Qwen 3.8-2.4t-a95b
Qwen3.8-2.4T-A95B is Alibaba's open-weight, sparse Mixture-of-Experts frontier model (2.4T total / ~95B active parameters) that serves as the downloadable backbone of the hosted Qwen3.8-Max flagship. It pairs a 1,048,576-token context window with hybrid (Gated DeltaNet + Gated Attention) attention, always-on test-time reasoning, and a strong agentic/coding profile — making it the first open-weight Max-class Qwen, positioned against GPT-6 Astra, Claude Opus 5, and Gemini 3.8 Pro.
Empirical Evaluation Results
Architectural Profile & Capabilities
Qwen3.8-2.4T-A95B (OpenRouter slug: qwen/qwen3.8-2.4t-a95b) is the open-weight checkpoint underlying Qwen3.8-Max, released publicly via Hugging Face in August 2026 (repository dated Aug 13, 2026) under the Qwen3.8-Max License. It is a causal-language, sparse Mixture-of-Experts transformer reporting 2.4 trillion total parameters with roughly 95 billion activated parameters per token — a sparsity ratio mirroring the Qwen3 line, with secondary deployment documentation citing ~512 experts and ~10 routed experts plus 1 shared expert per token. Reported attention is a hybrid scheme (Gated DeltaNet combined with Gated Attention) rather than plain MLA/GQA, consistent with 1M-class context ambitions. The variant is text-only on OpenRouter; the hosted Qwen3.8-Max derives from this backbone but adds vision input, a non-thinking mode, built-in tools, and a 1M-token default context — features NOT confirmed for the open-weight checkpoint itself, which appears to run with thinking always enabled. Priced at $2/1M input and $6/1M output, it targets self-hosting of a frontier-class agentic/coding model. On the verified Benchmark Heaven record it posts a SOTA-leaning profile: AA Intelligence Index 39.9, AA Coding Index 71.9, GPQA Diamond 93.5%, HLE (no tools) 42.4%, Terminal-Bench 2.1 at 82.0%, TAU-2 Airline 76.7%, AutomationBench 57.2%, and Creative Writing v3 at 1843 Elo — with measurable weakness on Terminal-Bench 4.0 (11.1%) and GDP PDF (15.6%). Empirical output speed measured ~38 tokens/sec, reflecting heavy reasoning-token overhead from always-on thinking. Note: the web corroboration layer did not expose independently verifiable exact-variant numeric scores beyond this ground-truth set, and several third-party tables conflate the open-weight checkpoint with Qwen3.8-Max columns; to preserve integrity, no proxy or family-derived figures have been appended.
2.4T total parameters / ~95B active parameters per token. Secondary deployment documentation cites approximately 512 experts (~10 routed + 1 shared expert activated per token), ~92 layers and 8,192 hidden size (uncorroborated by the primary model card). Attention is reported as a hybrid Gated DeltaNet + Gated Attention scheme rather than MLA/GQA. Context window: 1,048,576 tokens. Modality: text (OpenRouter); vision is an added capability of the hosted Qwen3.8-Max, not confirmed for this open-weight checkpoint. Thinking appears always enabled on the checkpoint (no documented non-thinking mode), whereas Max supports non-thinking. Weights distributed as full-precision safetensors (213 shards reported) plus an official FP8 variant; served via vLLM, NVIDIA NeMo, and Ascend stacks.
Recommended Workloads & Primary Use Cases
- •Always-on thinking imposes heavy test-time reasoning overhead; measured throughput is only ~38 tokens/sec, so latency-sensitive, high-QPS chat workloads are a poor fit.
- •1M-token context is powerful but prompt-processing (prefill) cost and TTFT scale sharply with input length — very long contexts risk high time-to-first-token on typical serving hardware.
- •Open-weight checkpoint differs from Qwen3.8-Max: no confirmed vision, no non-thinking mode, and built-in tools are a Max-side feature — agentic harness assumptions must be re-validated per-deployment.
- •Agentic/coding strengths are harness-dependent (model card reports Claude Code harness and Claude Opus 4.6 as judge); scores can shift materially with different scaffolds, and Terminal-Bench 4.0 (11.1%) and GDP PDF (15.6%) show sharp degradation on the newest/hardest suites.
- •Deployment economics are non-trivial: a 2.4T-parameter weight set demands large multi-GPU/expert-parallel clusters and substantial VRAM even though only ~95B parameters are active per token; license is a custom Qwen3.8-Max License (not fully open) with regional/usage conditions.
Calls to Qwen 3.8-2.4t-a95b are routed through strict Zero Data Retention inference channels. Prompts and outputs are never stored, indexed, or monitored by Alibaba.
More Models from Alibaba
Compare other engines available in this laboratory.