update readme

.gitattributes

@@ -36,3 +36,4 @@ saved_model/**/* filter=lfs diff=lfs merge=lfs -text
 tokenizer.json filter=lfs diff=lfs merge=lfs -text
 figs/efficiency.jpg filter=lfs diff=lfs merge=lfs -text
 figs/title.png filter=lfs diff=lfs merge=lfs -text
+figs/performance.png filter=lfs diff=lfs merge=lfs -text

README.md

@@ -24,7 +24,7 @@ pipeline_tag: image-text-to-text
 
 ## Introduction
 
-We introduce **Intern-S2-Preview**, an efficient 35B scientific multimodal foundation model. Beyond conventional parameter and data scaling, Intern-S2-Preview explores **task scaling**: increasing the difficulty, diversity, and coverage of scientific tasks to further unlock model capabilities. 
+We introduce **Intern-S2-Preview**, an efficient 35B scientific multimodal foundation model continued pre-trained from Qwen3.5. Beyond conventional parameter and data scaling, Intern-S2-Preview explores **task scaling**: increasing the difficulty, diversity, and coverage of scientific tasks to further unlock model capabilities. 
 
 By extending professional scientific tasks into a full-chain training pipeline from pre-training to reinforcement learning, Intern-S2-Preview achieves performance comparable to the trillion-scale Intern-S1-Pro on multiple core professional scientific tasks, while using only 35B parameters. At the same time, it maintains strong general reasoning, multimodal understanding, coding, and agent capabilities.
 
@@ -45,12 +45,12 @@ By extending professional scientific tasks into a full-chain training pipeline f
 
 We evaluate the Intern-S2-Preview on various benchmarks, including general datasets and scientific datasets. We report the performance comparison with the recent VLMs and LLMs below. 
 
-![performance](./figs/performance.jpeg)
+![performance](./figs/performance.png)
 
 
 > **Note**: <u>Underline</u> means the best performance among open-sourced models, **Bold** indicates the best performance among all models.
 
-We use the [OpenCompass](https://github.com/open-compass/OpenCompass/) and [VLMEvalKit](https://github.com/open-compass/vlmevalkit) to evaluate all models.
+We use the [OpenCompass](https://github.com/open-compass/OpenCompass/) and [VLMEvalKit](https://github.com/open-compass/vlmevalkit) to evaluate all models. For text reasoning benchmarks, Intern-S2-Preview is evaluated with a maximum inference length of 128K tokens, while for multimodal benchmarks, it is evaluated with a maximum inference length of 64K tokens.
 
 
 ## Quick Start
@@ -288,3 +288,109 @@ print(json.dumps(response.model_dump(), indent=2, ensure_ascii=False))
 ```
 
 > Note: We do not recommend disabling thinking mode for agentic tasks.
+
+
+## Agent Integration
+
+Intern-S2-Preview can be plugged into agent frameworks in two ways: connecting to a **self-hosted deployment**, or calling the **official InternLM API**. Below we cover both, with examples for agent frameworks (OpenClaw, Hermes, etc.) and for Claude Code.
+
+### 1. Self-hosted Deployment (LMDeploy as an example)
+
+First, serve the model with LMDeploy following the [Model Deployment Guide](./deployment_guide.md). The example below assumes the server is running at `http://0.0.0.0:23333`.
+
+#### Connecting Agent Frameworks
+
+Most agent frameworks (OpenClaw, Hermes, etc.) accept an OpenAI-compatible endpoint. Point them at the LMDeploy server base url `http://0.0.0.0:23333/v1`. 
+
+You can check the connection with the following command:
+
+```bash
+curl http://0.0.0.0:23333/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer EMPTY" \
+  -d '{
+    "model": "internlm/Intern-S2-Preview",
+    "messages": [
+      {"role": "user", "content": "Hello"}
+    ],
+    "temperature": 0.8,
+    "top_p": 0.95
+  }'
+```
+
+Or you can configure your agent framework with the environment variables
+
+```bash
+export OPENAI_API_KEY=EMPTY
+export OPENAI_BASE_URL=http://0.0.0.0:23333/v1
+export OPENAI_MODEL=internlm/Intern-S2-Preview
+```
+
+Remember to launch LMDeploy with `--tool-call-parser interns2-preview` so tool calls are parsed correctly.
+
+#### Connecting Claude Code
+
+LMDeploy exposes an Anthropic-compatible `/v1/messages` endpoint that Claude Code can talk to directly. Add the following to `~/.claude/settings.json`:
+
+```json
+{
+  "env": {
+    "ANTHROPIC_BASE_URL": "http://127.0.0.1:23333",
+    "ANTHROPIC_AUTH_TOKEN": "dummy",
+    "ANTHROPIC_MODEL": "internlm/Intern-S2-Preview",
+    "ANTHROPIC_CUSTOM_MODEL_OPTION": "internlm/Intern-S2-Preview"
+  }
+}
+```
+
+For a full walkthrough (curl verification, model routing, troubleshooting), see [LMDeploy × Claude Code](https://lmdeploy.readthedocs.io/en/latest/intergration/claude_code.html).
+
+### 2. Official Intern API
+
+If you do not want to self-host, you can use the official Intern API. Register at [internlm.intern-ai.org.cn](https://internlm.intern-ai.org.cn/) and create an API token (`sk-xxxxxxxx`).
+
+#### Connecting Agent Frameworks
+
+The service is OpenAI-compatible, so any agent framework works. You can set the base url to `https://chat.intern-ai.org.cn/api/v1` and the model name to `intern-s2-preview` in the cli or config file.
+
+You can check the connection with the following command:
+
+```bash
+curl https://chat.intern-ai.org.cn/api/v1/chat/completions \
+  -H "Content-Type: application/json" \
+  -H "Authorization: Bearer sk-xxxxxxxx" \
+  -d '{
+    "model": "intern-s2-preview",
+    "messages": [
+      {"role": "user", "content": "Hello"}
+    ],
+    "temperature": 0.8,
+    "top_p": 0.95
+  }'
+```
+
+Refer to the [Intern API documentation](https://internlm.intern-ai.org.cn/api/document?lang=en) for the current endpoint, available model names, rate limits, and advanced parameters.
+
+#### Connecting Claude Code
+
+Claude Code can route to the official Intern API by pointing `ANTHROPIC_BASE_URL` at the Intern Anthropic-compatible gateway:
+
+```json
+{
+  "env": {
+    "ANTHROPIC_BASE_URL": "http://chat.staging.intern-ai.org.cn",
+    "ANTHROPIC_AUTH_TOKEN": "your-api-token",
+    "ANTHROPIC_MODEL": "intern-s2-preview",
+    "ANTHROPIC_SMALL_FAST_MODEL": "intern-s2-preview"
+  }
+}
+```
+
+Then start claude code with the following command:
+
+```bash
+claude --model intern-s2-preview
+```
+
+For step-by-step setup, see [Intern API × Claude Code Integration](https://internlm.intern-ai.org.cn/api/document?lang=en).
+

deployment_guide.md

@@ -0,0 +1,116 @@
+# Intern-S2-Preview Deployment Guide
+
+The Intern-S2-Preview release is a 35B-A3B model stored in bfloat16 weight format. This guide provides deployment examples for the following configurations:
+
+- MTP speculative decoding (Recommended)
+- Basic serving without MTP
+- Long-context inference with YaRN RoPE configuration
+
+> NOTE: The commands below are reference configurations. Inference frameworks are under active development, so use the latest framework documentation and your local validation results when tuning production deployments.
+
+## LMDeploy
+
+Use the latest LMDeploy (>=0.13.0) with Intern-S2-Preview support.
+
+- Serving With MTP (Recommended)
+
+```bash
+lmdeploy serve api_server \
+    internlm/Intern-S2-Preview \
+    --trust-remote-code \
+    --backend pytorch \
+    --tp 2 \
+    --reasoning-parser default \
+    --tool-call-parser interns2-preview \
+    --speculative-algorithm qwen3_5_mtp \
+    --speculative-num-draft-tokens 4 \
+    --max-batch-size 256
+```
+
+- Basic Serving Without MTP
+
+```bash
+lmdeploy serve api_server \
+    internlm/Intern-S2-Preview \
+    --trust-remote-code \
+    --backend pytorch \
+    --tp 2 \
+    --reasoning-parser default \
+    --tool-call-parser interns2-preview
+```
+
+- Long-Context Serving
+
+For long-context inference, configure both `--session-len` and YaRN RoPE parameters. The following example uses a 512k context length:
+
+```bash
+lmdeploy serve api_server \
+    internlm/Intern-S2-Preview \
+    --trust-remote-code \
+    --tp 2 \
+    --backend pytorch \
+    --reasoning-parser default \
+    --tool-call-parser interns2-preview \
+    --session-len 512000 \
+    --max-batch-size 64 \
+    --hf-overrides '{"text_config": {"rope_parameters": {"mrope_interleaved": true, "mrope_section": [11, 11, 10], "rope_type": "yarn", "rope_theta": 10000000, "partial_rotary_factor": 0.25, "factor": 4.0, "original_max_position_embeddings": 262144}}}'
+```
+
+## vLLM
+
+Use the latest vLLM Docker image or source build with Intern-S2-Preview support.
+
+- Serving With MTP (Recommended)
+
+```bash
+vllm serve internlm/Intern-S2-Preview \
+    --trust-remote-code \
+    --tensor-parallel-size 2 \
+    --reasoning-parser qwen3 \
+    --enable-auto-tool-choice \
+    --tool-call-parser qwen3_coder \
+    --speculative-config '{"method":"mtp","num_speculative_tokens":4}'
+```
+
+- Basic Serving Without MTP
+
+```bash
+vllm serve internlm/Intern-S2-Preview \
+    --trust-remote-code \
+    --tensor-parallel-size 2 \
+    --reasoning-parser qwen3 \
+    --enable-auto-tool-choice \
+    --tool-call-parser qwen3_coder
+```
+
+## SGLang
+
+Use the latest SGLang Docker image or source build with Intern-S2-Preview support.
+
+- Serving With MTP (Recommended)
+
+```bash
+SGLANG_ENABLE_SPEC_V2=1 \
+python3 -m sglang.launch_server \
+  --model-path internLM/Intern-S2-Preview \
+  --trust-remote-code \
+  --tp-size 2 \
+  --reasoning-parser qwen3 \
+  --tool-call-parser qwen3_coder \
+  --mamba-scheduler-strategy extra_buffer \
+  --speculative-algo 'NEXTN' \
+  --speculative-eagle-topk 1 \
+  --speculative-num-steps 3 \
+  --speculative-num-draft-tokens 4
+```
+
+- Basic Serving Without MTP
+
+```bash
+python3 -m sglang.launch_server \
+    --model-path internlm/Intern-S2-Preview \
+    --trust-remote-code \
+    --tp-size 2 \
+    --reasoning-parser qwen3 \
+    --tool-call-parser qwen3_coder
+```