Model Scale and Activation Details
Alibaba released the weights for Qwen3.8-2.4T-A95B on ModelScope and Hugging Face. The model contains 2.4 trillion parameters in total. It activates 95 billion parameters for each token processed. This configuration follows the architecture established in the Qwen3.5 series.
The model supports a native context length of 262144 tokens. That length extends to 1.01 million tokens under the appropriate configuration. These numbers position the release as the first public distribution of weights from Alibaba's Max-level flagship line.
Qwen3.8-Max serves as the official production version built directly on Qwen3.8-2.4T-A95B. It became available on August 3 and adds capabilities such as visual input handling and built-in tools. The open weights release therefore provides access to the underlying model that supports those extended features.
A smaller variant, Qwen3.8-27B, is listed as forthcoming. The current release focuses on the large-scale mixture design that separates total parameters from active parameters per token. This distinction determines both training cost and inference efficiency for downstream users who download the weights.
Context Window and Architecture Base
Qwen3.8-2.4T-A95B contains 2.4 trillion total parameters and activates 95 billion of them for each token processed. The model follows the same underlying architecture as Qwen3.5 while delivering measurable gains on programming tasks, office workflows, scientific research, and extended agent operations that span multiple steps.
The context window starts at 262144 tokens in its native configuration. This length supports direct handling of lengthy documents or codebases without additional segmentation. The design also permits extension to 1.01 million tokens when longer sequences are required.
Local execution becomes feasible once combined device RAM and VRAM exceed 410 GB, according to tests using the Unsloth-Desktop tool. The architecture therefore balances scale with practical deployment constraints for users who meet the hardware threshold.
These choices position the model as Alibaba's largest open-weight release to date, with the context and parameter structure directly supporting the reported benchmark results in areas such as PaperBench and OSworld-Verified. Details on further architectural modifications beyond the Qwen3.5 base remain limited in the initial announcement.
Benchmark Results on PaperBench and OSworld
Information on Qwen 3.8 Max performance on PaperBench and OSworld has not appeared in the initial release materials from Alibaba. Details on these evaluations are still emerging.
The company instead emphasized results from other agentic coding benchmarks. On Terminal Bench and SWE-Bench Pro the model matches the level of Fable and GPT 5.6 Soul. It falls behind on more demanding suites such as Deep SWE 1.1. Alibaba also described a single continuous run that lasted more than 16 days. The system began with an empty repository and produced a complete coding harness named OMI CLI, including its own state machine, dispatcher, and testing loop.
These demonstrations point to reliable operation over extended periods without human intervention. The 2.4 trillion parameter model with 95 billion active parameters is presented as a direct rival to closed frontier systems in long-horizon coding work. Because the weights will become publicly available shortly after announcement, independent groups will soon be able to run the model on PaperBench, OSworld, and similar suites. Until those numbers are published, assessments rest on the agentic tasks already reported and on the autonomous demo described above.
Agentic Coding and Long-Running Task Performance
The Qwen3.8-2.4T-A95B model is positioned as a direct competitor to closed frontier systems such as Anthropic’s Fable line and OpenAI’s GPT 5.6 Soul, with its strongest results reported on coding and long-running agentic tasks. The open release of its weights marks the first time Alibaba’s Qwen team has made its flagship Max-level model available in full to developers and researchers worldwide.
Because the model runs locally once downloaded, teams can integrate it into extended workflows without depending on external API availability or policy changes. This permanence supports sustained agentic sessions that may span hours or days, where interruptions from service restrictions would otherwise break continuity. Researchers and startups gain immediate access to near-frontier capability at no recurring cost, enabling experiments that would be impractical under closed-model pricing or rate limits.
The source article notes that performance rivals Fable 5 in these domains, though detailed benchmark numbers for agentic coding remain limited in public reporting. Previews of the related Qwen3.8-Max are already accessible through Alibaba Cloud services including Token Plan, Qoder, and QoderWork, giving users an early view of the capabilities now shipping as open weights.
The decision to release the full 2.4 trillion parameter model carries the standard open-weight trade-off: broad experimentation accelerates progress for independent developers, yet removes any post-release control if misuse concerns emerge later. For agentic coding workloads, that trade-off favors those who need persistent, customizable access over those who prefer managed oversight.
Multimodal Features and Tool Integration
Alibaba's release notes for Qwen3.8-Max do not specify multimodal input or output capabilities. The available documentation focuses instead on language-based performance across extended sequences.
The model supports a native context window of 260,000 tokens, expandable to 1,010,000 tokens. This length directly addresses requirements for long-horizon agent workflows, where a single session may span multiple hours of interaction and thousands of intermediate steps. Alibaba states that the architecture builds on Qwen3.5 and delivers gains in programming, office automation, research assistance, and sustained agent tasks.
Details on explicit tool-calling interfaces or external API orchestration remain limited in the current announcement. The emphasis on independent completion of complex tasks over extended periods implies internal mechanisms for planning and state management, yet the company has not published concrete specifications for function calling formats, memory structures, or integration with external environments.
Developers testing the preview through Alibaba Cloud services can observe how the model maintains coherence across long traces, but formal documentation on multimodal encoders or standardized tool-use protocols has not yet appeared. Further technical reports will be required to clarify these dimensions.
Local Hardware Requirements for Inference
The Qwen3.8-2.4T-A95B model scales to 2.4 trillion total parameters, with 95 billion active per query. No official documentation has yet specified the hardware configurations needed to run inference on local systems. Details on memory footprint, quantization options, and minimum GPU or CPU setups remain unavailable from the release materials.
Alibaba has made the weights publicly available, marking the first time a Qwen-Max class model can be downloaded for local use. This step opens the possibility of running the model without cloud services, but the sheer parameter count means practitioners will need substantial resources even with optimizations such as mixture-of-experts routing that activates only the 95 billion parameters per forward pass.
The underlying architecture builds on Qwen3.5, and the training regimen emphasized multi-day workflows and nested directory structures. Those design choices affect how the model behaves during extended sessions, yet they do not translate directly into published guidance on VRAM consumption or throughput on consumer or workstation hardware. Independent benchmarks of inference speed or memory usage have not appeared alongside the weights.
Until Alibaba or third-party reports supply concrete measurements, anyone planning local deployment must treat the hardware requirements as an open variable. The parameter scale alone indicates that even sparse activation will demand high-end accelerators, and practical testing will be required to determine feasible batch sizes and context lengths on any given machine.
Comparison Against Closed-Source Flagships
Alibaba positions Qwen3.8-2.4T-A95B as the open-weight release of its Qwen-Max flagship. The model matches performance levels across agent harnesses tied to closed-source systems, including Claude Code and Codex. A bar chart from the release shows results in QwenWork, Claude Code, Codex, OpenClaw, and Hermes sitting close together, with no single environment showing a clear advantage.
Qwen states the model was not tuned for any one harness. Training expanded to cover multi-day workflows and varied directory structures during reinforcement learning, which supports consistent behavior when plugged into third-party agent setups. The model supports both OpenAI Chat Completions format and Anthropic API protocol, allowing direct substitution in existing codebases built for Claude or Codex.
Direct head-to-head benchmarks against closed models such as GPT-4o or Claude 3.5 Sonnet remain absent from the initial release materials. Details on this are still emerging. The open weights and broad API compatibility instead emphasize portability over claims of outright superiority in any single closed environment.
Deployment Channels on Hugging Face and ModelScope
Alibaba has placed the open weights for the 2.4 trillion parameter MoE model on Hugging Face under the Qwen organization. The repository supports direct downloads and integration with existing inference runtimes, including vLLM and SGLang. These frameworks allow operators to run the model with standard tooling rather than custom kernels, though the 95 billion active parameters per forward pass still require substantial GPU memory and careful expert routing configuration.
The model ships with a native context length of 262144 tokens that can extend beyond one million tokens. Practitioners deploying on Hugging Face therefore need to allocate KV cache capacity accordingly, especially for long-horizon agent workloads or large codebases. Early users report that the same checkpoints load without modification into the listed inference stacks, reducing the engineering effort typically required when moving from closed API models to self-hosted versions.
Details on ModelScope distribution remain limited in current announcements. The release notes focus on Hugging Face availability and Alibaba Cloud Model Studio access, leaving open questions about additional mirrors or region-specific hosting. Organizations evaluating production rollout should verify checksums and licensing terms directly from the Hugging Face repository before scaling beyond initial testing.
Upcoming Smaller Qwen3.8 Variants
Details on upcoming smaller Qwen3.8 variants are still emerging. The research material centers on the Qwen3.8-Max release, which uses a Sparse Mixture-of-Experts architecture with hybrid attention. This model holds 2.4 trillion total parameters yet activates only 95 billion for each request. Roughly 4 percent of the full parameter set runs at any time, which reduces inference cost while preserving overall capacity.
Alibaba scheduled the preview for July 19, 2026, and the production release for August 3, 2026. The model accepts text, images, and video as inputs and produces text output. It supports a context window of up to 1 million tokens and a maximum of 128K output tokens. Hybrid thinking mode is available, targeting complex coding, visual analysis, research, professional work, and long-horizon agent tasks.
Developers reach the model through Alibaba Cloud Model Studio or the Qwen3.8-Max API on GPT Proto, where it sits alongside other text, image, and video models under a single account. No comparable specifications appear for smaller variants in the current documentation. Parameter counts, context lengths, or release timelines for those models remain unspecified. As more announcements surface, the pattern of scaling total parameters while controlling active parameters may extend to lighter versions, though confirmation awaits further public statements.
Impact on Open-Weight Model Availability
Alibaba has released the Qwen3.8-2.4T-A95B model as open-weight. The release includes a sparse mixture-of-experts design with hybrid attention that totals 2.4 trillion parameters while activating 95 billion for each token processed. This configuration supports inputs across text, images, and video, produces text outputs, and maintains a context window reaching 1 million tokens with maximum output lengths of 128,000 tokens.
Nvidia has published initial inference measurements for the model on its GB300 NVL72 rack-scale system. The FP8 implementation, without further tuning, exceeds 4,000 tokens per second per GPU and 350 tokens per second per user. These figures establish early performance baselines for hardware optimized to run the architecture at scale.
The open-weight distribution also exposes hybrid thinking with adjustable reasoning effort, along with function calling, structured output, web search, and context caching. Batch inference and fine-tuning capabilities are not currently supported.
The combination of these specifications in an open-weight package allows organizations to evaluate and deploy the model on compatible systems without licensing restrictions from the developer. Performance data from Nvidia provides a concrete reference point for organizations planning hardware configurations around the 95 billion active parameter footprint.