🧔♂️ A friendly human may check it before it goes live. More news here
Chinese quant fund Ubiquant unveils code LLMs, rivals OpenAI
Beijing-based Ubiquant has announced a family of open-source, code-focused large language models (LLMs) that it says rival leading US systems in benchmark tests.
Ubiquant, one of China’s largest quantitative trading funds, reported that its IQuest-Coder-V1 models performed on par with or better than OpenAI’s GPT-5.1, and Anthropic’s Claude Sonnet 4.5, in several programming benchmarks, despite having fewer parameters.
The firm cited scores such as 76.2% on SWE-bench Verified, 49.9% on BigCodeBench, and 81.1% on LiveCodeBench for its 40-billion-parameter model, based on internal testing.
Ubiquant manages over US$10 billion in assets and was among the first Chinese quant funds to apply AI at scale for investment.
Other Chinese quant trading firms, including High-Flyer Quant, Goku Technologies, Mingshi, Wizard Quant, and Mengxi Investment, have also increased their AI development efforts in the wake of DeepSeek’s rise in the LLM field.
🔗 Source: South China Morning Post
🧠 Food for thought
Implications, context, and why it matters.
US parity claims need public leaderboard checks
- Ubiquant says IQuest-Coder-V1 hits 76.2% on SWE-bench Verified (a benchmark that evaluates whether a model can resolve real-world software issues in code repositories under controlled conditions) 1, which trails Claude Sonnet 4.5 77.2% 2. These are in-house runs so public checks are needed.
- They clarified and re-ran SWE-bench Verified 1. No public leaderboard entry exists so 76.2% may not match prompts, context limits, or tool settings.
- The data includes commits, pull requests, and review comments (developer feedback) rather than static snapshots 2. That could lift results if it generalizes beyond memorization.
- Without third-party checks on public leaderboards, developers will struggle to judge real performance versus marketing.
Dev tool vendors can explore cost gains if models prove viable
- AI infrastructure providers could adopt the 40 billion-parameter IQuest-Coder-V1 1 to cut API spend versus GPT-5.1 or Claude Sonnet 4.5 if the weights (the model parameters) are open and licensed for business use.
- Quantized builds include 4-bit 3 and 2-bit to 8-bit formats 4 for runs on local or edge hardware. Quantized means reducing numerical precision to shrink size and speed up inference, with minor accuracy trade-offs.
- Teams can serve it with vLLM (an open-source inference engine that serves models via an OpenAI-style API) for OpenAI-compatible endpoints 1 to plug into tools that call US APIs.
- Adoption still hinges on license terms for commercial use, measured costs or latency on common hardware, plus proof that results hold up outside controlled tests.
