[JustVugg]

Qwen/Qwen3-Coder-30B-A3B-Instruct (Apache-2.0, qwen3_moe) runs on the qwen36 engine: text, tool calls and streaming, through coli chat, coli serve and coli web. Its REAP prunes (same architecture, 103 experts) resolve to the same family under their own name.

What it took

Qwen3 MoE differs from the Qwen3.5/3.6 layers this engine was written for in four places: no DeltaNet (every layer attention), no attention output gate, no shared expert, and plain RMSNorm weights with rotary over the whole head.

• Engine (qwen36.c): shared_inter 0 is accepted for a routed model, and the shared expert is skipped at load and in both compute paths. Attention multiplied a missing gate as sigmoid(0) = 0.5 on every layer; without a gate the output now passes through. Qwen3.5/3.6/3.8 always carry the gate, so their numerics are unchanged: the 35B, 2.4T and 27B-dense tiny oracles are 16/16 before and after.
• Converter: for model_type: qwen3_moe it writes all-attention layer_types, shared_inter: 0, partial_rotary_factor: 1.0, zero_centered_norms: false, and M-RoPE sections that cover the rotary half (one position per token, so ordinary RoPE). It also reads num_local_experts, the name transformers 5 writes.
• Registry: qwen3_moe maps to the qwen36 family, the variants are named by geometry (Qwen3-Coder-30B-A3B, qwen3-coder-30b-a3b-colibri; Qwen3-Coder-REAP-25B-A3B), and the planner no longer requires DeltaNet keys when every layer is attention.
• Gateway: Qwen3-Coder's template is detected from the container (detect_chat_flavor) and rendered by render_chat_qwen3_coder: its XML tool block, its default system line, tool responses sharing one user turn, no thinking. enable_thinking is forced off for it (otherwise a request asking for it would file the whole answer as reasoning, as [Bug]: OLMoE streaming responses drop generated content #984 did on OLMoE). The Qwen parser reads its calls, and now reads a declared boolean written as True/False as a boolean, which is how the model writes it (seen on the real model; Qwen's own parser does the same).

Correctness

• Tiny oracle in CI: make_qwen36_tiny.py --geometry qwen3-coder-30b builds the model with transformers' Qwen3MoeForCausalLM (RMSNorm weights drawn around 1, so a zero-centered reading cannot pass); converted, the engine matches it 16/16 at caches of 1 and 8. Before the gate fix it matched 1/16, with each layer's attention output exactly half of transformers'.
• Template: tests/test_qwen3_coder_chat_template.py renders the release's chat_template.jinja (fixture, sha256 pinned) with jinja2 the way transformers does and compares byte for byte: single turns, system, history, tools with extra schema keys, an agent loop with parallel calls and tool results, a tool result last, and the open turn.
• Real checkpoint against bf16: a 325-token code question and answer, teacher forced, every position's logits against a reference that reads the bf16 release one layer at a time in f32 (that reference matches transformers to 4.8e-7 on the tiny model):

experts	dense trunk	top-1 = bf16	top-5 overlap	mean |Δ log p|
int4 gs64	int8 (default)	96.9%	91.1%	0.176
int4 gs64	f32	96.0%	91.7%	0.134
int8	int8	95.1%	95.2%	0.098
int8	f32	98.1%	97.8%	0.022

Measured

Ryzen 7 PRO 8700GE, 64 GB, NVMe RAID, nothing else running; CLI from a cold cache, 128 tokens, dense trunk int8:

container	experts cached per layer	decode	resident
int4 gs64 (19 GB)	128	8.5 and 9.6 tok/s (two runs)	15.2 GB
int8 (30 GB)	128	6.4 tok/s	25.1 GB
int4 gs64	32	5.1 tok/s	6.5 GB
int8	32	3.8 tok/s	9.6 GB

coli chat with --cap 128 and a warm cache streams a 577-token answer at 13 tok/s. Conversion takes under a minute per container.

On the real model, every front end

• Gateway: a code answer, enable_thinking: true answered in content (not reasoning), a tool call run_tests({"path": "/srv/app", "verbose": true, "workers": 4}) with both types restored, and streaming.
• coli chat (attached) and coli web:

One thing to know, not changed here: like the rest of this engine, coli serve/coli chat size the expert cache from --cap, not from --ram (docs/qwen36.md); without --cap the engine starts at the routing floor (8 experts per layer) and runs at about 2.7 tok/s. coli plan already computes the right cap (128 here); passing it through for this family is the launcher change discussed in #1823.

docs/qwen36.md documents the checkpoint, the commands, the quality and the speed tables.