Alibaba Ships Qwen3.6-27B, an Open-Weight Coding Model That Beats Its 397B MoE

Alibaba on Wednesday released Qwen3.6-27B, a dense 27-billion-parameter open-weight model under Apache 2.0 that tops its own 397B-parameter predecessor on every major agentic coding benchmark. The model scores 77.2 on SWE-bench Verified, 59.3 on Terminal-Bench 2.0, and 1487 on Alibaba's internal QwenWebBench, built on a hybrid Gated DeltaNet and Gated Attention layout across 64 layers. Weights went live on Hugging Face and ModelScope, with a 4-bit quantized version running in roughly 17 gigabytes of memory.

The numbers Alibaba is pushing

Qwen3.6-27B ships with benchmark gains that read like a correction rather than an increment. SWE-bench Verified lifts to 77.2 from the 75.0 posted by Qwen3.5-27B, and from the 76.2 posted by the much larger Qwen3.5-397B-A17B Mixture-of-Experts model. Terminal-Bench 2.0 reaches 59.3, matching Claude 4.5 Opus exactly under the same three-hour timeout harness. SkillsBench Avg5 jumps to 48.2 from 27.2 on the previous generation, a 77 percent relative swing.

The Claude comparison is the one Alibaba wants you to notice. On SWE-bench Verified, Claude 4.5 Opus still leads at 80.9. On Terminal-Bench 2.0, the 27B dense pulls even. On GPQA Diamond it lands at 87.8, a hair ahead of Opus at 87.0. The gap between a closed frontier model and a downloadable 55-gigabyte file has narrowed to a handful of points across the benchmarks that matter for coding agents. That is the tell.

Why a 27B dense beats a 397B MoE

The architecture is doing most of the work. Each block here starts with three Gated DeltaNet sublayers and caps off with one Gated Attention layer. Sixteen blocks, stacked. DeltaNet is a linear-attention trick: it scales O(n) with sequence length, not O(n²). That matters for long repositories, where quadratic attention chokes on memory. The standard attention layers sit on an asymmetric 24-query, 4-key/value setup. Smaller KV cache, lower VRAM at serve time.

And then there is Multi-Token Prediction. Qwen trained the model with multi-step MTP, enabling speculative decoding at inference. Candidate tokens get verified in parallel. Throughput climbs without quality loss. A dense 27B model running this architecture on a single high-end GPU can serve an agentic coding workload that previously required routing across a 397-billion-parameter MoE cluster. Alibaba also adds a "preserve_thinking" flag that keeps reasoning traces from prior turns in context, cutting redundant chain-of-thought generation inside multi-step agent loops.

Running it locally

The practical story lives downstream of the benchmarks. Simon Willison ran the Unsloth Q4_K_M GGUF build (16.8 gigabytes) on llama-server and clocked 25.57 tokens per second on a personal machine, generating 4,444 tokens in under three minutes. Quant footprints vary a lot. Q3 builds come in under 15 gigs. Q6 pushes past 25. Pick your trade-off. On a 32-gigabyte MacBook Pro the 4-bit build fits with room to spare, no thrashing, no swap.

Local inference still has honest limits. Developer reports on Qwen3.6-35B-A3B, the sibling MoE released six days earlier, describe the model handling utility functions and small refactors well, then getting shakier on long-range repository state. The 27B dense sits in the same neighborhood. For agentic loops demanding sustained context across a messy codebase, the frontier cloud models still hold the edge.

Alibaba's three-week release sprint

This is the fourth Qwen3.6 release in three weeks. Qwen3.6-Plus dropped April 2, followed by Qwen3.6-35B-A3B on April 16, the closed Qwen3.6-Max-Preview on April 20, and now the dense 27B on April 22. The cadence tells you something. Alibaba looks emboldened, and it is not betting on one architecture. The company is shipping one of each, MoE and dense, open and closed, hosted and API, and letting developers pick.

That matters against the broader backdrop. Earlier this month Alibaba released three closed-source models in three days, a signal the company is walling off the top tier for paid APIs. The 27B dense cuts the other way. Apache 2.0, commercial use permitted, no royalty. For teams building coding agents that can't route source code through a US vendor, the math just changed.

The benchmarks will face independent replication. The open weights are already on disk.

Kimi K2.6 did not release a coding model. It opened the control room.

On Monday, Moonshot AI put a familiar label on a less familiar move. Kimi K2.6 arrived as an open-source coding model, with a benchmark table, a Hugging Face page, a coding CLI, and the usual claims a

The Implicator

Cursor Acknowledges Kimi K2.5 as Composer 2 Base After Developer Spots Model ID

Cursor, the AI coding platform valued at $29.3 billion, publicly acknowledged on March 20 that its new Composer 2 model started from Moonshot AI's open-weight Kimi K2.5. The acknowledgment came less t

The Implicator

DeepSeek’s V3.1 lands quietly—and pushes the open-weight fight

💡 TL;DR - The 30 Seconds Version 🚀 DeepSeek quietly released V3.1 Tuesday—a 685-billion parameter open-source model that scored 71.6% on coding benchmarks, matching proprietary systems. 💰 Tas

The Implicator

Harkaram Grewal

New Delhi

Freelance correspondent reporting on the India-U.S.-Europe AI corridor and how AI models, capital, and policy decisions move across borders. Covers enterprise adoption, supply chains, and AI infrastructure deployment. Based in New Delhi.