Shanghai AI Lab releases Sage-A1.5, a 35B MoE agentic model with 256K native context

Shanghai AI Laboratory has released Sage-A1.5, a 35B-parameter Mixture-of-Experts agentic model with a native 256K-token context window, published under the InternScience organization on Hugging Face. The lab claims it reaches 95.1 on GAIA, above every larger model in its comparison set, and improves 17.6 points on SWE-bench Pro over its predecessor.

Specs

• Parameters: 35B total, Mixture-of-Experts. Active parameter count is not stated in the available material.
• Context window: 262,144 tokens (256K) natively. The model card says it can be extended to 524,288 (512K) by serving with YaRN at factor 2.0.
• Formats and serving: Hugging Face Transformers weights, compatible with vLLM and SGLang. An EAGLE-based multi-token prediction (MTP) speculative decoding configuration is documented for SGLang.
• Pricing: not yet disclosed. The model card links a free trial.
• Training cutoff, license, and release date: not stated in the available material.

What's new

Sage-A1.5 builds on A1 and targets long-horizon scientific problem solving. Shanghai AI Lab highlights three capabilities:

• Agentic orchestration: a lead agent delegates tasks to parallel subagents, each with an isolated context, to reduce interference between concurrent investigations.
• Verification Mode: an opt-in capability inside the model's chain of thought. The model pauses mid-trajectory to review what has worked and what has failed, then revises its approach in the same turn, with no external critic.
• Open scientific discovery: exploratory research workflows that produce inspectable research artifacts.

It also supports native function calling. The documented serving commands use the qwen3 reasoning parser and qwen3_coder tool-call parser. The evaluation notes treat OmniDocBench as a multimodal benchmark that Sage-A1.5 is scored on, which implies image input support, but the available material does not detail modalities.

Benchmarks (self-reported)

All figures below come from Shanghai AI Lab's own evaluation, not independent testing. The lab compares Sage-A1.5 against Agents-A1 and Ornith-1.5 (both ~35B-class) and the larger Kimi-K3, GLM-5.3, and GPT-5.6-sol.

Benchmark	Sage-A1.5
GAIA	95.1
XBench-DS-2510	92.0
OmniDocBench	91.5
PinchBench	90.0
IFBench	81.6
BrowseComp	81.7 (82.8 with agentic orchestration)
FrontierScience-Olympiad	81.0
AA-LCR	81.0
CMT	80.0
SWE-bench Multilingual	72.7
ChemCOTBench	66.5
MLE-Lite	65.2
SWE-bench Pro	61.2
HLE w/ tools	55.1
FrontierScience-Research	48.3

Versus Agents-A1, the lab reports gains of +17.6 on SWE-bench Pro, +14.0 on SWE-bench Multilingual, +13.1 on PinchBench, and +21.3 on MLE-Lite.

The claims are not uniform. On GAIA, Agents-A1 scores 96.0, above Sage-A1.5. The larger models still lead on several tasks:

• BrowseComp: Kimi-K3 at 91.2 and GPT-5.6-sol at 90.4.
• SWE-bench Pro: 64.6 for GLM-5.3 and GPT-5.6-sol.
• MLE-Lite: GPT-5.6-sol at 77.3.
• HLE w/ tools: GPT-5.6-sol at 64.5.
• CritPt: GPT-5.6-sol at 32.3, against Sage-A1.5's 20.0.
• ChemCOTBench: Kimi-K3 at 82.5.

SWE-bench results use the mini-SWE-agent harness at temperature 1.0, top-p 0.95, a 200K-token context, and automatic context compaction, in a fully offline environment. A technical report is linked from the model card.

What this means

Sage-A1.5 is a data point for efficiency in agentic models. A 35B MoE that lands within a few points of much larger systems on SWE-bench Pro (61.2 vs. 64.6) and GAIA is cheap enough to self-host. That matters for research groups that cannot send sensitive data to closed APIs.

The gaps are also informative. The model trails on knowledge-heavy and hardest-tier tasks such as CritPt, ChemCOTBench, and HLE w/ tools. This is consistent with the pattern that parameter count still buys breadth of knowledge, while agentic scaffolding closes the gap on search and orchestration-heavy tasks.

Treat the "overall SOTA" labels with caution until independent replication. The benchmark set is large and partly niche, the evaluation harnesses are the lab's own, and the comparison models' scores are not independently verified. The practical test will be how Verification Mode and subagent orchestration perform in third-party agent frameworks, which the open weights and vLLM/SGLang support now make possible.