Alibaba's Qwen team released Qwen3.8-2.4T-A95B as open weights on Hugging Face. It is a sparse mixture-of-experts model with 2.4 trillion total parameters and about 95 billion activated per step, built on a 92-layer hybrid attention backbone with 512 experts. It takes text input, supports function calling and a 1M token context window, and is served through Alibaba Cloud Model Studio at $1.65 per 1M input tokens and $4.951 per 1M output tokens in the Beijing region. Qwen3.8-Max is the hosted version with extra features such as vision input.
A workable 2400B-parameter MoE language model from Alibaba. Pulls ahead on graduate-level reasoning (GPQA) (93/100), so reach for it when that's the dimension that matters. Newly released, so production-readiness is still being shaken out.
Generated from this model’s benchmarks and ranking signals. Editor reviews refine it over time.
Access model weights, configuration files, and documentation.
Benchmark data sourced from: Hugging Face, Arena.ai, Artificial Analysis.
Average benchmark score against active parameters for every text model we track. Models higher up deliver more quality for their size.
The top devices for this model at 4-bit, ranked by fit and speed.
| Device | Grade | Speed | VRAM |
|---|---|---|---|
| GIGABYTE NVIDIA Vera Rubin NVL72 GIGAPOD AI DLC RackGigabyte | SS | 1381.8 tok/s | 2152.5 GB |
| ACEMAGIC M1A Pro (i9-13900HK + ARC A770)ACEMAGIC | FF | 0.5 tok/s | 2152.5 GB |
| Acer Veriton GN100 AI MiniAcer | FF | 0.3 tok/s | 2152.5 GB |
| AI TOP ATOM ATAGB10-9003Gigabyte | FF | 0.3 tok/s | 2152.5 GB |

Explore the Provider
Aggregate stats, leaderboard, release timeline, and benchmark coverage across every Alibaba model we track.

Explore the Family
The full Qwen family leaderboard with sizes, benchmark scores, and a release timeline.
| AMD Instinct MI300XAMD |
| FF |
| 5.2 tok/s |
| 2152.5 GB |
See how different quantization levels affect VRAM requirements and quality for this model.
| Format | VRAM Required | Quality | |
|---|---|---|---|
| Q2_K | 1648.5 GB | Low | |
| Q4_K_MRecommended | 2152.5 GB | Good | |
| Q5_K_M | 2392.5 GB | Very Good | |
| Q6_K | 2680.5 GB | Excellent | |
| Q8_0 | 3280.5 GB | Near Perfect | |
| FP16 | 5560.5 GB | Full |
Alibaba's Qwen3.8-2.4T-A95B marks a milestone in open-weight intelligence, delivering a flagship-grade model to self-hosted infrastructure. Operating as a sparse mixture-of-experts (MoE) model with 2.4 trillion total parameters and roughly 95 billion parameters activated per token, it brings the core text engine behind Alibaba's hosted Qwen3.8-Max directly to open deployment. While the managed cloud version adds native multimodal vision capabilities, the open weights provide developers with unconstrained control over an uncensored, full-scale reasoning engine.
The model is built for practitioners demanding frontier performance across complex coding, mathematical reasoning, agentic planning, and long-document processing. Equipped with an extensible 1,000,000-token context window, Qwen3.8-2.4T-A95B directly targets enterprise tasks that previously required proprietary APIs like Claude 3.5 Sonnet or GPT-4o. It represents the upper bound of open weights released to date, trading lightweight consumer accessibility for unmatched zero-shot capability and agent reliability.
Positioned at the apex of the open ecosystem, this release redefines what local AI deployments can achieve. It is not designed for everyday consumer workstations, but for private data centers, high-performance computing clusters, and enterprise teams seeking to avoid vendor lock-in without sacrificing cutting-edge reasoning depth.
Qwen3.8-2.4T-A95B utilizes a 92-layer hybrid attention architecture designed to balance long-context throughput with deep associative memory. Rather than relying entirely on standard multi-head or grouped-query attention, the backbone integrates 23 macro-blocks. Each block consists of three Gated DeltaNet linear attention layers paired with MoE feed-forward networks, followed by one Gated Attention layer paired with an MoE feed-forward network.
The attention configuration uses 64 query heads and 4 key-value heads with a head dimension of 256 for standard gated attention, alongside 128 linear attention heads for value and 16 for query-key in the DeltaNet layers. This hybrid structure prevents key-value (KV) cache bloat from crippling throughput across extended sequences, making the 1M token context window computationally viable.
The sparse mixture-of-experts design is the core driver of Qwen3.8-2.4T-A95B MoE efficiency:
This architectural layout yields significant operational advantages during inference. Because only 95B parameters activate per token pass, compute requirements mirror those of a mid-sized dense model. However, the memory footprint remains tied to the full 2.4T parameter registry. All 512 experts must reside in accessible memory (or high-speed tiers) to prevent severe routing latency penalties during expert dispatch.
Qwen3.8-2.4T-A95B is focused purely on text-based reasoning, autonomous execution, and instruction following. It does not process native image inputs like its hosted counterpart, but its text-based execution capabilities set a benchmark for open systems.
The model features native support for structured JSON generation, tool calling, and recursive environment feedback. On the OSWorld benchmark, it scores 86.1, demonstrating an ability to interact with operating system environments, execute multi-step shell commands, parse returned errors, and correct course without human intervention. The engine includes explicit thinking controls, allowing developers to set reasoning_effort and maintain reasoning state across tool calls using preserve_thinking.
Deploying Qwen3.8-2.4T-A95B for coding allows engineering teams to ingest entire code repositories within a single context pass. Ranking 4th on CodeArena, the model excels at cross-file refactoring, tracing complex dependency bugs, and writing comprehensive unit test suites. Its 1M token context capacity allows full framework documentation and active source code to sit simultaneously in the prompt, eliminating retrieval-augmented generation (RAG) fragmentation.
On academic evaluations, the model records a GPQA Diamond score of 92.6 and a PaperBench score of 93.0, reflecting strong performance on the Qwen3.8-2.4T-A95B reasoning benchmark suite. It processes raw scientific literature, validates mathematical proofs, and synthesizes cross-disciplinary research with minimal hallucination.
Attempting to run Qwen3.8-2.4T-A95B locally demands enterprise-grade hardware planning. A common question among engineers is how to run 2400B model on consumer GPU setups. The direct answer: you cannot run this model on a single consumer card like an NVIDIA RTX 4090 (24GB) or even a top-tier Apple M4 Max (128GB unified memory). The total weight tensor alone exceeds the memory ceiling of any standalone workstation.
Because all 2.4 trillion parameters must be addressed across the 512 experts, VRAM allocation is determined by weight size plus KV cache headroom:
For production clusters, the best quantization for Qwen3.8-2.4T-A95B is Q4_K_M or FP8. FP8 preserves full benchmark parity across reasoning and coding tasks. Q4_K_M cuts memory requirements to roughly 1.3 TB, making it possible to host on a single cluster of two 8x NVIDIA H100 (80GB) or H200 (141GB) nodes.
Selecting the best GPU for Qwen3.8-2.4T-A95B depends on scale:
When deployed across an optimized multi-GPU array using vLLM, SGLang, or TokenSpeed, Qwen3.8-2.4T-A95B tokens per second typically fall within:
For localized testing and automated cluster provisioning, Ollama and vLLM are the fastest deployment options:
1# Pull and initialize via vLLM across an 8-GPU node using 4-bit tensor parallelism2vllm serve Qwen/Qwen3.8-2.4T-A95B \3 --tensor-parallel-size 8 \4 --pipeline-parallel-size 2 \5 --quantization awq \6 --max-model-len 262144
For containerized setups running Ollama with custom Modelfiles, ensure distributed sharding is enabled across your physical cluster before pulling the GGUF equivalents.
Understanding where this model fits among mega-scale open weights requires looking at the primary alternatives in the local AI model 2400B parameters 2026 landscape.
DeepSeek-V3 features 671 billion total parameters with 37 billion active per token, utilizing a multi-head latent attention (MLA) architecture. DeepSeek is significantly easier to run: its weights fit comfortably in roughly 350 GB to 400 GB of VRAM at 4-bit precision, requiring half the nodes of Qwen3.8.
However, Qwen3.8-2.4T-A95B holds nearly four times the total parameter capacity and activates 95B parameters per step. This grants Qwen superior performance on long-horizon OS-level agent benchmarks (86.1 OSWorld vs 72.4 on DeepSeek-V3) and significantly deeper domain specialization across long documents. Choose DeepSeek for cost-effective enterprise hosting; choose Qwen3.8 if absolute frontier capability outweighs hardware cost.
Meta's Llama 3.1 405B is a dense model. Running Llama 405B requires computing all 405 billion parameters for every single token, resulting in heavier FLOP requirements and slower decoding speeds per token on equivalent hardware.
By contrast, Qwen3.8-2.4T-A95B computes only 95B parameters per step, delivering roughly three times higher token throughput once weights are loaded in memory. While Llama 3.1 405B requires less total VRAM to store weights (~230 GB at 4-bit vs ~1.3 TB for Qwen at 4-bit), Qwen provides a 1M token context window compared to Llama's 128k, alongside higher reasoning ceilings on mathematical and scientific benchmarks. Choose Llama 405B if your infrastructure has limited VRAM; choose Qwen3.8 when you have multi-terabyte cluster capacity and need high inference throughput alongside long-context execution.
Energy cost running this model locally vs flagship API pricing.
| Source | Cost per 1M tokens |
|---|---|
| No featured GPU in our directory fits this model. Check back as we add more hardware. | |
GPT-6.2 SolOpenAI · in $2.00 · out $10.00 | $4.40 |
Claude Fable 5.2Anthropic · in $8.00 · out $40.00 | $17.60 |
Gemini 4 NeonGoogle · in $3.00 · out $15.00 | $6.60 |
Grok 4.5xAI · in $2.00 · out $6.00 | $3.20 |
API prices blended at 70% input / 30% output.
Hardware amortisation not included. Run the full ROI calculator for payback math.
Cheapest current cloud rentals with at least 2153 GB VRAM, refreshed hourly.
No current rental listing covers this model’s VRAM requirement on the providers we track.