Stop Talking About DeepSeek R1 in the Abstract — Ten Founders Put Open-Source Models to the Test | BlueRun Ventures Perspectives
True knowledge comes from practice.
We know you're probably tired of hearing about DeepSeek R1. But when it comes to how founders can actually put open-source models — R1 included — to work, there are still very few early movers sharing their playbooks.
So BlueRun Ventures recently hosted a roundtable, inviting ten founders from its portfolio. Some had already integrated DeepSeek R1 and other open-source models into their products; others had deployed them for clients; still others had done deeper technical research.
The goal was to cut through the noise — from applications to underlying tech — and clarify what open-source models actually change, how they can channel greater intelligence into real business value, and what practical help that means for founders.
"After deploying DeepSeek R1, our business..."
Embodied Intelligence Company: We found that extending vision input on top of strong reasoning models like DeepSeek R1 significantly boosts their reasoning capabilities — even surpassing GPT-4o. So it's not just "slow thinking" that improves reasoning; multimodal fusion and cross-modal penetration also dramatically enhance a model's reasoning power.
If we can show that images, text descriptions, tactile feedback, and robot reinforcement trajectories can all be represented within a unified world model, then the path to AGI through VOA (virtual-to-reality and reality-to-virtual alignment) becomes much smoother. That's where our team is headed next. With modal fusion, language models would handle OOD (out-of-distribution) generalization tasks like cross-color grasping far more easily.
On modal fusion, three research directions stand out: 1) Explicitly learning causal relationships in representational space — current large models are still data-driven, with modal fusion relying more on statistical correlation than causation. 2) Causal intervention — by actively intervening, agents can better understand which environmental changes drive behavioral shifts, rather than depending on statistical similarity alone. 3) Integrating causal models and intervention capabilities into VOA — the ultimate goal is agents that freely switch between modalities and generalize in OOD environments.
SaaS Company: Our solutions serve two customer types: B-to-large-C (education, home renovation, automotive, real estate) and B-to-small-B (IP, tax and accounting). These industries rely heavily on phone and online sales. High-cost leads need meticulous management, but existing approaches are labor-intensive and struggle to process lead volume efficiently.
After integrating DeepSeek R1, we leveraged large models' ability to process massive unstructured data, converting recordings and chat logs into structured information. Sales scenarios demand extreme reasoning capability from models — not just data analysis, but actionable guidance on next steps, follow-up moves, and exact scripts for salespeople. With R1, our lead management efficiency improved tenfold, and customer response rates jumped two to three times.
R1 is performing particularly well in education. The AI-generated scripts are more logically structured and hit customer pain points directly — sales managers have been genuinely stunned. For application startups like us, success depends heavily on customer awareness and buy-in. That's one of R1's biggest values.
Social Product Company: Our experience with DeepSeek R1 so far suggests that the best monetization model right now combines AI with real humans to deliver services previously done by humans alone. Users fundamentally prefer interacting with real people; nobody wants to pay purely for AI itself.
That said, we can already see the impact powerful open-source software — R1 included — will have on gaming. Game creation has always had high barriers, but those will drop fast. Soon people will create fun games from just a few sentences describing their ideas.
We discovered a non-consensus insight while building AI social products. Most AI social apps today follow this pattern: user sends one message, AI replies with a long block of text. This is wrong — it's not how real social interaction works. The Stanford Town experiment showed us that once a user builds a relationship with an entity, the AI can continue simulating that entity's inner life and maintaining the relationship even when the user isn't actively engaging.
This is a fundamental shift AI enables. In traditional social software, the server basically forwards messages between two human users. In AI-driven social software, the server should host numerous "souls" that deliver better experiences.
Interactive Gaming Company: We've integrated two models: DeepSeek R1 and Claude 3.5. Comparing them, R1 leans toward logical information delivery, while Claude 3.5 emphasizes visual and sensory description.
Describing a "hitting someone" scene, for example: R1 uses logical language to fully narrate cause, process, and outcome, obsessively tracking whether the hit landed. Claude 3.5 cares less about the result, instead strictly following our requirements — say, distributing across the five senses with heavier visual weighting.
Initially we needed five agents working together. With R1, we consolidated two functions into one agent without performance loss. In fact, precision of language expression, degree of human-like style, and rigor in rule adherence all exceeded the five-agent setup. The biggest win was solving latency — we compressed it to millisecond levels. We also saw significant improvements in emotional perception, narrative approach, and hosting technique.
After DeepSeek, where do founders go?
Infrastructure Company A: Top technical talent can reproduce DeepSeek's approach from the paper's summary. So those with technical capability and capital should boldly invest in foundation model research. Without sufficient funding, focus on application development.
Whether you studied computer science matters less now — even humanities majors can start companies. The trend is shifting from heavy assets toward intellectual, algorithm-driven light-asset models. Tools like DeepSeek will only get better and cheaper; capital importance declines while creativity and execution speed become decisive.
Most importantly, R1 brilliantly exploited the "fear of being left behind by AI" that foreign mega-corporations and tech giants cultivated over the past two years. Open-sourcing at exactly the right moment, combined with genuine usefulness, created a star effect — the "momentum" (势) is what matters most. This will reinvigorate the startup ecosystem.
All great breakthroughs come from true openness. So when will we see truly open large models? As AI grows more important, governments worldwide will deepen their involvement, and that moment will gradually arrive in the foundation model space.
On the application side, many large companies with scale will build AI applications regardless of cost. In choosing models, they won't necessarily lock themselves into their own — they'll select flexibly based on actual needs. This sets a good example: when model costs drop, application development shouldn't cling to one model but should draw on the best of many.
Infrastructure Company B: We've observed three main customer demands for AI inference services: First, once model capability reaches a baseline, customers care about concurrency and elasticity — evening peaks for entertainment apps, Double Eleven, Black Friday. Second, cost and price-performance, since cost determines which scenarios are viable; customers want to unlock more use cases, whether paid apps or free ad-supported models. Third, latency — both end-to-end and time-to-first-token directly impact user experience. Digital human livestreaming, intelligent phone agents, and embodied intelligence are especially sensitive here.
We have some observations on DeepSeek's explosion. In my view, the optimal, fully open-source base model was bound to emerge eventually, driven by cost and price competition. Everyone in the open-source ecosystem benefits.
I'd also suggest paying more attention to distilled versions — many scenarios don't need the full V3 or R1. I've always believed each generation of large model revolution follows this pattern: first a closed-source breakthrough, then a large-parameter open-source model that rapidly raises industry awareness, then a wave of companies doing secondary distillation on top, finally driving costs down for mass adoption — and that mass adoption will inevitably be cloud-edge-device fusion.
We're in a price war environment. Startups should prioritize market voice and visibility; cost and price rationality can come second. Everyone is currently offering API services at a loss — nearly 100 companies claimed DeepSeek integration within two weeks. Painful short-term, but the cost-optimization curve and our overseas performance both show healthy market potential. At sufficient scale, the free-AI era arrives.
If model costs drop another tenfold in the coming year, I believe most applications won't need to build their own models. Currently many app companies spend 40-60% on token costs; in a year that could fall to 20%.
Data Services Company: Our core business is AI deployment in industries through a data-centric approach. We specialize in intelligently weaving enterprise local data into knowledge, embedding it as industry-specific agents into core workflows.
Three factors previously blocked large model adoption in B2B: inability to eliminate hallucinations; inability to guarantee data security and control; poor reasoning capability and explainability.
DeepSeek R1's biggest push for us was accelerating industrial deployment by rapidly educating users. Before this, state-owned enterprises and other major clients were still watching from the sidelines, uncertain about ROI and high-value use cases. R1's viral breakthrough made that education cost zero.
A Technical Lens on DeepSeek R1
Large Model Company: From a technical perspective, displaying pure RL reasoning processes is DeepSeek's most striking breakthrough — it also offers a pathway for AI to surpass human reasoning. Additionally, DeepSeek lacks the overly mechanical feel of previous models; this is a powerful and currently hard-to-replicate quality. Long-term, how to present this requires continuous refinement.
Currently RL uses only tens of thousands of data points. Why so few? Because RL requires extremely complex and challenging problems to trigger reasoning processes, and such data is very scarce. So it must rely on human effort to pose these complex questions. With current data volumes, compute demands are massive because RL efficiency is low — even lower than SFT. But in the future, as data scales and we discover appropriate scaling laws, clearly more compute will be needed to support this.
This wave of technological development is a bullish signal, especially for tech giants. Previously, large companies lacked confidence they could surpass OpenAI. Now that's changed — we'll see more investment and innovation ahead.
On foundation models, I focus on two points. First, whether RL's Scaling Law can be achieved. We know model intelligence improves with more compute, but whether data scaling achieves the same remains unclear. Clarifying this would give us another path to continuously improve model intelligence. Second, whether multimodality can lead us to AGI. If this path works, progress accelerates. In NLP, vision is already the easier task.
AI Product Company: DeepSeek is open-source, but openness isn't our primary selection criterion — fit matters. We need to remember that model technology isn't the goal; if it significantly reduces costs, it's worth using.
We evaluate three things: capability, speed, stability. 2025 is the year of Agent development, and the critical thing about Agents is that whatever model you use, it needs to perform well at key steps. The more freedom you give an Agent, the more likely it is to go off the rails — that's when you need a stable model guarding critical steps.
From a technical standpoint, companies trying to do both model and application together face extreme difficulty. Application builders don't want to be constrained by models, yet their self-built models often underperform.
For application development, we see several major directions ahead:
First, as Agents advance, reasoning models will be essential.
Second, real-time APIs and interaction. The essence is efficient, end-to-end real-time APIs with reduced response times, with voice as input. This enables entirely new interaction experiences, especially in multimodal scenarios where models need to better handle social norms like conversational interruptions and context switching.
Finally, Operator and Container-related developments. Through optimized context engineering, developers can improve task completion efficiency without depending on specific models. For example, the Cursor tool optimizes programming context so developers can flexibly switch models rather than being locked into one.
Embodied Intelligence Company: DeepSeek R1 can't be called comprehensively superior to American products — it shows advantage on one dimension, a distinctive large model. Several points deserve demystification:
First, R1's filtering process involved substantial human and engineering optimization effort. This is actually one of the key factors behind its strong performance, but it's been overlooked. Second, DeepSeek R1 uses product-thinking to improve user interaction feel — it summarizes your previous questions in its chain of thought, then places what it doesn't know into the output. So while it appears model capability has improved, the underlying reasons may be less "glamorous" than people imagine.
That said, DeepSeek shows us three likely breakthroughs in the next one to three years. First, traditional large model prediction and RL will deeply integrate, possibly not through current fusion methods but by internalizing long-horizon reasoning results into the model's fast thinking. Second, DeepSeek's MoE technology will see wide adoption, driving large model application costs to minimum levels faster than people expect. Third, future applications won't rely on single models — different model types will combine into Agent systems delivered as products.
Should current large model companies keep developing cutting-edge models while also building applications? From our foundation model development experience, if application barriers are high enough, relying solely on base models can't satisfy product requirements. Take GPT: even with base models near product-ready standards, bridging that "last centimeter" gap required fine-tuning at enormous cost — sometimes one-tenth or one-fifth of developing a model from scratch.
This means product design must weigh from the start whether such massive foundation models are truly needed, and whether their cost and capability tradeoff makes sense. Rather than blindly pursuing model capability limits, then discovering in application that costs are prohibitive and having to rebuild from scratch.
Multimodal DeepSeek Arrives! Yaodong Yang's Team and HKUST Release Align-DS-V, Powering Lingchu DS-VLA Alignment and Expansion | BlueRun Ventures Portfolio News
Luchen Tech Completes New Funding Round | BlueRun Ventures Portfolio
Galaxy Universal Unveils GraspVLA, World's First End-to-End Embodied Grasping Foundation Model, Defining New Paradigm for All-Synthetic Large-Scale Pre-training | BlueRun Ventures Portfolio News
