline based on experiment results — handling 30–50% of the research workflow that previously required multiple human researchers.
Strategic Context
The self-evolution narrative is compelling and strategically deliberate. In an increasingly commoditised LLM landscape — where pricing has raced to the floor and benchmark differentiation is narrowing — MiniMax has staked out a distinct positioning: the model that builds itself. This is a meaningful branding and technical claim that distances M2.7 from pure benchmark competition with GPT-5.x or Gemini 3.x.
MiniMax becomes the second Chinese startup to release a proprietary cutting-edge LLM in recent months, following z.ai with its GLM-5 Turbo, amid reports that Alibaba’s Qwen team is also shifting to proprietary development. The broader trend — Chinese labs moving from open-weight to proprietary releases — signals growing confidence in commercial monetisation and reduced reliance on community adoption as the primary growth vector.
MiniMax’s M2.7 update was released ahead of a next-generation M3 expected in the second half of 2026, suggesting the company is maintaining a rapid cadence of major model releases.
Market Impact
MiniMax Group, which had just released a new self-evolving agent, extended earlier gains to climb as much as 29% to a record high during the week, buoyed by the dual catalyst of M2.7’s launch and Jensen Huang’s bullish commentary on the OpenClaw ecosystem. The market is clearly treating MiniMax’s agent-native positioning as a premium valuation driver.
Capital has poured into Hong Kong–listed MiniMax over the past week. The buying spree sent its share price soaring from around HK$800, hitting an intraday peak of HK$1,320, with multiple daily jumps exceeding 20%, marking the highest level since the company’s IPO.
For enterprise buyers, the pricing signal is significant: M2.7 is priced at $0.30 per million input tokens and $1.20 per million output, unchanged from M2.5. At a price point approximately 16x cheaper than Claude Opus 4.6 on input tokens while claiming comparable SWE-Pro performance, MiniMax is directly targeting cost-sensitive enterprise agent deployments.
Tech Angle
M2.7 supports a 204,800-token context window and runs at approximately 60 TPS in standard configuration, with a MiniMax-M2.7-highspeed variant reaching approximately 100 TPS at higher throughput cost.
Key benchmark claims (all company-reported, pending independent verification):
- In the SWE-Pro benchmark, M2.7 scored 56.22%, nearly matching the best performance of Anthropic’s Claude Opus, MiniMax said. In professional office work, its ELO score on GDPval-AA ranked just behind Claude Opus 4.6, Claude Sonnet 4.6, and OpenAI’s GPT-5.4, while surpassing GPT-5.3.
- In a separate test, MiniMax ran M2.7 through 22 MLE Bench Lite machine learning competitions with 24-hour autonomous runs. The best attempt earned 9 gold, 5 silver, and 1 bronze medal, averaging a 66.6% medal rate across three trials — tying Google’s Gemini 3.1 but sitting behind Opus 4.6 at 75.7%.
- Notable technical details include multi-agent collaboration, high skill adherence rates (97% across 40+ complex skills), and performance scores near the industry’s best on benchmarks such as SWE-Pro (56.22%) and VIBE-Pro (55.6%).
One important caveat: not all independent, third-party benchmarks show improvement for M2.7 over M2.5. On BridgeBench, M2.5 scored 12th place while M2.7 scored 19th place — a regression that underscores the risk of over-relying on company-curated evaluations.
Alongside M2.7, MiniMax built and open-sourced OpenRoom, an interaction system based on an agent harness that places AI interaction within a Web GUI space where everything is interactive, with characters proactively engaging with their environment — most of the code written by AI.
Source: VentureBeat — https://venturebeat.com/technology/new-minimax-m2-7-proprietary-ai-model-is-self-evolving-and-can-perform-30-50 Source: MiniMax Official Blog — https://www.minimax.io/news/minimax-m27-en Source: CnTechPost — https://cntechpost.com/2026/03/18/minimax-releases-next-gen-ai-model-m2-7-self-evolution-capabilities/
Story 2: The OpenClaw Rally — MiniMax as China’s Agentic AI Proxy
Strategic Context
The OpenClaw phenomenon has become the defining market narrative for Chinese AI stocks in Q1 2026, and MiniMax is its most leveraged pure-play beneficiary. OpenClaw is an agent that uses large language models to do things like hail a ride and book restaurants, going beyond traditional chatbots by enabling software agents to complete tasks, make decisions, and act with limited user input.
Jensen Huang on Tuesday said that OpenClaw was “definitely the next ChatGPT,” calling it a transformative step in expanding what users can achieve with AI. MiniMax and Zhipu are among China’s rising “AI tigers” — companies building large language models to rival the likes of OpenAI and Anthropic — and both have been ramping up their agentic AI offerings with tools built on OpenClaw.
MiniMax’s own OpenClaw spinoff — MaxClaw — is now part of a competitive cluster that includes Tencent’s WorkBuddy, MiniMax’s MaxClaw, and MoonShot’s Kimi Claw, as local governments joined in with grants of up to 10 million yuan for “one-person companies” building OpenClaw-powered applications.
Market Impact
MiniMax shares have climbed more than 600% from its IPO earlier this year, with the stock hitting an intraday peak of HK$1,320. The company is now unambiguously the premier vehicle for investors seeking exposure to the agentic AI wave in China’s public equity markets.
However, the rally has also introduced new risks. Authorities in Beijing have started to limit the use of OpenClaw within government agencies and some of the country’s largest banks, suggesting policymakers are monitoring the technology’s rapid adoption. This regulatory caution could temper institutional adoption just as MiniMax’s M2.7 is architected to maximise OpenClaw workflow integration.
Source: CNBC — https://www.cnbc.com/2026/03/18/china-ai-zhipu-minimax-after-nvidia-jensen-huang-openclaw-comments.html Source: Bloomberg — https://www.bloomberg.com/news/articles/2026-03-18/china-s-openclaw-stocks-rise-as-nvidia-calls-it-the-next-chatgpt Source: Fortune — https://fortune.com/2026/03/14/openclaw-china-ai-agent-boom-open-source-lobster-craze-minimax-qwen/
Story 3: FY2025 Earnings — The Underlying Growth Story
Strategic Context
Published 2 March and still reverberating through analyst models this week, MiniMax’s first full-year results as a public company confirmed the growth narrative investors had priced in at IPO — and then some.
Total revenue increased by 158.9% year over year to US$79.0 million, with more than 70% derived from international markets. Gross profit increased by 437.2% year over year to US$20.1 million. Gross profit margin was 25.4%, an improvement of 13.2 percentage points compared to the same period of 2024.
As of end-2025, MiniMax had cumulatively served more than 236 million users across over 200 countries and regions, as well as 214,000 enterprise customers and developers from more than 100 countries.
A strategically important disclosure: internally, MiniMax’s agent interns now support nearly 90% of employees, with use cases spanning software development, data analysis, operations management, talent recruitment, and sales and marketing — making MiniMax a live testing ground for AI-native organisational capabilities. In January 2026, this was productised as the MiniMax Agent AI-native workspace.
Market Impact
MiniMax’s shares surged as much as 21% in its first earnings since a landmark debut in Hong Kong. The results underscore solid demand for low-cost, open-source Chinese models that rival the best of Silicon Valley. The adjusted net loss of US$250.9 million remains a concern — the company is not yet profitable — but the trajectory of margin expansion and the accelerating token consumption (M2 series daily token consumption grew 6x between December 2025 and February 2026) suggests the economics are moving in the right direction.
Source: MiniMax IR / PRNewswire — https://www.minimax.io/news/minimax-global-announces-full-year-2025-financial-results Source: Bloomberg — https://www.bloomberg.com/news/articles/2026-03-02/china-ai-pioneer-minimax-more-than-doubles-sales-in-hot-market
Strategic Outlook: What to Watch
| Theme | Near-Term Signal | Risk / Opportunity |
|---|---|---|
| M2.7 Independent Verification | Third-party SWE-Pro and VIBE-Pro audits | Self-reported benchmarks; BridgeBench regression is a red flag to watch |
| M3 Release Timeline | H2 2026 target confirmed by multiple sources | Maintaining 90-day major model cadence is operationally demanding |
| OpenClaw Regulatory Risk | Beijing limiting OpenClaw in banks and agencies | Could blunt enterprise adoption of MaxClaw/M2.7 integrations domestically |
| OpenRoom Ecosystem | Open-source community traction on GitHub | High-extensibility framework; community adoption velocity is key proof point |
| Path to Profitability | Token consumption growing 6x; gross margin at 25.4% | Adjusted net loss still US$250.9M; monetisation of agent workloads is the unlock |
| Weight Release for M2.7 | Possible after compliance review (per company) | Open-weight release would dramatically accelerate developer adoption globally |
Summary Scorecard
| Dimension | Status | Signal |
|---|---|---|
| Model Innovation | ✅ Leading | M2.7 self-evolution is a genuine architectural milestone |
| Market Position | ✅ Strong | 600%+ stock gain from IPO; record highs mid-week |
| Revenue Growth | ✅ Exceptional | 159% YoY; 70%+ international; gross margin doubling |
| Profitability | ⚠️ In Progress | US$250.9M adjusted net loss; margin trajectory positive |
| Regulatory Risk | 🔴 Elevated | OpenClaw restrictions in Beijing; geopolitical uncertainty |
| Benchmark Credibility | ⚠️ Mixed | SWE-Pro claims compelling; BridgeBench regression and self-reporting caveat |
