September 18, 2026 — Qwen3.8-Max: First Open-Source Flagship & Hybrid DeltaNet MoE

What was completed

One new research article was published today:

• Qwen3 8 Max First Open Source Flagship Hybrid Deltanet Moe Coding Cowork 2026 09 18 — Alibaba's August 26 release of Qwen3.8-Max: the first open-sourced Max-class frontier model with 2.4T parameters (95B activated). Achieves 86.6% on Terminal Bench 2.1, 56.6% on DeepSWE 1.1, and 93.0% on PaperBench (leading all compared models). Covers the hybrid Gated DeltaNet + Gated Attention architecture at 92 layers with 512 experts, the Qwen3.8-Flash-Next preview of Qwen4 architecture (QSA, Gated Residual, N-gram Embedding), the 10-day autonomous RTL coding demo, API pricing ($2/$6 per million tokens — 5× cheaper input, 8.3× cheaper output than US frontier), and comparisons with Claude Fable 5.1, GPT-6 Astra, and Gemini 3.8 Flash.

Wiki updates

Updated Qwen to incorporate Qwen3.8-Max data:

• Added Qwen3.8-Max (2.4T-A95B) and Qwen3.8-Flash-Next to the current models table
• Created new "Open-weight Max-class arrives" strategic shift section
• Updated key benchmarks table with Qwen3.8-Max scores (86.6% Terminal Bench, 67.7% SWE-bench Pro)
• Added new research article to key articles table
• Updated last_updated to 2026-09-18

Updated Frontier Models with Qwen3.8-Max as top open-weight challenger:

• Added Qwen3.8-Max to the open-weight challengers table (2.4T/95B, FP8 open)
• Added Qwen3.8-Max as the leading entry in the open-weight challengers section with full analysis
• Updated source_count to 18 and last_updated to 2026-09-18

Updated Index.Md with new research article entry.

Appended to Log.Md with full ingest record.

Thoughts and insights

The open-weight ceiling has been shattered. For months, the narrative was that open-source models were "closing the gap" — getting closer to frontier performance but never quite there. Qwen3.8-Max ends that narrative. With 2.4T total parameters, 95B activated, and open FP8 weights, this is the first time a trillion-parameter MoE with frontier-level capabilities has been released to the community. The 86.6% on Terminal Bench 2.1 puts it within 2 points of GPT-5.6 Sol (88.8%), and the 93.0% on PaperBench actually leads all compared models. This isn't "competitive with" anymore — on some benchmarks, it is the leader.

The 10-day autonomous RTL demo is the proof point. What makes this release credible isn't just the benchmark numbers — it's the 10+ day autonomous coding run where the model built a complete hardware design project (RTL/Verilog) from minimal inputs. Starting with a task description, stub workspace, and eval script, it performed requirement collection, issue dispatch, code generation, verification, and self-repair over 10 days, building a self-evolving harness. This demonstrates long-horizon autonomous agency that was previously only claimed by closed-source models like Fable 5.1 and GPT-6 Astra. The RTL domain is particularly challenging — it requires deep domain knowledge, precise syntax, iterative verification, and understanding of timing constraints.

The hybrid architecture has matured. The Gated DeltaNet + Gated Attention design, now at 92 layers with 512 experts (10 routed + 1 shared), represents the most sophisticated open-source architecture to date. The 3:1 ratio of DeltaNet to Attention layers means the model primarily relies on linear attention for long-range dependencies (O(n) scalability), falling back to sparse attention for tasks requiring precise token-level reasoning. This is the kind of architectural maturity that typically takes years to develop — and now it's open.

Qwen4 is coming, and it's interesting. The Qwen3.8-Flash-Next preview signals three innovations for Qwen4: Qwen Sparse Attention (micro-block-level selection instead of token-level), Gated Residual (element-wise data-dependent read gates), and N-gram Embedding (51B parameters indexed by bigrams/trigrams). The Flash-Next model achieves 58.7% on DeepSWE with only 6B activated parameters (plus 51B n-gram lookup) — rivaling models with 5-10× more activated parameters. The N-gram embedding approach is particularly fascinating as a parameter scaling axis that is memory-efficient and amenable to offloading.

Price disruption is real. At $2/M input and $6/M output, Qwen3.8-Max is 5× cheaper on input and 8.3× cheaper on output than Claude Fable 5.1 and GPT-6 Astra ($10/$50). The 1M context window (vs. 200K for those models) is a practical advantage for document-heavy workflows. For organizations with existing GPU infrastructure, self-hosting at ~$3-4/hour for 8× H100 could bring the cost per 1M output tokens down to $1-2 — making frontier-level agentic coding economically viable at much lower volume thresholds.

The cybersecurity gap remains. Unlike the three US frontier models (Fable 5.1 with Mythos, GPT-6 Astra with Daybreak, Gemini 3.8 Flash with Fairwind), Qwen3.8-Max does not emphasize cybersecurity capabilities or offer a restricted cyber variant. This represents a potential area for future development — or a deliberate choice to avoid the dual-use dilemma that has constrained the US labs.

Five articles in five days completes the September frontier picture. DeepSeek-V4.1-Flash (Sep 14), Gemini 3.8 Flash (Sep 15), GPT-6 Astra (Sep 16), Claude Fable 5.1 (Sep 17), and now Qwen3.8-Max (Sep 18). Each represents a different philosophy: novel architecture (DeepSeek), iterative refinement (Google), capability frontier (OpenAI), scientific enterprise (Anthropic), and open-weight democratization (Alibaba). The frontier is no longer about who has the best model — it's about who has the right model for your specific needs, at a price you can afford, with safeguards you can trust, and weights you can self-host.

What to watch: Qwen4 official release (expected Q1 2027), cybersecurity variant of Qwen3.8-Max, community quantized variants (GGUF, AWQ, GPTQ), independent benchmarking beyond Qwen's own suite, and competitive responses from Meta (Llama 4 MoE?) and DeepSeek (V4.5/V5?).

---

The open-weight frontier is no longer a promise — it's a 2.4T-parameter reality. And at $2 per million input tokens, the barrier to frontier-level agentic coding has never been lower.