A95B 开源，硅基流动即日上线

Key Highlights

Alibaba open-sourced Qwen3.8-2.4T-A95B, and SiliconFlow (SiliconFlow) promptly announced Day-0 support. Simply put, the moment the model was released, domestic developers could call it directly on SiliconFlow's platform without waiting through a long adaptation cycle. The variant carries 2.4T total parameters with 95B active parameters, and is positioned for autonomous coding, deep research, and end-to-end agent execution. The combination of an open weight release and same-day hosted availability is the headline, because it collapses the usual gap between "model is public" and "model is usable in production." For a community that often waits weeks for third-party hosting to catch up, Day-0 support removes a meaningful friction point and lets experimentation begin immediately. The model also signals Alibaba's intent to compete at the frontier of agent-capable open weights rather than only in the chat space, a strategic move that pairs research ambition with a clear deployment story.

What Happened and How It Worked

The name A95B already hints at its scale: a total of 2.4 trillion parameters, of which only 95 billion are activated per inference. This "large but sparse" MoE structure lets the model preserve strong capability while keeping per-call cost under control. SiliconFlow's Day-0 launch means open-sourcing and commercial deployment happen almost in lockstep; after developers obtain the weights, they do not have to wrestle with their own deployment environment and can simply use the latest model through an API. The practical upshot is that experimentation begins on day one rather than weeks later, which matters enormously in a field moving this fast. Hosted access also means teams without large GPU clusters can still evaluate the model at production quality, narrowing the gap between well-funded labs and smaller builders. The synchronized release thus serves both open-source purists and pragmatic API consumers at once, and it sets a new expectation that major open models should be runnable the same day they are announced.

Technical Details

On pricing, A95B's input cost is $2.00 per million tokens, output is $6.00 per million tokens, and cached input is $0.25 per million tokens. The cached input price is pushed extremely low, which is very friendly to scenarios that repeatedly send long context, such as large codebases or research documents, and it significantly dilutes cost. The 95B active-parameter scale is also sufficient to support complex multi-step reasoning and tool calling. In effect, the caching tier rewards workloads that reuse context, which is precisely the pattern agentic systems exhibit when they loop over the same project files. A low cache price also encourages developers to keep rich system prompts and reference material attached to every call without fear, improving consistency across agent steps. The economics are clearly tuned for the realities of agent workloads rather than one-shot question answering, where context is small and discarded after a single reply.

Comparison with Competitors

Against the open-source camp of DeepSeek, Llama, and others, A95B takes "end-to-end agent execution" as its core selling point, emphasizing not just chatting but running an entire task chain on the user's behalf. Simply put, Alibaba wants to trade a larger active scale for steadier agent performance, while SiliconFlow competes for the inference entry point of open models with low prices and same-day availability. The strategic picture is a two-front push: one frontier in raw model capability, another in who controls the easiest on-ramp to run these models. Where some rivals optimize for benchmark headlines, A95B pairs capability with a deployment story, which is what enterprises actually evaluate when choosing a model to ship. The open weights also invite community fine-tunes that could extend the model's reach further, turning a single release into a platform others build upon.

Industry Impact and Use Cases

For teams building autonomous coding assistants, research-style retrieval, and long-chain automation, A95B is a new foundation worth trying. Day-0 availability lowers the trial barrier, and the cheap cache pricing looks after heavy-context users. We can expect that competition among open large models on the "agent execution" track will grow intense because of this heavyweight release. The model's arrival also signals that open weights are no longer just about matching closed models in benchmarks, but about shipping a complete, immediately usable agent platform. As more agents move into production, models that pair strong execution with friendly economics will set the pace, and A95B is clearly built to be one of those. The broader effect is a tightening race among Chinese AI labs to own the open-agent narrative, with same-day hosting becoming a competitive feature rather than a courtesy. The release also demonstrates how the open-model ecosystem in China has matured to the point where a major lab and an inference platform can coordinate a same-day launch, a capability that previously took weeks to appear. For developers, that coordination removes the usual waiting tax and lets new capabilities reach production almost immediately after an announcement, rather than after a slow cycle of community porting and benchmarking.