---
title: "Bearish calls overdone: MiniMax isn't that dire"
type: "Topics"
locale: "en"
url: "https://longbridge.com/en/dolphin/post/43609158.md"
description: "China’s independent model developer $MINIMAX-W(00100.HK) released its H1 results after the HK close on Aug 26. The company has been gradually losing ground in 2026.Using Mar and Jun as key checkpoints, MiniMax has moved almost one-way down. Its market cap fell from $55bn to $13.5bn.With the actual numbers now out, is it really that bad? Let’s take a closer look (for the earnings call Trans, click here): 1) Revenue: in a breakout year for coding..."
datetime: "2026-08-26T14:31:57.000Z"
locales:
  - [en](https://longbridge.com/en/dolphin/post/43609158.md)
  - [zh-CN](https://longbridge.com/zh-CN/dolphin/post/43609158.md)
  - [zh-HK](https://longbridge.com/zh-HK/dolphin/post/43609158.md)
author: "[Dolphin Research](https://longbridge.com/en/dolphin.md)"
generator: "portal-rs"
---

# Bearish calls overdone: MiniMax isn't that dire

China’s independent AI model vendor $MINIMAX-W(00100.HK), which has steadily fallen behind since 2026, released its H1 results post-close on Aug 26. Marking Mar and Jun as key inflection points, MiniMax slid from $55bn to $13.5bn in market cap. With the actual print out, is it really that bad?

Let’s dive into the details (for the earnings call Trans, please click [here)](https://longbridge.cn/zh-CN/dolphin/post/43606351?channel=SH000001&invite-code=032064&app_id=longbridge&utm_source=longbridge_app_share&locale=zh-CN&share_track_id=78d31dfb-16e8-4bca-a7f7-e32b6558a508https://longbridge.cn/zh-CN/dolphin/post/43606351?channel=SH000001&invite-code=032064&app_id=longbridge&utm_source=longbridge_app_share&locale=zh-CN&share_track_id=78d31dfb-16e8-4bca-a7f7-e32b6558a508):

**1) Revenue: In a breakout year for Coding, consumer AI apps are pivoting — B-side now contributes over 60%.**

MiniMax, which used to monetize AI via Intl consumer channels, saw a structural shift in H1 2026 revenue mix. Key data points:

**1\. Total revenue:** $117mn, with YoY growth accelerating to 283%.

**2\. Products:** The AI app suite (Minimax, video generator Hailuo AI, audio generator Minimax Audio, Talkie — mainly Talkie and Hailuo) delivered $43mn, up 101% YoY. This was faster vs. Q4’s 82%, largely driven by Hailuo.

**3\. Enterprise services:** Model-as-a-service for enterprises — selling multi-modal API access and Coding Plans — generated $74mn. YoY growth accelerated further to 703% vs. nearly 300% in Q4 last year.

Street expectations were mostly below $70mn. With M3’s failure already priced in, this print is actually decent.

**c. Key disclosures from management:**

-   **ARR in Aug exceeded $800mn; 80% from enterprise users (vs. 30% last year).**
-   Q2 revenue rose 81.8% QoQ vs. Q1.
-   Token consumption in Jul hit 20x Jan levels, driven by post-CNY agent usage and M3’s multi-modal capabilities.
-   Token growth also came from new use cases; interactions evolved from human-agent to agent-vs.-agent (much faster than human-agent), with per-user consumption rising rapidly.

Note the company’s full-year ARR target is $1bn. M3 and agent interactions are rapidly burning tokens, pushing Aug ARR past $800mn, vs. sell-side estimates around $600mn amid heavy skepticism.

Also, the new video model H3 only launched in late Jul and has earned solid feedback in video generation. With Keling’s ARR seemingly stalling, H3 could capture more token consumption ahead.

**2) Overseas revenue share fell to 60%+**

MiniMax’s other identity as a Chinese independent model vendor has been monetizing AI apps overseas. As B-side’s mix rose, overseas share slid to 61%.

Given the visible drop in product revenue driven by overseas AI apps, it is still unclear whether overseas’ contribution within single API + coding B-side services is rising or falling. Even within model-as-a-service, Dolphin Research believes MiniMax will inevitably keep pushing Intl markets.

**4) How much do current revenues cover prior-gen training? Coding upcycle lifts all models.**

As base models refresh annually, a year’s training spend effectively buys one year of service life. Model economics can be gauged by comparing same-year direct/indirect revenues vs. prior-year training investment.

The B-side Coding boom has improved payback even for MiniMax, whose coding execution lags. Current revenues recover a much larger share of the prior-gen training bill.

Using total R&D in H1 last year (mainly training + R&D headcount) as a rough proxy for 2026 training spend, H1 2026 recognized revenue is about 94% of H1 2025 training cost. Training payback is no longer the core issue.

**3) Enterprise services: API price war crushes MiniMax’s GPM**

Historically, enterprise services carried higher GPM (token inference cloud cost vs. MiniMax’s token pricing). Most enterprise API integrations were paid, with overall GPM near ~70% in the first three quarters of last year.

In H1 2026, even as B-side mix surged, company GPM fell to 18%. That’s down from 30% in Q4, failing to hold even the 20% level.

Dolphin’s estimate suggests it may only be around 25%. With model capability not strong enough and DeepSeek offering extreme value-for-money, MiniMax is squeezed in the middle; B-side scales, but margins materially deteriorate in the API price war.

Token inference is the primary cost driver. Unit compute cost per token had been falling industry-wide; compute inflation is more of an H2 2026 issue.

In other words, benefits from lower token unit cost and higher consumption are captured by SOTA vendors like Claude. Mid-tier models mainly trade price for volume, leaving MiniMax with rising revenue but falling GPM.

Compounding this, the flagship M3 launched in H1 failed to keep pace in the crucial 2026 coding race. Pricing also stumbled — without sufficient communication, the unit switched from per-call to per-token, reducing transparency and poorly bridging legacy entitlements.

Prices were raised only to be cut permanently by 50% shortly after. Capability gains did not translate into pricing power, reflecting both lower smart premium vs. peers and pricing execution missteps.

After the raise-then-slash saga, the new model did not receive extra pricing credit for longer context windows. Since this pricing reset happened after Jun 15, H2 will hinge on whether API price cuts or domesticized compute cost reductions win out.

Management is confident GPM will improve in H2. They guided GPM to trend up in H2 and beyond 2026.

**4) Losses: absolute widened, loss ratio narrowed — not bad**

With less than $120mn in H1 revenue, operating loss (ex. financial asset revaluation and D&A) was $330mn, doubling YoY in absolute terms. The loss ratio narrowed to ~300% of revenue, which is reasonable for this stage.

Large models appear to have decent GP on paper, but the biggest investment — training — sits in R&D. R&D often runs at 3–5x revenue; as long as rapid training iterations continue, breaking even is tough ([click for reasons](https://longbridge.cn/zh-CN/topics/38121039?channel=SH000001&invite-code=032064&app_id=longbridge&utm_source=longbridge_app_share&locale=zh-CN&share_track_id=f43b7d6b-1801-4315-b9aa-dc5d8bc44318)).

H1 MiniMax R&D (mostly training) surged to $300mn, or 2.6x revenue. Sales expense fell 18% YoY, but G&A doubled (larger management team and external services), resulting in a $330mn operating loss.

**Dolphin Research’s overall take**

Since the last earnings, Hong Kong’s model duo MiniMax and Zhipu diverged. Zhipu, backed by domestic coding leadership, still holds $60bn+ in market cap even after a pullback, while MiniMax is at $13.5bn — below private-market Keling’s $18bn valuation.

Behind the slump, Dolphin Research sees four resonating factors:

**1) Core issue — M3 underperformed in real-world use**

Pricing power depends on [intelligent scarcity](https://longbridge.com/zh-CN/dolphin/post/42052740), while monetization this cycle centers on B-side coding. MiniMax and OpenAI represent C-side models, Anthropic and Zhipu the B-side approach.

Against this backdrop, MiniMax’s M3 released on Jun 1 emphasized native multi-modality during pretraining. Based on Dolphin’s checks, text accounted for ~15–20%, with most data as mixed image-text, charts, web screenshots, and document-like sources.

In post-training task execution (coding focus), success rates lagged SOTA and trials were excessive. Coding this year stresses 1M-long context task decomposition and agent collaboration, producer–validator adversarial loops to iteratively generate and fix code, and human-like GUI operations.

**2) M3 iterations fell behind**

More importantly, M3 as the flagship fell behind on benchmarks and iterated too slowly. Zhipu’s GLM 5.2 launched 12 days later debuted near global No.3 and China No.1, still firmly second-tier today, while M3 has clearly dropped back.

M3’s parameters are small — total 428bn, with 23bn activated — vs. GLM-5.2 (744bn/40bn), DeepSeek V4 (1.6tn/49bn), and Kimi/Qwen (each >2tn). Releasing such a lightweight flagship in Jun stood out.

The gap widened as Zhipu’s GLM 5.3 on Aug 14 again reached the top cohort, while MiniMax’s next-gen model has yet to launch. Management admitted a slower cadence due to pursuing a multi-model route within M3 (M3 language + H3 video) and expanding a self-operated compute cluster in Q1 2026.

**3) M3 launch pricing missteps**

Given the multi-modal training tilt, coding capability was weaker than peers. On launch, pricing misfired: vs. M2.7, M3 doubled price, and subscriptions rose sharply.

Without warning, the Coding Plan (rate-limited, no monthly token cap) was replaced with a Token Plan (Plus at RMB 49/mo = 600mn tokens ≈ 12,000 calls), upsetting legacy subscribers. Heavy users found plans insufficient, sparking controversy in the developer community.

**4) To make matters worse: financing + lock-up expiry**

Cornerstone and pre-IPO investors’ lock-up expired on Jul 8, expanding free float 10x. It rose from 5.44% to 54.38%, increasing selling pressure.

In Jul, the company raised about $2.1bn (HK$16bn) via placement and CB. The placement equals ~11% of total shares, and a fully converted CB adds ~6%, for up to 17% dilution.

Model iteration and ARR acceleration must both validate to offset dilution. Dual proof points are required.

These 1)–4) factors resonated, and the market extrapolated a temporary lag into a structural deficit, pricing MiniMax as falling from the first to the second tier. The stock kept sliding.

**This time, the company disclosed Aug ARR above $800mn, implying only ~17x PS — pessimism looks overdone. The market had expected its $1bn annual ARR target to be out of reach.**

Dolphin estimates ARR rose from $400mn in May to $800mn in Aug, implying 25%+ MoM. Linear extrapolation suggests $1bn by year-end is almost a matter of time.

**With three catalysts below potentially resonating in H2, the odds of upside are not low. If ARR reaches $1.5bn by year-end, at 25–30x PS, MiniMax could plausibly double from today’s depressed expectations.**

**Bottom Line — model competition is about staggered leadership, not a single model staying SOTA or a structural lag. Linear pricing off this mindset creates both long and short opportunities.**

**Dolphin Research’s key H2 upside catalysts to watch:**

**1) Multi-modal investment: upside optionality**

MiniMax lagged in coding partly because resources were diverted to multi-modal, with little visible payoff until H3 video launched in late Jul. In video, ByteDance’s Seedance 2.0 first reached SOTA, followed by H3 and Seedance 2.5 in late Jul.

H3’s edge is value-for-money; capability is slightly behind ByteDance but ahead of Keling and Google’s Veo. Given Keling’s current monthly revenue of ~$42mn, H3 likely has room to ramp.

That said, meaningfully capturing big-tech video model revenues won’t be easy. Both CSPs have broad app ecosystems; with weaker distribution, MiniMax may need more time.

**2) M3.1: can coding narrow the gap?**

Per news flow, M3.1 focuses on post-training, effectively patching coding. It should launch in Aug–Sep; watch whether coding meaningfully improves.

**3) Flagship: trillion-parameter Minimax M3pro incoming?**

Per pipeline, M3pro is slated for Sep–Oct, MiniMax’s first trillion-parameter model. Total params jump from 428bn to 2.7tn, with 60bn activated, and the company is reportedly researching 5tn scale.

Surprisingly, management said the core pursuit is cost efficiency rather than SOTA. Communication in Mar and May targeted ‘Opus-level capabilities, entering the global first tier,’ but the stance has shifted toward pricing, cost efficiency, and inference speed as key user drivers.

Dolphin’s view: while scaling laws still hold (more params ⇒ more intelligence), model intelligence remains the key competitiveness. Value-for-money matters more once SOTA progress slows and capability gaps narrow.

Strategically, MiniMax appears to diverge from Anthropic’s text SOTA, OpenAI’s multi-modal SOTA, and DeepSeek’s extreme value in text. MiniMax may now aim for relative leadership in multi-modal at extreme cost-performance.

Under this choice, the company pursued multi-modal R&D and expanded self-operated compute even in a text-first phase. Dolphin questions whether this came too early, impairing intelligence gains, especially as value-for-money competition is intense — e.g., Kimi’s K3 sets a new performance/price benchmark for future open-source models ([details here](https://longbridge.com/zh-CN/dolphin/post/42853339)).

\<End here>

Related posts:

[‘Large Models’ 360% Losses — Why Is MiniMax Still in Favor?’](https://longbridge.cn/zh-CN/dolphin/post/38991861?channel=SH000001&invite-code=032064&app_id=longbridge&utm_source=longbridge_app_share&locale=zh-CN&share_track_id=d12c396f-771c-4dea-95d4-a409f3db69dc)

[‘Deep Dive: MiniMax vs. Zhipu — Large Models, Compute Intensity, and Financing Stamina’](https://longbridge.cn/zh-CN/topics/38121039?channel=SH000001&invite-code=032064&app_id=longbridge&utm_source=longbridge_app_share&locale=zh-CN&share_track_id=f43b7d6b-1801-4315-b9aa-dc5d8bc44318)

**Risk disclosure and statement:** [**Dolphin Research Disclaimer & General Disclosure**](https://support.longbridge.global/topics/misc/dolphin-disclaimer)

### Related Stocks

- [00100.HK](https://longbridge.com/en/quote/00100.HK.md)
- [GOOG.US](https://longbridge.com/en/quote/GOOG.US.md)
- [GOOGL.US](https://longbridge.com/en/quote/GOOGL.US.md)
- [GOOGN.US](https://longbridge.com/en/quote/GOOGN.US.md)
- [GGLL.US](https://longbridge.com/en/quote/GGLL.US.md)
- [GGLS.US](https://longbridge.com/en/quote/GGLS.US.md)
- [GOOY.US](https://longbridge.com/en/quote/GOOY.US.md)
- [GOOX.US](https://longbridge.com/en/quote/GOOX.US.md)
- [GOOP.US](https://longbridge.com/en/quote/GOOP.US.md)
- [GOU.US](https://longbridge.com/en/quote/GOU.US.md)

---
> **Disclaimer: This article is for reference only and does not constitute any investment advice.**