Xiaomi's Trillion-Parameter Model Tops Open-Source Charts, OpenAI Undercuts Claude at Half Price, DeepSeek Bets on Huawei Chips

The AI industry saw three noteworthy developments on the same day: Xiaomi open-sourced a trillion-parameter model that topped the open-source leaderboard, OpenAI launched a half-price model to undercut Anthropic, and DeepSeek's founder publicly bet on Huawei's domestic chips. Behind these three storylines lie overlapping battles over open-source competition, price war logic, and compute independence.

Xiaomi's Trillion-Parameter Model Tops Open-Source Charts — But Open-Source First ≠ Global First

Xiaomi officially released and open-sourced the Pro version of its MiMo V2.6 series, with a total parameter count of 1.02 trillion. It scored 46 points on AI benchmark leaderboards, making it the highest-scoring open-weight model — tied with xAI's Grok 4.7.

But a critical distinction needs to be made: topping the open-source category is not the same as being number one globally. The leading closed-source models currently score 53 points each, leaving a 7-point gap between them and Xiaomi's model. That gap may look modest on a benchmark chart, but in practice it often corresponds to a significant difference in real-world performance.

Also worth watching is how the cost figures are being misread. Xiaomi disclosed that the reinforcement learning phase cost approximately $2.62 million — but note, this covers only the reinforcement learning stage, not total training costs. Overseas headlines have already compressed this into "a trillion-parameter model trained for $3 million." This is nearly a replay of the DeepSeek^Ad V3 narrative: the technical report stated that formal training cost roughly $5.58 million (excluding prior research and ablation experiments), which eventually got boiled down to "built a frontier model for $6 million." Today's $2.62 million figure is an even narrower slice of total costs, leaving even more room for misinterpretation.

The real highlight is openness: weights, training code, and technical report are all released under the MIT license, with virtually no restrictions on commercial use. That's a genuine win for the open-source community and downstream developers.

---

Background: What is an "open-weight model"? An open-weight model is one where the vendor publicly releases the model's parameter files (weights), allowing anyone to download, deploy, and modify them — as opposed to closed-source models accessible only via API. The MIT license is among the least restrictive open-source licenses, permitting commercial use, modification, and redistribution with only a copyright notice requirement. This stands in sharp contrast to licenses that allow open access but restrict commercial use (such as Meta's early Llama license). The combination of fully open weights and the MIT license means companies can directly fine-tune Xiaomi's model for commercial products without paying licensing fees or facing usage restrictions — particularly attractive for small and mid-sized teams with sufficient compute but limited R&D capacity.

Background: What is the "reinforcement learning phase"? In the large model training pipeline, the reinforcement learning phase is typically the final step. Its purpose is to use reward signals to steer the model toward outputs that better align with human preferences — commonly via RLHF or similar methods. Its computational demands are far smaller than the earlier pre-training phase, which involves iterating over massive text corpora hundreds of billions of times and typically accounts for over 90% of total training costs. Disclosing only the $2.62 million reinforcement learning cost is technically accurate, but it's easy for readers unfamiliar with the training pipeline to mistake it for the full cost.

---

OpenAI Undercuts Claude at Half Price: A Familiar Price War Playbook

On the morning of September 22 (Pacific Time), Anthropic and OpenAI released new models roughly 90 minutes apart, igniting a direct price war.

Anthropic moved first with Opus 5.5, priced about 20% lower than its previous generation and roughly 60% cheaper than the top-tier closed-source models above it. Anthropic also claimed the new model could reduce costs for equivalent tasks by around 40%, though the company itself cautioned that the figure "varies by task" and encouraged developers to test it themselves.

About 90 minutes later, OpenAI launched two models — Sol and Luna — priced at exactly half of Opus 5.5's non-cached rate. The targeting was unmistakable.

A note of caution is warranted: there is currently no publicly available, unified benchmark that shows which company's models cost less to complete an identical task. All benchmarks released so far are vendor-published, lacking third-party standardization.

This battle has already played out in full in China. After DeepSeek led a wave of price cuts, Alibaba, Baidu, and Tencent all joined in, with reductions exceeding 90% at their peak — yet within 16 months, one vendor led a price increase of over 70%, and others followed. The domestic cycle has already written its ending: collective price cuts, followed by collective price hikes. Whether the overseas price war follows the same trajectory remains to be seen.

---

Background: What is "non-cached price"? "Non-cached price" is a critical detail in understanding API pricing. Major LLM APIs typically offer a "prompt caching" feature: when the same system prompt or context is called repeatedly, the cached portion is charged at a lower rate (typically 10–25% of the base price). The choice between "non-cached" and "cached" pricing when making comparisons can produce results that differ by several times. OpenAI's decision to benchmark Sol and Luna at half of Opus 5.5's non-cached price is a legitimate — but self-serving — framing choice. In real-world usage, developers frequently leverage caching, so the actual cost differential may be far smaller than 2x. This is precisely why the lack of a "unified-baseline cost comparison" is emphasized as the core obstacle to understanding the true substance of this price war.

---

DeepSeek Bets on Huawei Chips: A Wager That Must Pay Off

According to U.S. tech media reports, DeepSeek founder Liang Wenfeng told investors on September 21 that training the company's largest models on Huawei and other domestic chips is "one of the biggest bets we're making — and it has to work."

Why "has to"? Because the options have narrowed. NVIDIA's China data center shipments dropped to essentially zero in the February-to-April quarter of this year, compared to $4.6 billion in the same period the prior year. When external supply is cut off, turning to domestic chips is no longer a choice — it's a necessity.

Reports indicate Huawei expects to deliver a new batch of training chips as early as Q4 this year or Q1 next year. Liang Wenfeng believes Huawei's chip performance will catch up to NVIDIA within a few years. DeepSeek is currently training an ultra-large model with a parameter count far exceeding today's mainstream models.

But the transition comes at a cost. According to earlier reporting by the Financial Times, DeepSeek encountered stability and interconnect issues when attempting to train on Huawei's Ascend hardware and ultimately reverted to NVIDIA. And according to a previously leaked investor meeting transcript, Liang Wenfeng calculated that training a top-tier large model on Huawei Ascend would require roughly 4x the number of chips compared to NVIDIA.

This means domestic substitution is not just a question of whether it works — it's whether the efficiency and cost gap can be closed. The more decisively DeepSeek commits to this path, the clearer the signal that the external environment no longer allows hesitation.

---

Background: Huawei Ascend chips Huawei Ascend is Huawei's AI training and inference chip series, with the 910B and the latest 910C as its flagship models. Compared to NVIDIA's H100/H800, Ascend has closed much of the gap in raw single-chip compute, but the key differences lie in two areas: interconnect bandwidth and software ecosystem. NVIDIA's NVLink/NVSwitch technology enables thousands of GPUs to cooperate efficiently at scale, while Ascend's cluster interconnect solution lags behind in both stability and bandwidth during large-scale training — which is the technical root cause of the "stability and interconnect issues" mentioned in the article. On the software side, NVIDIA's CUDA ecosystem has over a decade of accumulated, highly optimized operator libraries; Huawei's corresponding CANN framework started later, and some deeply optimized training techniques require re-adaptation. This is also a major contributor to the estimated 4x chip count disadvantage.

---

The Common Logic Behind Three Storylines

Viewed together, these three developments reveal several defining themes in today's AI industry:

Open source is catching up, but the gap is real. Xiaomi's trillion-parameter model topping the open-source leaderboard is a milestone — but the 7-point gap with top closed-source models is a reminder that open-source first and globally first are still meaningfully different.

Price wars are competitive tactics, not endgames. OpenAI's half-price shot at Anthropic looks aggressive, but without a unified cost benchmark, it's hard to know who's actually cheaper to use. China has already demonstrated the full cycle of price cuts followed by price hikes — overseas markets will likely face the same commercial gravity.

Compute sovereignty has become a strategic imperative. DeepSeek's bet on Huawei is, at its core, a forced pivot driven by cut-off external supply. Whether domestic chips can match NVIDIA in efficiency will directly determine the ceiling on large model training in China.

These three storylines point to a shared conclusion: AI competition has expanded well beyond model capability benchmarks into a full-spectrum contest over cost structures, open-source ecosystems, and compute supply chains.