Castform推出一个新平台，使开发者能够通过强化学习后训练开源模型，在检索任务上以100倍更低的成本击败GPT-5.6-Sol。[1]该平台与Neon数据库集成，将企业内部数据转化为训练任务，使小规模开源模型能够匹配或超越闭源前沿模型的性能。[1]
根据相关信息，一次典型的GPT-5.6-Sol多轮搜索请求需耗时超过10秒，成本约为0.03美元。[1]相比之下，通过该平台优化的开源模型大幅降低了推理成本和延迟。[1]Castform平台使用强化学习后训练技术，使开发者无需处理机器学习和GPU的底层细节。[1]为了应对突发工作负载，该平台与Neon的Lakebase Search功能集成，利用动态计算扩展。[1]Neon的分支功能为每个推断创建隔离的数据库状态。[1]
Castform has introduced a platform enabling developers to optimize open-source models through reinforcement learning post-training, achieving performance comparable to or exceeding GPT-5.6 Sol on retrieval tasks while reducing costs by a factor of 100 [1]. The platform abstracts away machine learning and GPU implementation details, allowing developers to focus on training without specialized technical expertise [1].
The cost and performance differential is significant in practical applications [1]. A typical multi-turn retrieval request using GPT-5.6 Sol requires more than 10 seconds and costs approximately $0.03 per query [1]. Castform's approach enables smaller open-source models to match these capabilities while dramatically lowering both inference expenses and latency [1].
The platform integrates with Neon's Lakebase Search and leverages dynamic compute scaling to handle traffic spikes [1]. Neon's branching feature creates isolated database states for each inference operation, supporting the infrastructure requirements of the reinforcement learning workflow [1].