Ornith-1.0-35B: 1M Context + MTP + Vision

Mirrors: Hugging Face | ModelScope (full quant ladder always available on ModelScope)

DeepReinforce's Ornith-1.0-35B (35B MoE, 3B active, Qwen3.5 family) with YaRN RoPE scaling baked into the GGUF metadata for a 1,048,576-token context window, needle-certified through the full range, now shipping MTP speculative decoding and a vision tower. Weights are bit-identical to the source builds; only rope metadata differs, so llama.cpp and Ollama apply the extension with no extra flags.

Files: MTP-first

This Hugging Face repo intentionally carries only the MTP builds. The full 10-quant ladder (IQ2_M through bf16, every quant needle-tested) lives on the ModelScope mirror.

Every file, every mirror

Nothing was discontinued: every quant is one click away. Hugging Face carries the curated picks, ModelScope always carries everything, and Ollama serves ready-to-run tags.

On Ollama every tag ships with the vision tower bundled and the 1M rope metadata baked in.

Needle-in-a-haystack certification

Native 262,144, YaRN factor 4 to 1,048,576: 10/10 needles at every rung through 1M, f16 KV, temperature 0 (results.jsonl). The APEX-Compact build was ladder-tested separately: 50/50 from 64K through 524K on the RTX 5090 (results-apex.jsonl); its 786K and 1M rungs are running on a 128 GB Mac at publish time and this card will update when they land.

MTP: same output, more tokens per second

Ornith's base family (Qwen3.5) ships a multi-token-prediction layer in the official checkpoints; Ornith's own GGUF releases dropped it, and community builds by wang-yang (Q6_K) and SC117 (APEX) restored it. The trunk verifies every drafted token, so output is identical to standard decoding, only faster.

Measured here (RTX 5090, 16K ctx): Q6_K 267.9 to 382.9 tok/s (+43 percent). APEX-Compact gains +14 percent; its custom imatrix shifts token probabilities away from the draft head, lowering acceptance, and its strength is instead the 17 GB footprint. Both builds were re-baked to 1M, coherence-gated, and NIAH-spot-checked here before shipping.

llama-server -m ornith-1.0-35b-1M-MTP-Q6_K.gguf \
  -c 1048576 -np 1 --jinja \
  --spec-type draft-mtp --spec-draft-n-max 3 \
  --mmproj mmproj-ornith-35b-f16.gguf

Ollama (1M and vision work; Ollama has no speculative decoding yet, so MTP adds no speed there):

FROM ./ornith-1.0-35b-1M-MTP-Q6_K.gguf
RENDERER qwen3.5
PARSER qwen3.5
PARAMETER num_ctx 262144

How this was built

1. 1M context: YaRN rope-scaling metadata (factor 4.0, original context 262,144) written into the GGUF header with gguf-py. No weight changes, no fine-tuning.
2. MTP: adopted the community GGUFs carrying the official Qwen3.5 MTP layer, verified provenance and license, re-baked 1M metadata, then gated: coherence check, speed A/B, needle spot-checks with speculation active.
3. Certification: multi-needle harness (10 needles, depths 5 to 95 percent, temperature 0, seeded prompts) against llama-server, f16 KV only. The harness ships in this repo (niah_test.py) along with prefill timing data (prefill_timing.jsonl).

For base-model capability benchmarks see the official Ornith-1.0-35B card; weights here are bit-identical apart from rope metadata, so base capability carries over modulo quantization.

How to actually use a 1M-context model

Habits that measurably help, from our RULER, hop and adherence testing across this fleet:

1. Re-state standing instructions near the end of long prompts; recency beats depth.
2. One big reference dump beats a long accumulated conversation. Fresh session per task.
3. After any compaction or summarization, repeat your active rules yourself.
4. Prefill at 500K+ takes real time on any hardware; stage your questions accordingly.
5. Know your quant: the results tables on this card show what each quant actually holds at depth; pick the strongest one your memory allows.

Credits

Ornith-1.0: DeepReinforce (MIT), post-trained on Qwen 3.5. MTP layer: Qwen. MTP GGUF builds: wang-yang and SC117. 1M YaRN extension, vetting, and certification: SatGeze.

Repo history note: the commit history was squashed (July 2026) to reclaim storage from deleted large files, since Hugging Face bills every revision including deleted ones. The real build history of this repo is longer than the log shows. Method, tooling, and full records: aviary-1m.