NVFP4 (4-bit e2m1, per-16 FP8-e4m3 scales) GGUF of ornith-ai/Ornith-1.5-35B-A3B, quantized from the official BF16 GGUF release (token embeddings and output head Q5_K, norms F32). 71.9 GB BF16 → 20.2 GB. First NVFP4 quantization of this model in either format at original publication (2026-08-19, ~34.5 h after the model dropped).

This revision ships a continued-trained MTP head. The vendor's multi-token-prediction head is functional at chain depth 1 and collapses deeper — measured next-token top-1 on held-out own-generation data: depth 1 0.80, depth 2 0.27, depth 3 0.13 (the trunk's RL loop moved; the 1-layer head did not follow). We continued-trained the head on the trunk's own generations with depth-3 chain-rollout (teacher-forced tokens, self-recursive hidden carrier — the exact recurrence speculative decoding runs at serve time). Same held-out measurement after training: 0.81 / 0.58 / 0.43. Serve-level draft acceptance (K=3, greedy, same-window interleaved A/B): 0.431 vs 0.352 for the vendor head; code-generation probes 0.540 vs 0.388. Trunk, embeddings, lm_head and every non-head tensor are unchanged official bytes; a draft head can never change output — the target verifies every drafted token — it only moves acceptance and speculative speed.

Built as a serving artifact for memra, a from-scratch Rust + CUDA inference engine for RTX Blackwell (sm_120a) with per-request exactness gates: speculative, graphed, and batched serving are gated byte-identical to plain decode. A single 96 GB Blackwell card fits these weights with room for the full 262,144-token native context.

NVFP4 is not an upstream llama.cpp tensor type: this file runs on memra and on the NVFP4 branch of avifenesh/llama.cpp. For upstream-llama.cpp/Ollama use, the official Q4_K_M–Q8_0 GGUFs are the right pick.

mtp-Ornith-1.5-35B-A3B-NVFP4-frspec-owngen32768.gguf (0.9 GB, sha256 46f0dd4c…73fe899) — standalone speculative-decode draft from the trained MTP block: lm head masked to the top 32,768 of 248,320 rows, ranked by the model's own generations (external text used as prompts only), then requantized in the mask-first order — NVFP4 head, Q4_K_M block. Ranks ship as ornith15-ranks-owngen-32768.txt (drives memra's load-time head mask on safetensors trunks via MEMRA_FRSPEC_TRIM) and as .gguf. A mask can never change output; it moves draft acceptance (measured 0.393 vs 0.431 embedded at K=3) against a smaller head read.

head-Ornith-1.5-35B-A3B-frspec-owntrim-q5k-32768.gguf (46 MB, sha256 d7c47026e232c0d19240e6ecb3ec9702751fc68dec82909bec79b909846b3b4d) — the top-32,768 own-generation-ranked rows of this artifact's own Q5_K output.weight, byte-gathered (zero requant) plus the d2t map. memra's MEMRA_FRSPEC_TRIM consumes it (or the ranks file directly) to trim the DRAFT lm_head from 248,320 to 32,768 rows (~221 → ~29 µs per draft step). Verify stays full-vocab, so output is unchanged by construction; measured serve effect on one RTX PRO 6000 Blackwell (ABBA, N=6/shape, 2026-08-22): single-stream short prompts +10%, shared-prefix c8 +2.9%.

The early spec-off posture is superseded. With the continued-trained head + the self-trim, memra serves Ornith-1.5 spec-on by default, and since v0.105.0 the automatic depth table serves cached long prompts at K=5 when a rank-trimmed head is loaded — the strongest measured cell (single-stream cached-long 326–354 tok/s; ~273 at the old K=2, same card). Other measured serve cells on one RTX PRO 6000 Blackwell WS (vendor sampling T=0.6/top-p 0.95/top-k 20): 8-turn agentic session 11.6–12.9 s total; shared-prefix c8 ~915 tok/s aggregate; c16 ~700 tok/s aggregate; warm-turn TTFT 25–40 ms.

Pair the card with a current-generation host CPU. This model's architecture forces the sampled speculative path, and that path is host-issue-bound rather than GPU-bound: with MEMRA_SPEC_PHASE=1 the round splits verify-issue 50–58%, host-commit 24–26%, draft 19–25% and verify-wait 0.0% — the GPU is never what the host waits for. Measured consequence, same GPU / same binary / same artifact: a Zen 3 host (~3.4 GHz) serves 195–230 tok/s while a Zen 5 host serves 326–354. A cold host also reads low until its clocks ramp. Receipts: memra research/orndecode-20260822/.

Serve with memra:

The standalone masked draft (mtp-…-frspec-owngen32768.gguf) remains published for external-attach engines; on memra the self-trim of the embedded head is the faster route (the external attach forgoes the resident graph-draft path). Built with tools/make-trimmed-draft.sh + frspec-owngen from the memra repo.