InfoQ 中文 · 10/8/2026, 10:49:05
Reflection and Mistral Launch New Models to Counter Chinese Open-Source Dominance
US startup Reflection and French firm Mistral launched Beam and Large 4 models respectively, explicitly targeting Chinese open-source leaders like DeepSeek and GLM. Data shows Chinese models account for 41% of Hugging Face downloads, while the share of tokens processed by open-weight models on Vercel surged from under 10% to 56%, signaling their rapid transition into mainstream infrastructure.
SOURCE COVERAGEFull Original Report
Contents6 sections
Reflection and Mistral recently released Beam and Mistral Large 4 in close succession, benchmarking against China’s leading open models. Hugging Face data shows that Chinese models account for 41% of platform downloads, establishing them as the primary reference point for global open-source developers. The productionization of open models is accelerating: their share of tokens processed via Vercel AI Gateway has surged from less than 10% at the end of last year to 56% by August this year.
Reflection’s technical team originates from DeepMind’s core RLHF and reward modeling groups. Chinese open models now lead globally in download volume and engineering deployment speed. Open models are transitioning from research tools to mainstream production infrastructure.
This article is suitable for AI infrastructure engineers, open-source model application architects, and LLM product leads.
The Western open-model camp is beginning to pick up speed.
On October 5 local time, Reflection—a US startup dubbed the "American DeepSeek"—released its first open-weight model, Beam. One day later, French company Mistral launched Mistral Large 4, its largest model to date. Both companies chose highly consistent benchmarks for their releases: GLM, Kimi, DeepSeek, and Qwen.
Hugging Face data indicates that over the past year, Chinese models have accounted for approximately 41% of total model downloads on the platform, surpassing US models in both monthly and cumulative download volumes. Models such as DeepSeek, Alibaba’s Qwen, Moonshot AI’s Kimi, Zhipu AI’s GLM, MiniMax, and Xiaomi’s MiMo have successively pushed the capability limits of open models, making Chinese open models the target for global open-source developers to catch up with.
On the other hand, open models are no longer just "niche alternatives" used by researchers and geeks. Production traffic data from Vercel AI Gateway shows that by August this year, open-weight models were processing 56% of the platform's tokens, compared to less than 10% in December of last year. Consequently, a global race for open models is inevitable.
US and European Open Models Debut Simultaneously
"American DeepSeek" Submits Its First Attempt: A Gap Remains
Reflection, founded in 2024 and heavily anticipated, was established by two former Google DeepMind executives: CEO Misha Laskin, who previously worked on Gemini’s reward modeling, and CTO Ioannis Antonoglou, who contributed to AlphaGo and AlphaZero before leading Gemini’s Reinforcement Learning from Human Feedback (RLHF) efforts.
However, Reflection did not initially aim to be the "American DeepSeek." The company originally focused on autonomous coding, aiming to use reinforcement learning to build "super-intelligent systems" capable of autonomously completing complex tasks. They also assumed that sufficiently powerful open models would continue to emerge in the West, allowing Reflection to build directly upon them.
What truly changed the company’s trajectory was DeepSeek V3. According to Reflection’s founders, after Chinese open models advanced rapidly while no equivalent alternatives appeared in the West, the company decided to train frontier open models itself, gradually clarifying its goal as "bringing the open-model frontier back to the West."
Two years later, Beam represents Reflection’s first serious submission.
Beam is a Mixture-of-Experts (MoE) model with 501 billion total parameters, activating 23 billion parameters per token. It is primarily designed for coding, reasoning, and agent workloads. The model was pre-trained on 23.8 trillion tokens. Reflection utilized 10,500 Nvidia GB300 GPUs for four weeks of high-compute reinforcement learning, generating over 100 million rollouts. The model is currently undergoing final red-teaming; Reflection plans to release the model weights, technical report, and development tools in October.
Based on test results published by Reflection, Beam’s biggest highlight is not absolute capability but inference efficiency. The company claims that on certain advanced reasoning tasks, it can achieve performance close to GLM-5.2 using only one-quarter to one-third of the inference compute.
However, when compared horizontally with China’s latest generation of models, Beam still lags significantly behind true parity.
For example, in the agentic coding benchmark DeepSWE v1.1, Beam scored 44.4, nearly matching GLM-5.2’s 44.0, but falling short of GLM-5.3’s 61.0, Kimi K3’s 68.0, and DeepSeek V4.1 Flash’s 74.2. On Terminal Bench v2.1, Beam scored 80.1, whereas GLM-5.3, Kimi K3, and DeepSeek V4.1 Flash achieved 88.2, 88.3, and 90.6 respectively. On Humanity's Last Exam (HLE), Beam scored 36.2, lower than GLM-5.3’s 42.3 and Kimi K3’s 46.9.
Therefore, Beam currently brings US open models back into proximity with China’s top-tier models, but it primarily catches up to the previous generation (GLM-5.2). There remains roughly a one-generation gap between Beam and China’s latest round of open models.
Ioannis stated that the company's goal for the next few years is to eventually merge the "open frontier" and the "closed frontier" into a single frontier. In 2027, Reflection plans to release larger-scale models while continuing to expand its reinforcement learning compute resources. Similar to models like GLM, it intends to continuously launch multiple RL-enhanced versions built on top of its base models.
Misha, however, believes that this roadmap is underpinned by a broader judgment: Reinforcement Learning (RL) has evolved from a domain-specific technique used in AlphaGo and AlphaZero into a foundational method capable of driving continuous progress in general-purpose models, coding models, and Agent models.
In his view, current models are either already a form of AGI or clearly situated on a direct path toward AGI. The core challenge is no longer just about making models stronger, but rather how to engineer and productize these capabilities, and disseminate them through a multipolar, more open ecosystem.
Mistral Reclaims Its Position Among Top Open Models
Compared to Reflection, which has just released its first base model, European model developer Mistral faces significantly greater pressure.
Two to three years ago, it was arguably the most representative company in the open model space. However, with the rapid iteration of models like DeepSeek, Qwen, Kimi, and GLM, Mistral’s presence on leading capability benchmarks has steadily declined. In an interview this September, CEO Arthur Mensch was directly asked whether Mistral had fallen behind the US and China, given that its models had dropped out of the top 20 on Artificial Analysis rankings.
Mensch responded at the time that a new generation of large models was imminent, noting that the company had already planned the next two product generations for the coming year. He added that newly raised capital would be heavily invested in expanding training compute, potentially increasing available training compute by approximately 20 times.
Two weeks later, Mistral Large 4 arrived.
Mistral Large 4 features 52 billion active parameters and 1.05 trillion total parameters, equipped with a 1.6-billion-parameter visual encoder. It natively supports multimodal inputs and over 160 languages, with a context window reaching 1 million tokens. Notably, it was trained from scratch using approximately 3,800 Nvidia Grace Blackwell GPUs in Mistral’s European data centers. The model currently offers an API preview, with full weights scheduled for release by the end of October.
According to Mistral’s official Coding Agent Index, Large 4 surpasses DeepSeek V4 Pro and Qwen 3.8 Max but trails Kimi K3. In blind tests conducted in collaboration with third-party firm Surge AI, professional reviewers rated Large 4’s code quality at 3.74—lower than Claude Opus 5’s 4.22, but higher than Kimi K3’s 3.59, GLM-5.3’s 3.60, and GLM-5.2’s 3.40.
Mistral is clearly targeting "Enterprise Agents" rather than just developers. Across 657 business workflows spanning applications like Gmail, Google Sheets, Slack, and Salesforce in AutomationBench, Mistral claims Large 4 outperforms Kimi K3, MiMo-V2.6-Pro, and DeepSeek V4 Pro, though it falls short of GLM-5.3. For scientific coding, cybersecurity, and certain financial and legal tasks, Mistral also asserts that it has reached the top tier among open models.
This time, Mistral specifically highlighted its cybersecurity capabilities. The company states that Large 4 has entered the global top five in the Artificial Analysis Cyber Index, achieving an 82% success rate on a task involving reproducing and fixing real-world vulnerabilities—the highest score in that test. On Cybench, it completed 93% of challenges. Mistral deliberately emphasized that some closed-source models, due to safety refusal mechanisms, can barely answer certain vulnerability reproduction tasks, whereas open-weight models allow enterprises to define their own security policies and deploy them in private clouds or local environments.
Of course, much of this data currently comes from Mistral’s internal testing, and Large 4 remains in public preview; the actual weights have not yet been released.
Validating Business Models Initially Piloted by Chinese Vendors
Beyond model capabilities, both Mistral and Reflection are pursuing commercial paths initially validated by domestic Chinese vendors.
Reflection represents a very "American-style" business story: It aims to build a comprehensive AI infrastructure provider, starting from base models and extending upward to inference, Agents, software, and enterprise deployment, while diving downward into compute management and cluster scheduling. Long-term, it even plans to own its data centers.
He argues that every layer—from models and applications to cloud computing and bare metal—will eventually become commoditized. Companies that sustain high profit margins are often those that vertically integrate the entire technology stack. Misha even proposed a distinctly "Silicon Valley-esque" formula:
Revenue Potential = Intelligence Density × Compute Scale × Enterprise Trust.
Misha noted that open models themselves may not directly generate revenue, but they create massive demand for inference. Just as Kubernetes itself is open source yet drove the growth of the cloud market, open models will spur enterprises to purchase compute, inference services, and complete AI infrastructure. Therefore, what Reflection truly sells is the capability to build a full suite of intelligent infrastructure around Beam.
It even compares itself to early-stage players who solved similar problems: By addressing challenges such as training large models, managing GPU clusters, and optimizing inference costs, Reflection positions itself as providing essential infrastructure solutions.
Arthur argues that small, efficient open models capable of customization and deployment better align with enterprise needs than general-purpose models solely pursuing maximum scale. Mistral is simultaneously building its own models, an enterprise customization platform, and European compute infrastructure, aiming to become a supplier where training, inference, and deployment can all occur within the European legal and infrastructural framework.
In his view, European enterprises cannot base their core AI capabilities entirely on whether foreign vendors are willing to continue providing services, nor can they hand over their technological foundation to models from other countries. Therefore, Mistral not only trains its own models but also builds data centers in Europe, seeking to control the entire chain from compute and models to enterprise deployment.
Attitude Toward Domestic Open Models
One company is based in Europe, the other in the US, but recent public statements by their founders indicate they are forming increasingly similar judgments about open models: The AI market has shifted from "who provides the best API" to "who owns the models and controls the infrastructure."
On the question of "why we must build open models," Reflection explains that the AI market is moving from "renting intelligence" to "owning intelligence."
Misha compares closed-source APIs to "renting a house." Over the past three years, as the AI commercial market was just forming, enterprises calling upon models from companies like OpenAI, Anthropic, and Google were essentially renting intelligence. This model was simple and efficient enough for the early stages of the market. However, as AI truly enters core business operations, things begin to change.
"Startups that were small a few years ago have grown into large enterprises, and these large enterprises are now genuinely using AI themselves; naturally, a market for ownership has emerged," Misha stated. If a company wants to truly own intelligence, it must be able to deploy models on its own infrastructure, control data, modify models, and customize them—which means the models must be sufficiently open.
Ioannis approaches this from a technical perspective, noting that the gap between open models and the cutting edge of closed-source models is narrowing, and that "there is no technical reason why open models must always lag behind." In his view, given sufficient talent and compute resources, open models are fully capable of reaching parity with closed-source models. Previously, enterprises avoided open models primarily because "they weren't good enough"; now, a batch of Chinese open models can directly replace closed-source models for many tasks, making the premise for commercial migration viable.
Mistral CEO Arthur Mensch’s judgment is very close to Reflection’s on this point. He believes that as open models improve, most enterprises no longer need to pay additional API premiums for closed-source models. He recently stated that approximately 99% of enterprise scenarios Mistral encounters can be handled by open models.
So why did Chinese models take the lead over the past year?
Both companies share a similar assessment: The leadership of Chinese open models is not merely due to one company accidentally releasing a great model, but rather the result of combined commercial incentives, industrial environment, and training resources.
Misha noted that the US actually once had the strongest open models—Llama 3 being a clear milestone. But at that time, the US AI commercial market was still dominated by APIs, so Meta or other companies lacked strong enough commercial motivation to continuously release frontier capabilities openly.
The situation in China is different. DeepSeek V3 was a significant turning point; it was the first time Chinese open "intelligence" made a large-scale global impact, followed by a more concentrated push for open models in China. Misha believes that a substantial part of this momentum comes from the fact that building one's own open-weight ecosystem inherently holds industrial and strategic value.
"In reality, these Chinese labs hadn’t made much money until recently. It’s only now that their model capabilities are strong enough, and the market has matured enough to pay for open intelligence, that they are starting to generate real revenue," Misha said.
Arthur offers a more engineering-focused explanation: While the compute resources of top Chinese labs are lower than those of the largest US AI companies, they are still higher than Mistral’s. Meanwhile, they utilize internet data as extensively as possible, and their iteration methods have been industrialized. These conditions allow Chinese labs to iterate rapidly.
Are Open Models More Dangerous?
Although both companies actively advocate for open models, neither describes itself as an extreme open-source proponent who believes "all models should be unconditionally open."
Reflection’s core security argument is that openness itself enhances system safety. Misha borrows "Linus’s Law" from the open-source software world: Given enough eyeballs, all bugs are shallow. In his view, closed AI labs with only hundreds of security researchers cannot cover all long-tail issues in extremely complex models. If open models allow hundreds of thousands of researchers to participate in inspection, attack, and repair, the entire ecosystem may actually become safer.
The incident where an agent in OpenAI’s internal cybersecurity evaluation escaped the test environment and entered Hugging Face systems is frequently cited by Reflection. Misha believes this demonstrates that a single security philosophy is unreliable: Closed models might fail to execute certain defensive tasks due to safety guardrails, whereas open models can provide an alternative set of tools.
Ioannis further proposes the idea of "using AI to defend against AI." As models become stronger, future anomalous agents can be detected and constrained by another AI system. The more models and providers there are, the higher the system redundancy; if the strongest capabilities are concentrated in just one or two companies, it creates a "single point of failure."
However, Reflection also acknowledges that openness is not unlimited. Ioannis explicitly states that once models reach a certain capability level, deployment methods require greater caution; Misha also notes that after models exceed a certain risk threshold, industry and government should jointly decide under what conditions they can be released and whether access controls are necessary.
Arthur recently publicly rebutted claims regarding superintelligence runaway risks. Rather than advocating for active deceleration, he prefers strengthening agent monitoring, security engineering, and practical governance.
This stance partially overlaps with Anthropic’s position. Anthropic is more cautious about the high-risk diffusion of open weights but similarly advocates that open and closed-source models should face consistent capability-based safety thresholds.
Reference links:
https://reflection.ai/blog/introducing-beam?utm_source=chatgpt.com
https://mistral.ai/news/mistral-large-4/?utm_source=chatgpt.com