Qwen3.8-2.4T-A95B Benchmarks vs Opus 4.8 and GPT-5.6 Sol
Qwen3.8-2.4T-A95B benchmark results compared against Opus 4.8, Fable 5, and GPT-5.6 Sol across coding-agent and general-agent tests.
What is Qwen3.8-2.4T-A95B and how does it compare to Opus 4.8 and GPT-5.6 Sol?
Qwen3.8-2.4T-A95B is an open-weight mixture-of-experts language model from Alibaba’s Qwen team, built on the Qwen3.5 architecture, with 2.4 trillion total parameters and 95 billion activated per token. On released benchmark tables it beats or closely trails Opus 4.8 and GPT-5.6 Sol on several coding-agent and general-agent tests, notably scoring 93.0 on PaperBench and 86.6 on Terminal Bench 2.1, while landing behind Fable 5 on most agentic tasks. It marks the first time Qwen has released a Max-class model as open weights.
TL;DR
- Qwen3.8-2.4T-A95B is a mixture-of-experts model with 2.4 trillion total parameters and 95 billion active parameters per token, using 512 experts with 10 routed plus 1 shared expert activated at a time.
- The model natively supports a 262,144-token context window, extensible up to roughly 1,010,000 tokens, and pairs Gated DeltaNet linear attention layers with Gated Attention layers in a 23-block repeating pattern.
- On Terminal Bench 2.1, Qwen3.8-Max scores 86.6, ahead of Opus 4.8 (84.6) and Fable 5 (84.6), and behind GPT-5.6 Sol (88.8).
- On PaperBench, Qwen3.8-Max leads the comparison group outright with 93.0, ahead of GPT-5.6 Sol (90.5), Fable 5 (88.8), and Opus 4.8 (80.3).
- Qwen3.8-Max trails Fable 5 on most general-agent benchmarks like CoWorkBench, WorkSpaceBench, and JobBench, suggesting its strength is more concentrated in coding and research-style tasks.
- Compared to its predecessor, Qwen3.7-Max, the new model shows large jumps across nearly every benchmark, including a jump from 21.6 to 56.6 on DeepSWE 1.1.
- The model ships with flexible reasoning controls, including a
reasoning_effortparameter and apreserve_thinkingflag for carrying reasoning context across turns.
Plans first. Then code.
Remy writes the spec, manages the build, and ships the app.
What are the headline coding-agent benchmark results?
Qwen3.8-Max was tested across a set of coding and software-engineering benchmarks alongside Opus 4.8, Fable 5, and GPT-5.6 Sol (max setting). The results are mixed but generally competitive:
- Terminal Bench 2.1: Qwen3.8-Max scores 86.6, beating Opus 4.8 and Fable 5 (both 84.6) and trailing only GPT-5.6 Sol (88.8).
- SWE-bench Pro: Qwen3.8-Max reaches 67.7, ahead of GPT-5.6 Sol (64.6) and Qwen3.7-Max (60.6), but behind Fable 5 (80.0) and Opus 4.8 (69.2).
- DeepSWE 1.1: Qwen3.8-Max hits 56.6, a large improvement over Qwen3.7-Max’s 21.6, though still behind GPT-5.6 Sol (73.0) and Fable 5 (70.0).
- NL2Repo-Bench: Qwen3.8-Max scores 55.9 versus Opus 4.8’s 69.4; no scores were listed for Fable 5 or GPT-5.6 Sol on this test.
- FrontierSWE: Qwen3.8-Max posts 73.5, ahead of Opus 4.8 (70.0) but behind Fable 5 (88.8).
- PaperBench: Qwen3.8-Max tops the field at 93.0, ahead of GPT-5.6 Sol (90.5), Fable 5 (88.8), and Opus 4.8 (80.3).
- AndroidBench: Qwen3.8-Max scores 75.1, ahead of GPT-5.6 Sol (74.0) and Opus 4.8 (69.8), but behind Fable 5 (84.5).
- QwenSWEBench: Qwen3.8-Max scores 80.7, ahead of GPT-5.6 Sol (73.5) but behind Fable 5 (86.3) and Opus 4.8 (84.0).
- QwenQoderBench: Qwen3.8-Max scores 58.4, behind Opus 4.8 (62.7) and Fable 5 (63.1), but ahead of GPT-5.6 Sol (53.8).
The pattern across coding benchmarks is consistent: Qwen3.8-Max is rarely the top scorer, but it’s almost never far off the leader, and it clears GPT-5.6 Sol or Opus 4.8 on several individual tests.
How does it perform on general-agent tasks?
The general-agent category, covering multi-step workflow and task-completion benchmarks rather than pure coding, tells a somewhat different story. On CoWorkBench, Qwen3.8-Max scores 74.8, close behind Fable 5’s 75.9 and ahead of Opus 4.8 (72.3) and GPT-5.6 Sol (71.5). On WorkSpaceBench, it scores 67.7, again trailing Fable 5 (68.7) but ahead of the other two models. On JobBench, the gap widens: Qwen3.8-Max’s score wasn’t fully listed in the available data, but Fable 5 clearly leads at 57.4 versus Opus 4.8’s 48.4 and GPT-5.6 Sol’s 45.4.
The broader takeaway is that Fable 5 consistently performs strongest on general-agent, real-world workflow style benchmarks, while Qwen3.8-Max’s advantages show up more clearly in coding-specific and research-oriented tasks like PaperBench and Terminal Bench.
What’s under the hood architecturally?
Qwen3.8-2.4T-A95B is a mixture-of-experts (MoE) transformer with 2.4 trillion total parameters, of which 95 billion are activated for any given token. That activation ratio, roughly 4%, is what makes a model this large practical to run: compute cost scales with active parameters, not total ones.
The architecture repeats a block structure 23 times, each block combining three Gated DeltaNet-to-MoE units followed by one Gated Attention-to-MoE unit. Gated DeltaNet is a linear attention mechanism, used here for most of the sequence-mixing work, with Gated Attention (a more conventional attention mechanism) inserted periodically. This hybrid linear/full-attention design is common in newer large models because pure full attention becomes expensive at long context lengths, while pure linear attention alone can lose some modeling quality.
Built like a system. Not vibe-coded.
Remy manages the project — every layer architected, not stitched together at the last second.
The MoE layer routes each token to 10 experts out of a pool of 512, plus one shared expert that’s always active. This routing pattern (sparse routed experts plus a dense shared expert) is intended to combine the efficiency of sparse MoE with more stable, always-on capacity. Context length is 262,144 tokens natively, with support for extension up to roughly 1,010,000 tokens, putting it in the range of the largest long-context models currently available.
Is Qwen3.8 worth using over Opus 4.8 or GPT-5.6 Sol?
Whether Qwen3.8-Max is a better choice than Opus 4.8 or GPT-5.6 Sol depends heavily on the workload. For terminal-based coding agents and research-paper-style tasks (Terminal Bench, PaperBench), it’s competitive with or better than both closed models. For deeper software engineering work involving large pull-request-scale changes (SWE-bench Pro, DeepSWE), it lags behind Fable 5 and sometimes Opus 4.8, though it clearly outperforms its own predecessor, Qwen3.7-Max, by a wide margin on nearly every test.
The fact that Qwen3.8 is released as open weights is itself a meaningful differentiator. Opus 4.8 and GPT-5.6 Sol are closed models accessible only through their respective APIs, while Qwen3.8-2.4T-A95B’s weights can be self-hosted using inference engines like vLLM or SGLang. That changes the calculus for teams that care about cost control, data residency, or fine-tuning access, even if raw benchmark scores are a percentage point or two behind a closed frontier model on any given test.
For teams that want a managed version without running the weights themselves, Qwen also offers Qwen3.8-Max as a hosted API product with additional features like vision input, non-thinking mode, and a default 1M-token context window.
Frequently Asked Questions
What does “2.4T-A95B” mean in the model name?
It describes the parameter structure: 2.4T refers to 2.4 trillion total parameters across the model, and A95B refers to 95 billion parameters actually activated for any given token, since it’s a mixture-of-experts model that only routes each token through a subset of its experts.
How does Qwen3.8 compare to its predecessor, Qwen3.7-Max?
Qwen3.8-Max shows substantial gains over Qwen3.7-Max across nearly every listed benchmark. The largest jump is on DeepSWE 1.1, rising from 21.6 to 56.6, but similar large gains appear on FrontierSWE (40.7 to 73.5) and PaperBench (64.8 to 93.0).
Is Qwen3.8 open source?
Qwen3.8-2.4T-A95B is released as open weights on Hugging Face, in a format compatible with common inference engines like vLLM and SGLang. A separately branded hosted version, Qwen3.8-Max, is also available through Qwen Cloud with extra managed features.
Which benchmark categories does Qwen3.8 lead in?
Among the tests compared, Qwen3.8-Max’s strongest relative results are on PaperBench, where it leads all four models listed, and Terminal Bench 2.1, where it beats Opus 4.8 and Fable 5 and trails only GPT-5.6 Sol.
Does Qwen3.8 support long context windows?
Yes. It natively supports up to 262,144 tokens of context and can be extended to approximately 1,010,000 tokens, which is among the longer context ranges available in current large language models.
