Update README.md

README.md

@@ -13,7 +13,7 @@ We're introducing Ling-3.0-flash, our next-generation native hybrid reasoning mo
 
 + **Native Hybrid-Linear Architecture:** Ling-3.0 adopts a native hybrid linear attention architecture from the very start of pretraining (5:1 alternating stacking of Kimi Delta Attention (KDA) and MLA), upgraded with KDA fine-grained diagonal gating and 1/64 sparse MoE. With 124B total parameters and 5.1B activated parameters, it achieves a synergistic leap in long-context efficiency and computational cost.
 + **Remarkable Efficiency & Performance:** Engineered for speed, compute efficiency, and production deployment, Ling-3.0-flash delivers class-defying performance against both larger SOTA competitors and previous-generation flagships. Activating only 5.1B parameters per token, it provides impressive reasoning, instruction following, and long-context capabilities to empower complex agentic workflows in production environments.
-+ **Comprehensive Agentic Evolution:** Tailored for real-world productivity workflows, the model incorporates 10,000+ interactive training environments to achieve end-to-end closed-loop execution across Coding, General, and Deep Research Agent tasks. It natively integrates the SGLang HiCache + Mooncake hierarchical caching architecture (featuring physical dual-pools and a cluster-shared L3 cache), eliminating redundant recomputation during long-horizon interactions and reducing Time to First Token (TTFT) by 60% to over 80% in long-input scenarios.s the SGLang HiCache + Mooncake hierarchical caching architecture (featuring physical dual-pools and a cluster-shared L3 cache), eliminating redundant recomputation during long-horizon interactions and reducing Time to First Token (TTFT) by 60% to over 80% in long-input scenarios.
++ **Comprehensive Agentic Evolution:** Tailored for real-world productivity workflows, the model incorporates 10,000+ interactive training environments to achieve end-to-end closed-loop execution across Coding, General, and Deep Research Agent tasks. It natively integrates the SGLang HiCache + Mooncake hierarchical caching architecture (featuring physical dual-pools and a cluster-shared L3 cache), eliminating redundant recomputation during long-horizon interactions and reducing Time to First Token (TTFT) by 60% to over 80% in long-input scenarios.
 
 <!-- 这是一张图片，ocr 内容为：SWE-BENCH MULTILINGUAL TERMINAL-BENCH 2.1 SWE-BENCHPRO TAU3-BANKING-AA LL088000LL 65 6 71.2 76.5 75.9 AP 28.0 60 0 72.4  73.3 56.6 56.2  56.3 71.0 53.9 70 22.9 55 57.0 50 47.9 48.3 60 56.7 14.6 45 11.3 71.3 39.3 39.0 70.1 50 8.9 40 42.7 40 34.1 35 30 30 WIDESEARCH MCP-ATLAS SKILLSBENCH BROWSECOMP 品质88品质品等学导新品 品民品品导品只只口 豆豆复品品导导导 品品&品品定品名品 6 4 79.5 米 0 75.2 73.6 74.4 70.2 69.0 82.0 66.7 6 65.5 62.2 53.5 75.8 61.2 73.2 71.7 74.0 55.2 53.6 52.6 49.4 G 20.3 19.5 71.9 31.3 MULTI-AGENT IFBENCH SYSBENCH MRCR-256K MULTI-IF 100 雪饼 110 110 87.7  89.3 79.2 06 82.9 84.6 86.2 84.6 84.8 75.7 82.3 93.6 91.4  93.9 90793.3949 72.6 84.3 90 品8888元 81.1 86.5 86.2 69.0 70 67.3 70 60 56.6 05 30 0 30 10 ULIO RING-2.6-1T(XHIGH) STEP-3.7-FLASH(HIGH) LING-3.0-FLASH MINIMAX-M2.7 WALBAN DEEPSEEK-V4-FLASH-PREVIEW(MAX) CLAUDE-SONNET-4.6(MAX) GPT-5.4-MINI(HIGH) NEMOTRON-3-SUPER-120B-A12B NOTE:THINKING MODE IS ENABLED BY DEFAULT. -->
 ![](https://intranetproxy.alipay.com/skylark/lark/0/2026/png/23157180/1785831264180-d6ca4404-acef-4424-84db-fbc5a4c6db5f.png)
@@ -62,50 +62,77 @@ We have conducted a comprehensive evaluation of Ling-3.0-flash across multiple a
 >     - <font style="color:rgb(16, 18, 20);">Draco: Scored based on official rubrics per question, with the final score calculated as the average across all questions using </font>Claude Opus 4.6<font style="color:rgb(16, 18, 20);"> as the scoring model.</font>
 >     - BrowseComp (Single-Agent): Evaluated using a resume strategy for context management: once the context reaches a 64K-token threshold, the trajectory is summarized, the original history is discarded, and execution is resumed from the summary.
 >     - BrowseComp (Multi-Agent):<font style="color:rgb(16, 18, 20);"> Evaluated on English and ZH Revised datasets using an internal multi-agent search harness based on SearchSwarm/Tongyi DeepResearch, configured with </font>`temperature=0.85, top_p=0.95, max_tokens=8K`<font style="color:rgb(16, 18, 20);">, and main/sub-agent context windows of 128K and 64K, respectively.</font>
->
+
+
+
+## Quantized Models
+We evaluate the quantized models using several datasets. The FP8 quantized model is applied via the blockwise quantization, and INT4 and FP4 models are applied via groupwise quantization with routed experts weights.
+
+|dataset	| BF16	| FP8	| INT4	| **FP4** |
+| -- | -- | -- | -- | -- |
+|GPQA-diamond	| 84.97	| 84.00	| 83.65	| 82.42 |
+|IFBench |	74.49	| 73.40	| 72.20	| 72.33 |
+|SciCode |	41.24	| 40.37	| 39.35	| 39.79 |
+|ArcPrize |	68.75	| 67.18	| 67.56	| 64.16 |
+
+
+
 ## Quickstart
 #### Install our SGLang
 ```bash
 pip install uv
+
 uv venv ~/my_ling_env
+
 source ~/my_ling_env/bin/activate
+
 git clone -b ling_v3_support https://github.com/inclusionAI/sglang_ling_v3.git
-cd sglang
+
+cd sglang_ling_v3
+
 pip install --upgrade pip
+
 pip install -e "python"
 ```
 
 #### Run Inference
-Here is the example to run Ling-3.0-flash with 4 GPUs, where the master node IP is `${MASTER_IP}` and server port is `${PORT}`:
+Here is an example to run Ling-3.0-flash with 2 Blackwell GPUs, where the master node IP is `${MASTER_IP}` and server port is `${PORT}`:
 
 **Server**
 
-As our model is trained with MTP, we recommend to turn on MTP (i.e., --speculative-algorithm NEXTN) in inference for lower latency.
+Since the model is trained with MTP, we recommend enabling MTP during inference (i.e., --speculative-algorithm NEXTN) for lower latency.
 
 ```bash
 export SGLANG_ALLOW_OVERWRITE_LONGER_CONTEXT_LEN=1
 python -m sglang.launch_server \
-    --model-path $MODEL_PATH \
-    --dist-init-addr $MASTER_IP:2345 \
-    --port $PORT \
+    --model-path "$MODEL_PATH" \
+    --trust-remote-code \
     --nnodes 1 \
-    --mem-fraction-static 0.8 \
+    --dist-init-addr "$MASTER_IP:2345" \
+    --port "$PORT" \
+    --tp-size 2 \
+    --ep-size 1 \
     --max-running-requests 64 \
-    --tp-size 4 \
+    --max-mamba-cache-size 320 \
     --chunked-prefill-size 8192 \
-    --allow-auto-output-truncate \
+    --context-length 262144 \
+    --random-seed 308534008 \
+    --attention-backend trtllm_mla \
+    --disable-flashinfer-autotune \
+    --mem-fraction-static 0.85 \
+    --fp8-gemm-backen cutlass \
     --tool-call-parser ling3 \
     --reasoning-parser ling3 \
-    --context-length 262144 \
-    --speculative-algorithm NEXTN \
-    --max-mamba-cache-size 320 \
+    --moe-runner-backend flashinfer_mxfp4 \
+    --flashinfer-mxfp4-moe-precision default \
     --enable-fp32-lm-head \
-    --disable-shared-experts-fusion
+    --disable-shared-experts-fusion \
+    --speculative-algorithm NEXTN
 ```
 
 **Client**
 
-It is recommended to use the sampling params (`<font style="color:rgb(16, 18, 20);">temperature=0.6, top_p=0.95, top_k=20</font>`) and turn on `enable_thinking`for better performance.
+We recommend using the sampling parameters `temperature=0.6`, `top_p=0.95`, and `top_k=20`, and enabling `enable_thinking` for better performance.
 
 ```bash
 curl -s http://${MASTER_IP}:${PORT}/v1/chat/completions \