Sage-A1.5: Shanghai AI Lab Open 35B MoE with Subagents and Verification Mode

Shanghai AI Laboratory has released Sage-A1.5, an open-weight 35B mixture-of-experts model aimed at long-horizon agentic work and scientific research. The weights are on Hugging Face as InternScience/Sage-A1.5 under Apache 2.0, with a technical report on arXiv (2610.28578). The model card describes it as built on Qwen3.6-35B-A3B and the lab's earlier Agents-A1 model.

The card makes two kinds of claims: concrete features you can try (parallel subagents and an opt-in verification mode), and benchmark numbers that put a roughly 3B-active model next to much larger systems. This post covers both, with the caveat that every score is self-reported.

TL;DR: Sage-A1.5 at a glance

What is new in Sage-A1.5?

The lab frames the model around three capabilities.

1. Agentic orchestration. Instead of one agent working a task serially, the model can delegate independent sub-questions to subagents with isolated contexts, then challenge assumptions and synthesize results.
2. Verification. An opt-in self-review that lives in the model's own chain of thought: it pauses mid-trajectory to review what it has established and what has failed, then revises direction in the same turn, "with no external critic."
3. Open scientific discovery. Exploratory research with inspectable artifacts, beyond answering predefined questions.

Compared with its predecessor Agents-A1, the card lists gains on engineering and agentic workloads: SWE-bench Pro up 17.6 points, SWE-bench Multilingual up 14.0, PinchBench up 13.1 and MLE-Lite up 21.3. The card says it was developed with an "RSI Data Flywheel" using real scientific discovery workflows and tens of thousands of synthesized environments, though we could not check the details in the report in the time available.

How do the benchmark numbers compare?

The card compares Sage-A1.5 against Agents-A1, Qwen3.6-35B-A3B, Ornith-1.5 (all roughly 35B), and larger Kimi-K3, GLM-5.3 and GPT-5.6-sol. A selection:

Read the pattern, not the headline. Sage-A1.5 leads on search, instruction following and the science-research benchmarks. The larger models still lead on SWE-bench Pro, HLE with tools, MLE-Lite and ChemCOTBench. The card's own summary says it "narrows the gap" with frontier systems rather than closing it.

Why these numbers need caution

• Self-reported. The card says it reports other models' results "from their original technical reports," and evaluates with the same protocol only where a benchmark was missing. That mixes harnesses and settings across rows.
• Harness matters. For SWE-bench the card names the mini-SWE-agent harness at temperature 1.0, with a 200K context and context compaction, run offline. Scores from other harnesses are not strictly comparable.
• Newer, less-known benchmarks. Several rows (FrontierScience, CMT, PinchBench, XBench-DS-2510) are not standard references; consistency across labs is hard to judge.
• Orchestration changes the result. BrowseComp is 81.7 without and 82.8 with agentic orchestration, so which mode a headline number uses matters.

For a similar self-reported open-model case, see our write-up of Ornith-1.5 on SWE-bench Verified, claimed vs verified, and the small-model landscape in our Haiku 5.5, GPT-6 Luna and DeepSeek V4.1 Flash comparison.

How subagent orchestration works

A decision fork signpost, illustrating how Sage-A1.5 splits a task across parallel subagents

Orchestration is off by default. You turn it on by exposing a spawn_subagent tool to the model. When a task splits into independent investigations, the model calls the tool with a list of tasks; your harness runs each in parallel and returns structured results.

Details from the card:

• Each task needs three fields: task, context and expected_output. Subagents cannot see the main conversation.
• Each subagent returns JSON with answer, status, evidence and unresolved_issues.
• A subagent that cannot resolve its task returns UNKNOWN with a status of unknown or insufficient_evidence, instead of guessing, so check status before using an answer.
• Harness-side limits: max_tasks_per_spawn defaults to 5 and subagent_max_turns to 20; oversized batches are rejected, not truncated.
• By default subagents inherit the main agent's tools; set subagent_tools to narrow them.

The model does not run subagents itself. You need a harness that executes the tool calls. If you build agent loops, our agent loop architecture guide covers the surrounding plumbing, and the Sakana agentic scientist release is a comparable open research-agent stack.

Verification Mode in practice

An audit of model behavior, illustrating Sage-A1.5 Verification Mode reviewing its own progress

Verification Mode is off unless you pass chat_template_kwargs with verification set to true in the request. The chat template then appends a verification protocol to the end of the system message. Constraints from the card:

• A non-empty system message is required; otherwise the server returns HTTP 400.
• Pass a JSON boolean, not the string "true".
• It works with or without tools.
• The self-review appears in reasoning_content, so read that as well as content.

This is a useful design for audits: you can log the reasoning trace. It is also a reminder that a self-review inside the same model is not an external check; it can reproduce the model's own blind spots.

How to run it

The card gives launch commands for both common servers. With vLLM, install from the main branch, then serve with a 262,144-token context, the qwen3 reasoning parser, auto tool choice and the qwen3_coder tool-call parser. A text-only flag skips the vision encoder to free KV-cache memory, and a multi-token prediction configuration is provided. SGLang commands mirror this, including an EAGLE speculative-decoding variant. To reach 512K context, both servers need a YaRN override with factor 2.0.

Recommended sampling: temperature 1.0, top_p 0.95, top_k 20, min_p 0.0, presence penalty 1.5.

For local use, official Q4_K_M, Q6_K and Q8_0 GGUF repos and an FP8 build appear under InternScience on Hugging Face, and the community has already posted MLX and other quantizations. Posts about the launch put the Q4_K_M footprint at roughly 20 GB (we did not verify this figure on the repository page), which would fit a high-memory laptop or a single consumer GPU with offloading. Third-party uploads, including an "abliterated" cybersecurity variant that someone has already posted, are not from the lab; treat unofficial builds with the usual care.

What people are asking

Is a 3B-active model really competitive with frontier systems? On the card's own table it is competitive on search and science-style evaluations and behind on the hardest engineering and tool-heavy ones. Mixture-of-experts models keep inference cost low because only a fraction of parameters run per token, but memory still has to hold all 35B parameters, which is why quantized builds matter.

Does orchestration make it faster? The card presents orchestration as a quality feature: parallel investigations with isolated contexts that cross-check each other. Parallel subagents can reduce wall-clock time on search tasks but multiply tool calls and token spend, and the card gives no cost or latency numbers. Measure both before enabling it.

Is it safe to hand it tools? Subagents inherit tools by default. If the main agent has write access, so do the subagents. Narrow subagent_tools, and treat web pages read by subagents as untrusted input that can carry prompt injection. A guard layer such as AgentBeam, from the explainx.ai team, is aimed at stopping agents before dangerous actions, which matters more once many subagents run unattended.

How does it differ from its Qwen base? The card says it builds on Qwen3.6-35B-A3B and Agents-A1. In its own table, Qwen3.6-35B-A3B scores far lower on the research-style rows (2.9 on FrontierScience-Research versus 48.3), which is the clearest sign of what the extra training changed, though the table cannot separate data, recipe and harness effects.

Who should try it

• Researchers and tinkerers who want an Apache-licensed agent model with built-in parallel decomposition.
• Teams comparing open models for science workflows, where its FrontierScience and CMT results are the draw.
• Not obviously teams needing top-tier software engineering: GPT-5.6-sol, GLM-5.3 and Kimi-K3 still lead SWE-bench Pro in the card's own table.

Also note the model is derived from Qwen. If your policy treats Chinese-origin models differently, see our coverage of the Qwen censorship audit and the broader China AI export catalogue.

What to check before you rely on it

1. Reproduce two or three benchmark rows with the lab's open evaluation code in the Agents-A1 repository.
2. Run your own tasks with orchestration on and off, and verification on and off, and compare cost and latency.
3. Check status fields on subagent results in production, not just answers.
4. Read the technical report for training data details and potential contamination notes.

Related reading

• Ornith-1.5 claimed vs verified
• Small models comparison, October 2026
• Sakana agentic scientist v3
• Agent loop architecture
• Kimi K3 on vLLM, claimed vs verified
• Qwen censorship audit
• Official: Sage-A1.5 on Hugging Face and the technical report

Details are accurate as of October 22, 2026 and come from the model card; scores are self-reported by the lab.