Shanghai AI Lab Open-Sources Sage-A1.5: 35B MoE Model Gets Day-0 vLLM Support, Deploy Agents Locally

Shanghai AI Lab has made another move in the open-source model space. Its newly released Sage-A1.5 model debuts with a 35B total parameter, 3B active parameter Mixture-of-Experts (MoE) architecture, positioned around "agentic intelligence," and emphasizes that it can be deployed and run directly in local environments. More importantly, the inference framework vLLM announced support on the day of release, so developers can start the service with a single command.

Model Architecture: 35B Total Parameters, 3B Active MoE Design

Sage-A1.5 uses a MoE (Mixture-of-Experts) architecture with 35B total parameters, but only about 3B parameters are activated per inference. This design strikes a balance between model capacity and inference cost: the total parameter scale determines the model's knowledge capacity and expressive power, while fewer active parameters means the cost of a single forward pass is significantly reduced.

For local deployment scenarios, this characteristic is especially important. If a traditional dense model reached the 35B level, inference would require loading all weights and performing full computation, placing high demands on VRAM and compute. The MoE architecture, through sparse activation, makes it possible for developers to run larger-scale models on consumer-grade or single-GPU workstations. Sage-A1.5's 3B activation level is roughly equivalent to inference overhead close to that of a small dense model, yet it can draw on the knowledge carried by 35B parameters.

The core idea of the MoE (Mixture-of-Experts) architecture is to split the model's feed-forward network into multiple parallel "expert" subnetworks and introduce a router/gating mechanism that dynamically selects a small number of experts for each input token to participate in computation. Therefore, "total parameters" is the sum of all weights, while "active parameters" is the number of parameters a single token actually passes through. This sparse activation allows the model to expand parameter scale (improving knowledge capacity) while keeping the FLOPs of a single forward pass far below the level of the total parameters. The cost is that VRAM still needs to hold all expert weights, and whether routing load is balanced and how expert parallelism is partitioned can significantly affect actual throughput. Models such as DeepSeek^Ad and Mixtral adopt similar approaches; Sage-A1.5's 35B/3B is a configuration with a very low activation ratio.

Day-0 vLLM Support: Start the Service with One Command

On the day the model was released, vLLM announced support for Sage-A1.5. Developers can start the inference service directly with the following command:

vllm serve InternScience/Sage-A1.5

vLLM is a high-throughput inference engine widely used in the open-source community, known for optimizations such as PagedAttention and continuous batching. Day-0 support means that when the model weights are released, the inference-side toolchain is already ready, and developers do not need to wait for community adaptation or write inference code themselves. For teams hoping to quickly validate model capabilities and build agent applications, this greatly lowers the barrier to entry.

From an ecosystem perspective, synchronized adaptation between model releases and inference frameworks is becoming an implicit standard in open-source model competition. Whether a model can be supported by mainstream inference frameworks on its release day directly affects how quickly it is adopted. Sage-A1.5 has gotten off to a good start on this front.

vLLM is an open-source large model inference and serving engine initiated by a team at UC Berkeley and now maintained by the community. Its key innovation, PagedAttention, borrows the idea of virtual memory paging from operating systems, dividing the KV Cache into fixed-size blocks and allocating them on demand, greatly reducing VRAM fragmentation and thereby significantly improving concurrent throughput; continuous batching dynamically schedules at the sequence level, allowing new requests to be inserted into an executing batch at any time. These optimizations have made vLLM a mainstream choice for production deployment. So-called Day-0 support means that on the day the model weights are made public, vLLM has already merged adaptation code for the corresponding model structure, tokenizer, and chat template, so users can launch an OpenAI-compatible service with vllm serve without modifying the source code.

Agentic Intelligence: Positioned for Agentic Scenarios

From the official description, Sage-A1.5 is positioned as "agentic intelligence"—intelligence capabilities oriented toward agent scenarios. This positioning aligns with the current trend in large model applications: more and more development needs are no longer single-turn Q&A, but rather having models perform multi-step reasoning, call tools, and complete complex tasks within an environment.

Agent scenarios have different requirements for models than chat models. They require strong instruction-following ability, multi-step planning ability, and stability in tool-calling formats. Whether Sage-A1.5 has made targeted optimizations in these dimensions is currently limited in public information, but judging from its naming and promotional framing, agent capabilities are the core selling point. After deploying locally with vLLM, developers can verify it themselves in scenarios such as tool calling and task planning.

Agentic intelligence generally refers to a model's ability to autonomously complete long-horizon tasks in a "perceive—plan—act—observe" loop, rather than merely generating single-turn responses. Its typical technical forms include: stably outputting tool-call requests in structured formats (such as JSON), maintaining task state across multi-turn interactions, reflecting and correcting based on tool return results, and performing multi-step planning under paradigms such as ReAct and function calling. Common benchmarks for evaluating such capabilities include tool-calling tasks (such as BFCL) and interactive environment tasks (such as WebArena and SWE-bench). These scenarios place different demands on models than chat: format stability, long-context state retention, and error recovery are often more critical than sheer breadth of knowledge.

The Significance and Limitations of Local Deployment

"you can serve locally today" is the focus emphasized in this release. Local deployment means data does not leave the internal network, inference costs are controllable, and latency is predictable—attractive to teams that are privacy-sensitive or require deep customization. Combined with the sparse activation characteristics of MoE, Sage-A1.5 is more feasible on local hardware than a dense model with the same parameter count.

However, 35B total parameters still means the model weight files are not small, and sufficient VRAM or memory is needed when loading. The actual throughput performance of the MoE architecture on different hardware also depends on expert parallelism strategy and batching configuration. vLLM's support solves the question of "whether it can run"; "how fast and how efficiently it runs" still requires developers to test in practice based on their own hardware.

When running a 35B-level MoE model locally, VRAM requirements need to be viewed in two parts: weight storage is roughly proportional to the total parameter count (in FP16, 35B roughly requires tens of GB, and quantization can significantly compress this), while compute overhead is related only to active parameters. Therefore, a common approach is to use post-training quantization such as AWQ or GPTQ to compress weights to around 4bit, then run them on a single-GPU or dual-GPU workstation. Expert parallelism refers to distributing different experts across multiple GPUs, with cross-GPU communication during routing, which introduces additional latency; when batch size is small, MoE compute utilization may also be low. These factors together determine the gap between "can run" and "runs efficiently," and actual throughput needs to be measured in practice based on hardware and batching configuration.

A New Variable in Open-Source Model Competition

Shanghai AI Lab continues to invest in open-source models, and Sage-A1.5 is one of its moves aimed at agent scenarios. The combination of 35B MoE plus 3B activation targets the middle ground of "capable enough, cost controllable." Day-0 vLLM support shows the maturity of the release process—coordination between model teams and the inference ecosystem is becoming tighter.

For developers, this means one more locally deployable agent foundation option; for the open-source model landscape, the rapid adaptation of MoE architectures and inference frameworks is turning "usable upon release" into a new default expectation. Whether Sage-A1.5 can deliver convincing performance on agent tasks will determine whether it can stand out from the many open-source models.

One-line Summary

Shanghai AI Lab releases Sage-A1.5, a 35B total parameter, 3B active MoE model focused on agent scenarios and local deployment, with Day-0 vLLM support.

Paragraph Summary

Shanghai AI Lab has launched the open-source model Sage-A1.5, using a Mixture-of-Experts (MoE) architecture with 35B total parameters and only about 3B activated, balancing model knowledge capacity with single-inference overhead and making local deployment possible. The model is positioned as "agentic intelligence," targeting agentic scenarios such as multi-step reasoning and tool calling. On release day, the mainstream inference engine vLLM announced Day-0 support, so developers only need a single vllm serve command to start a local service compatible with the OpenAI API, without waiting for community adaptation. The article argues that synchronized adaptation between model releases and inference frameworks is becoming an implicit standard in open-source model competition; MoE sparse activation lowers the hardware barrier, but 35B weight size and expert parallelism efficiency still determine actual throughput, requiring developers to verify in practice based on their own hardware.