On the night of August 12 (Beijing time), the Alibaba Cloud ModelScope community account published a brief announcement: the Qwen team has officially released the full weights of Qwen3.8-2.4T-A95B — the first Qwen-Max-class model to be completely open-sourced. The model has 2.4 trillion total parameters with 95 billion activated per token, a native 262,144-token (256K) context extendable to 1,010,000 tokens (1M), text-only input, and a mandatory thinking mode. It ships day-one with Transformers / vLLM / SGLang / TokenSpeed inference stacks.
How This Open Release Compares in Scale
| Model | Total Params | Active Params | Native Context | Open / Closed | |-------|-------------|---------------|----------------|---------------| | Qwen3.8-2.4T-A95B | 2.4T | 95B | 256K (ext. 1M) | Open | | Moonshot Kimi K3 | 2.8T | 104B | 1M+ | Semi-open (mainly distilled) | | DeepSeek V4-Pro | 1.6T | ~49B | 1M | Semi-open (partial architecture) | | Meta Llama 4 Behemoth | ~2T | ~100B | 128K | Closed | | xAI Grok 4.6 | 1.5T | 8–16B | 256K | Closed | | Claude Fable 5 Max | Undisclosed | — | 200K | Closed | | Microsoft MAI-Thinking-1 | 1T | 35B | 256K | Closed |
Placed in this table, the takeaway is blunt: by August 2026, the "scale gap" between open and closed camps has shrunk from 10x to about 1.5x. Qwen's open release is a 2.4T total / 95B active combination; Kimi K3 (2.8T / 104B) is larger but semi-open (the main model is a distilled derivative); DeepSeek V4-Pro (1.6T / 49B) is lighter; Meta's Llama 4 Behemoth has been 2T+ all along but stays closed.
Qwen3.8's 2.4T sits in the same tier as Kimi K3's 2.8T, one tier above DeepSeek V4-Pro's 1.6T, and in the same tier as closed models with undisclosed totals like Fable 5 Max and GPT-5.6 Sol.
This marks the "de-budgetization" of open models — open releases used to be mini versions of closed flagships; this time Qwen open-sourced a Max-class model directly.
Three Architecture Details Worth Knowing
1. 512 experts, 10 routed + 1 shared per token.
This sets a new 2026 MoE benchmark. Ling-3.0-tiny (08-11) activates 8 routed + 1 shared = 9 per token; Ling-3.0-flash (08-09) has 5.1B active / 124B total = 1/24 sparsity; Kimi K3 uses 256 experts with 12 active per token. Qwen3.8 goes straight to 512 experts / 10+1 active = 1/218 sparsity — the densest expert pool with the sparsest activation among current Chinese open-source MoEs. That means:
- More expressible expert combinations at the same total parameter count
- Lower per-inference compute due to sparse activation
- Harder routing learning (512-choose-10 is a ~5.2×10¹¹ combinatorial space)
- Vision input (image understanding)
- Non-thinking mode (direct answers without reasoning traces)
- Built-in tools (search, calculator, code execution)
- Default 1M-token context
- 08-08 NVIDIA Cosmos 3 (foundation model, 1.3B data points, OpenMDW 1.1 commercial license)
- 08-08 Unitree lists on the STAR Market at 219x PE
- 08-09 Ant Group open-sources Ling-3.0-flash (5.1B active / 124B total / 1/64 sparsity, with domestic inference stack SGLang HiCache + Mooncake)
- 08-11 Ant Group open-sources Ling-3.0-tiny (1.3B active / 7.9B total / 128 routed experts)
- 08-12 Alibaba open-sources Qwen3.8-2.4T-A95B
- 08-12 NVIDIA Nemotron 4 trillion-parameter R&D info leaks
- DeepSeek: V4-Pro 1.6T is already semi-open — will V5 (if it exists) be fully open or split closed?
- Moonshot Kimi: K3's 2.8T is mostly open-sourced via distillation; full open-sourcing lags half a beat.
- Meta Llama: Llama 5 will likely keep the dual-track "open model vs closed flagship" approach.
- Closed camps (OpenAI / Anthropic / Google): Can they keep the "closed = premium" label against Qwen's explicitly open Max-class posture?
- No vision input / no native tool calls — text-only, forced thinking. Multimodal or tool-calling requires paid cloud Qwen3.8-Max or your own modifications.
- PaperBench / IFBench scores are self-reported by Qwen — no independent third-party verification yet.
- 512-expert routing learning curve unverified independently; fine-tuning reproduction costs are non-trivial.
- Commercial license not spelled out — currently only points to a "Qwen3.8-specific license file" with undisclosed terms.
- Hardware barrier excludes individual developers — unless using extreme quantization at a quality cost.
- Same tier as Kimi K3 — Qwen doesn't clearly win on total parameters; it wins on "fully open" and "native 256K with forced thinking" as product posture.
- https://modelscope.cn/Qwen (ModelScope homepage; specific model page URL not stated)
- https://finance.sina.cn/tech/2026-08-13/detail-ininarri3875962.d.html (Sina Finance / IT Home: Qwen3.8-2.4T-A95B open-sourced)
- https://www.163.com/dy/article/L46N0SML0556I485.html (NetEase: Alibaba officially releases Qwen3.8-2.4T-A95B weights)
- https://www.donews.com/news/detail/8/6668953.html (DoNews: Alibaba Cloud open-sources Qwen3.8-2.4T-A95B)
- https://setupai.cc/update/alibaba-opens-1ea180b0 (SetupAI: technical architecture analysis)
- https://wpnews.pro/news/alibaba-releases-open-qwen3-8-with-2-4t-total-95b-active-parameters (Web Pulse: open weights and Qwen3.8-Max differences)
2. 92 layers: 69 linear attention + 23 full attention.
Continuing the hybrid attention approach since Qwen3.5: 69 layers use linear attention to shrink KV cache memory, 23 layers use full attention to preserve precision. This is what makes 256K / 1M long context feasible on finite memory — pure full attention can't fit a 1M-context KV cache at all.
3. Multi-step Multi-Token Prediction (MTP) training.
MTP, proven in the DeepSeek V3 / R1 lineage, trains the model to predict multiple tokens at once for speculative decoding at inference. Qwen3.8 ships with a built-in MTP draft head usable directly by vLLM / SGLang / TokenSpeed.
A Conflicting Set of Numbers
The key to understanding open LLMs is separating "can I run it" from "can I afford to run it":
| Format | Size | Hardware | |--------|------|----------| | BF16 full precision | ~4.9TB | DGX-class (H100 / RTX PRO), 8+ GPUs | | FP8 quantized | ~2.5TB | H100 / H200, 1–8 GPUs | | Unsloth TQ1_0 (1-bit) | <400GB | Single RTX 4090 / Mac Studio (M2 Ultra 192GB) | | 4-bit GPTQ | ~1.2TB | Single RTX 5090 / Mac Pro |
Qwen officially recommends FP8. The 1-bit Unsloth TQ1_0 compresses 4.9TB down to under 400GB, but throughput is low and long-context quality degrades significantly.
After the earlier "local-run" narrative (Qwen3.5 30B, Qwen3.6 27B), crossing the Qwen3.8-2.4T-A95B line creates a new pattern: open weights, but hardware requirements return to the data center. You can read the code, do research, and fine-tune — but you can only run quantized versions. This is a "sovereign AI / national AI team" path, not an "indie developer" path.
Mandatory Thinking Mode: The Open Version's Capability Seal
The open Qwen3.8-2.4T-A95B requires thinking mode — every response is automatically wrapped in thinking tags, with no non-thinking option and no vision input. The cloud Qwen3.8-Max adds on top of the same base weights:
In other words, Qwen3.8-A95B is a "half-finished core" — weights, general language ability, long context, and agent-task foundations are open, while vision / tools / fast responses stay in the paid cloud version. Researchers get pure-text long-context and agent work from the open weights; product builders pay for cloud Qwen3.8-Max. Qwen wins both ways: community from open source, revenue from the cloud tier.
One Piece of the H2 2026 "Sovereign AI" Puzzle
August 2026 so far:
Qwen3.8-2.4T-A95B is not "just another open release" — it is the first time Qwen open-sourced a Max-class model. Previously the largest open Qwen was Qwen3-235B (235B total, 22B active), a tier below the closed Qwen3.5-Plus flagship at 720B. Opening 2.4T directly locks Qwen's open-source ceiling at Max parity.
Implications for other players:
Limitations and Unknowns
What Qwen-Max Going Open Means
Qwen3.8-2.4T-A95B is, piece by piece, a stack of hardware-account numbers: largest Chinese open model, forced thinking, native 256K, MTP, 512-expert sparsity. Each is a continuation of an existing trend; combined, they form an inflection point — and that inflection is named "Qwen decided to fully open-source its Max-class model."
It signals that in H2 2026, competition among Chinese open LLMs shifts from "whose active parameters are smaller and cheaper to run" to "who dares to open more total parameters, and whose release posture is more open." After Qwen dumped the full weights on the night of August 12, DeepSeek / Kimi / MiniMax / ByteDance / Zhipu all face the same stress test: do you dare open-source a Max-class model?
This was unthinkable in 2025 (Qwen3-Max: 720B total, 20B active, closed — nobody dared). After August 12, 2026, the global open-LLM narrative shifts from "US majors + closed Qwen Max" to "Qwen's open Max + how everyone else responds."
Qwen's move here isn't a technical breakthrough — it's a commercial posture. The open-source camp is starting to pressure the closed players back.
---
Sources