LIT-135: Qwen3.8-27B

Qwen Team, Alibaba (2026) — Qwen3.8-27B model card and the release post at qwen.ai (blog?id=qwen3.8, whose own title is “Qwen3.8-Max: A New Bar for Coding and Cowork”). No paper; the third url:-only source in the record.

Key takeaways

• Released 14 August 2026 under Apache 2.0: 27.78B parameters, dense, with a 27-layer vision encoder for image and video, a 262,144-token native context extensible to 1M, and a multi-token-prediction head shipped in the checkpoint for EAGLE-style speculative decoding.
• The language model is a 64-layer, 5120-hidden hybrid: 48 Gated DeltaNet linear-attention layers (48 value heads, 16 query/key heads, 128-dim) and 16 gated full-attention layers (24 query heads, 4 KV heads, 256-dim) — three linear layers to one global, so only 16 layers carry a KV cache. full_attention_interval is 4 and layer_types spells the pattern out literally; attn_output_gate is true and output_gate_type is swish, which is where SOTA-134 lands in this checkpoint. Rotation is partial (partial_rotary_factor 0.25) with an mRoPE section of [11, 11, 10] over a 10M theta.
• The configuration is the same as Qwen3.6-27B’s. Fetched from the Hub and compared directly, the two config.json files differ in exactly one key — transformers_version, 4.57.1 against 5.8.0.dev0 — which is a stamp of the library that serialised the file, not a fact about the model. Every architectural field is identical: model_type, the 64 entries of layer_types and their 3:1 pattern, hidden size, head counts and dimensions on both attention paths, attn_output_gate, vocabulary, max_position_embeddings, the MTP layer, and the whole vision block. What changed in four months is post-training. The 3.8 card re-evaluates both in one harness — Claude Code, with the baselines re-run on its own refined benchmark — and reports Terminal Bench 2.1 (Terminus) 73.0 vs 63.4, SWE-bench Pro 61.7 vs 53.5, QwenSWEBench 79.0 vs 49.3, with smaller gains on plain reasoning. QwenSWEBench is in-house.
• Thinking mode is on by default and can be disabled per request. reasoning_effort takes xhigh, medium or low, and the default is xhigh — the deepest of the three, described as being for complex tasks demanding thorough analysis — so the out-of-the-box configuration is the expensive one. preserve_thinking is likewise on by default, carrying reasoning context forward from earlier messages.

Standing in the anthology

Every architectural number above is read from the checkpoint’s own config.json, not from coverage of it; the benchmark comparisons are the model card’s own re-evaluation of both generations in one harness, which is the card’s claim rather than an independent one.

Filed as the current dense instance of the design Qwen3-Next introduced (LIT-136) and Qwen3.5 and 3.6 carried: Gated DeltaNet (LIT-137) for three layers in four, gated attention (LIT-138) for the fourth. It is the second independent adopter of the 3:1 hybrid in the record, with a different linear module from Kimi’s (LIT-133), which is what let SOTA-132 stop calling that replication missing. The pretraining recipe for this checkpoint is still not published, and nothing here is sourced to the card beyond the architecture — but the generation’s design work now is (LIT-152), for a different model in the same family.

No practice of its own: a model report is adoption, not evidence (ADR-017). It describes a recipe in full and isolates none of it, so what it supplies is that somebody shipped these choices — which is a claim about the field, and belongs in the consensus_note: of the practices whose designs it carries. It is cited that way by the hybrid-interleave and multi-token-prediction practices, and it is the sharpest adoption datum in the record for the claim that a generation’s gains can come entirely from post-training: the architecture is identical to its predecessor four months earlier.