This is a REAP-pruned and oMLX-quantized derivative of Qwen/Qwen-AgentWorld-35B-A3B.
The goal of this package is practical local agent use on Apple Silicon with 16 Gb unified memory. Tool calling was treated as the release gate, so the chat template is kept both as chat_template.jinja and embedded in tokenizer_config.json, with tool_parser_type set to qwen3_coder.
REAP Pruning
Method: REAP expert pruning
Metric: reap
Calibration rows: 154
Experts per MoE layer: 256 -> 56
Experts pruned per layer: 200
Prune fraction: 78.12%
The pruning pass used tool-calling-oriented calibration rows so the removed experts are the lowest-scoring routed experts for this intended agent workload.
The upstream config declares an MTP layer, but the upstream checkpoint published for Qwen/Qwen-AgentWorld-35B-A3B does not include mtp.* tensors. These artifacts therefore publish a self-consistent non-MTP config instead of advertising a missing draft head.
Compatibility
Expected local targets:
oMLX 0.4.4 or newer
LM Studio with MLX model loading and OpenAI-compatible tool calls
Use greedy decoding for strict tool use and eval runs:
{"temperature": 0, "top_p": 1}
Tool-Calling Verification
Swival core tool suite: 5/5 passed on 2026-06-24 with deterministic greedy decoding and zero failed tool calls.
Swival all-built-ins suite: 5/5 passed on 2026-06-24 with deterministic greedy decoding and zero failed tool calls.
The Swival suites exercise real file, edit, command, planning, checklist, snapshot, grep, outline, URL fetch, batched reads, and shell-tool dispatch through an OpenAI-compatible server. A direct /v1/chat/completions smoke also checks that the server returns structured tool_calls, not just plain XML text.
Known Limitation
Not a good thinking model - I highly recomend disabling thinking.
Long-Context Notes
The 64k no-TurboQuant KV path passed a direct long-context tool smoke at 64,901 prompt tokens with a structured read_file tool call and no hard memory-pressure transition in the oMLX log.
The 128k TurboQuant KV path passed a direct long-context tool smoke at 127,997 prompt tokens with a structured read_file tool call. On the tested oMLX 16 GB guard, this top-of-window stress run still produced hard memory-pressure warnings during prefill, so use it as a functional 128k proof rather than a claim of wide memory headroom. Keep concurrency at 1 for this mode.
