Model Overview
Description:
The RadixArk Qwen3.8-2.4T-A95B-NVFP4 model is the quantized version of Qwen/Qwen3.8-2.4T-A95B, a causal language model with 2.4T total parameters and 95B activated parameters. The quantization was produced at RadixArk using NVIDIA Model Optimizer, following an experts-only NVFP4 recipe.
Run on SGLang: launch command and per-platform recipes in the Qwen3.8 cookbook.
Third-Party Community Consideration
This model is not owned or developed by RadixArk. It is a quantized derivative of Qwen's model; see the upstream Qwen3.8-2.4T-A95B model card for the source model's capabilities, training information, limitations, and license.
License/Terms of Use:
Deployment Geography:
Global
Use Case:
Developers looking to take off-the-shelf, pre-quantized models for deployment in AI Agent systems, chatbots, RAG systems, and other AI-powered applications.
Release Date:
Hugging Face 08/07/2026 via https://huggingface.co/RadixArk/Qwen3.8-2.4T-A95B-NVFP4
Model Architecture:
Architecture Type: Transformers (Mixture-of-Experts) Network Architecture: Qwen3.8-2.4T-A95B Number of Model Parameters: 2.4T total, 95B activated
Input:
Input Type(s): Text Input Format(s): String Other Properties Related to Input: Native context length up to 262,144 tokens.
Output:
Output Type(s): Text Output Format: String
Software Integration:
Supported Runtime Engine(s):
- SGLang
Supported Hardware Microarchitecture Compatibility:
- NVIDIA Blackwell (this checkpoint was produced and validated on B300 SXM6)
Preferred Operating System(s):
- Linux
Model Version(s):
Quantized with NVIDIA Model Optimizer, commit 87c9f8cf83021957d1a1a575c90c9a4eaaf7ef0c.
Training, Testing, and Evaluation Datasets:
Calibration Data:
The conversion used deterministic synthetic token calibration. See CONVERSION.md for the complete conversion and validation contract.
Training Dataset:
RadixArk did not train or fine-tune this checkpoint. Training information is inherited from the upstream Qwen3.8-2.4T-A95B model card.
Evaluation Dataset:
The model was evaluated on GSM8K, AIME 2026, SWE-bench Verified, and Terminal-Bench 2.1.
Post Training Quantization
This model was obtained by quantizing Qwen/Qwen3.8-2.4T-A95B with an experts-only NVFP4 recipe. Only linear operators in the routed-expert MLP blocks are quantized; other components retain the source checkpoint precision. See CONVERSION.md for tensor-level details.
Usage
The recommended SGLang configuration for a single 8×B300 node is:
SGLANG_ENABLE_MOE_DEFERRED_FINALIZE=1 \
SGLANG_FLASHINFER_MNNVL_CUTEDSL_AR_FUSION=1 \
sglang serve \
--trust-remote-code \
--model-path RadixArk/Qwen3.8-2.4T-A95B-NVFP4 \
--tp-size 8 \
--context-length 200000 \
--preferred-sampling-params '{"top_k": 20}' \
--attention-backend trtllm_mha \
--page-size 64 \
--linear-attn-prefill-backend flashinfer \
--linear-attn-decode-backend flashinfer \
--moe-runner-backend flashinfer_trtllm \
--mamba-radix-cache-strategy extra_buffer \
--mamba-ssm-dtype bfloat16 \
--speculative-algorithm DSPARK \
--speculative-draft-model-path RadixArk/Qwen3.8-2.4T-A95B-DSpark \
--mem-fraction-static 0.80 \
--max-running-requests 128 \
--chunked-prefill-size 8192 \
--max-prefill-tokens 8192 \
--cuda-graph-backend-prefill breakable \
--cuda-graph-max-bs-prefill 8192 \
--cuda-graph-backend-decode full \
--cuda-graph-max-bs-decode 128 \
--reasoning-parser qwen3 \
--tool-call-parser qwen3_coder \
--host 0.0.0.0 \
--port 30000
For other deployment topologies and hardware-specific configurations, see the SGLang Qwen3.8 cookbook.
This checkpoint is text-only and requires thinking mode. Multimodal inputs and enable_thinking=false are not supported.
Evaluation
The benchmark results below were produced with this NVFP4 checkpoint on an 8×B300 TP8 SGLang deployment.
| Benchmark | Evaluation protocol | Score |
|---|---|---|
| GSM8K | Full 1,319-example split, sgl-eval | 97.73% |
| AIME 2026 | 30 problems × 16 repeats, sgl-eval | 98.75% pass@1; 100% pass@16 |
| SWE-bench Verified | 500 instances, agentic evaluation with unbounded steps | 70.6% pass@1 (353/500) |
| Terminal-Bench 2.1 | 89 tasks, Claude Code agent | 87.64% pass@1 (78/89) |
GSM8K used zero-shot evaluation with recommended sampling (temperature=1.0, top_p=0.95). AIME 2026 used 16 independent samples per problem, for 480 generations in total. SWE-bench Verified used the official test split and the SWE-bench harness to score submitted patches. For Terminal-Bench 2.1, Claude Code multimodal requests were disabled.
Model Limitations:
The base model was trained on data that contains toxic language and societal biases originally crawled from the internet. Therefore, the model may amplify those biases and return toxic responses especially when prompted with toxic prompts. The model may generate answers that may be inaccurate, omit key information, or include irrelevant or redundant text producing socially unacceptable or undesirable text, even if the prompt itself does not include anything explicitly offensive.
Ethical Considerations
RadixArk believes Trustworthy AI is a shared responsibility and we follow established policies and practices that enable development for a wide array of AI applications. When downloaded or used in accordance with our terms of service, developers should work with their internal model team to ensure this model meets requirements for the relevant industry and use case and addresses unforeseen product misuse.