Cyber-Tiel-Coder-35B-A3B-MLX-Attention8-BF16Recurrence-MTP

Cyber-Tiel Coder 35B-A3B — Attention8 / BF16 Recurrence / native MTP

A 26.17 GiB MLX/MTPLX release of the abliterated Ornith 1.5 35B-A3B parent. It was streamed from pinned Huihui BF16 weights (7269f0953054d68b3cacbd9ad729921258732007) and uses the Cyber-Tiel publisher's Sharp chat template (c1471894c53a204b24fce56d673e9ce52c17310b). This is an independent mixed-precision quantization, not the publisher's oQ4e weights or cyber importance-matrix quantization.

Immutable weight revision: 1e6025469055332ebea7685477fc9a1ae22a6937. integrity.json gives SHA-256 and size for every file at that revision, including its initial README; this card-only edit leaves the weights unchanged.

The architecture has 40 layers (30 linear attention, 10 full attention), 256 experts with 8 active per token. The artifact contains 17 text shards (23.76 GiB), one vision shard (0.83 GiB), and a BF16 native-MTP sidecar (1.57 GiB). Source and model weights are MIT licensed.

Suggested runtime: MTPLX 2.11.2 with MLX 0.32.2, sustained native MTP at depth 2, 65,536-token context, unquantized KV cache, Sharp tokenizer template, and thinking off. Images require MTPLX MTP generation mode. Example:

mtplx serve --model Shiftedx/Cyber-Tiel-Coder-35B-A3B-MLX-Attention8-BF16Recurrence-MTP \
  --profile sustained --generation-mode mtp --depth 2 \
  --context-window 65536 --reasoning off

Local checks: On an Apple M4 Max with 64 GiB unified memory, temperature 0 and thinking disabled, code smoke passed 3/3 (~99 generated tokens/s), a harder coding/reasoning/context set passed 6/10 (~80 tokens/s), and two image questions passed 2/2. A short MTPLX tuning set measured D2 at 108 tokens/s versus AR at 51 tokens/s. These are local prompt-set measurements, not a general quality score. The full BF16 parent's runtime numerical parity was not measurable on this host; MTPLX marks this artifact unverified pending that baseline. Review generated code before use.