DEV.co
Open-Source LLM · protoLabsAI

Ornith-1.0-9B-MTP-GGUF

Ornith-1.0-9B-MTP-GGUF is a 9-billion-parameter language model quantized in GGUF format with a built-in multi-token prediction (MTP) head for speculative decoding acceleration. It runs on consumer and data-center GPUs via llama.cpp with 1.38–1.73× throughput gains depending on quantization and hardware. The model itself is a Qwen3.5-9B hybrid fine-tune, available in 10 quantization variants ranging from 3.9 GB (very low VRAM) to 18.4 GB (reference precision). MIT license, no access gate.

Source: HuggingFace — huggingface.co/protoLabsAI/Ornith-1.0-9B-MTP-GGUF
Unknown
Parameters
mit
License (OSI-approved)
Unknown
Context (tokens)
55.1k
Downloads (30d)

Key facts

Objective fields from the source. Values we can't verify are shown as “Unknown” rather than guessed.

FieldValue
DeveloperprotoLabsAI
ParametersUnknown
Context windowUnknown
Licensemit — OSI-approved
Modality / tasktext-generation
Gated on HuggingFaceNo
Downloads55.1k
Likes60
Last updated2026-07-04
SourceprotoLabsAI/Ornith-1.0-9B-MTP-GGUF

What Ornith-1.0-9B-MTP-GGUF is

Ornith-1.0-9B is a Qwen3.5-9B base with linear + full attention layers, fine-tuned by deepreinforce-ai. The MTP variant adds a KL-distilled draft head (15 tensors, 1 layer) grafted into the trunk before quantization. The head enables llama.cpp's lossless multi-token self-speculative decoding (no separate draft model required) via `--spec-type draft-mtp`. GGUF quantization profiles span importance-matrix (IQ) and k-quantized (K) rungs. On Blackwell (RTX 50xx), NVFP4 format (6.6 GB) achieves ~306 tok/s with MTP vs. ~201 tok/s base; on Ampere, Q4_K_M (5.8 GB) hits ~153 tok/s MTP vs. ~105 tok/s base. Acceptance rates hold ~0.65–0.84 across quants at n-max 3. Requires llama.cpp ≥ b9616 (Qwen3.5 arch support, NVFP4 type 40 for Blackwell variant). Last modified 2026-07-04.

Quickstart

Run Ornith-1.0-9B-MTP-GGUF locally

Load the open weights with 🤗 Transformers and generate — the same model, self-hosted.

quickstart.pypython
from transformers import pipelinepipe = pipeline("text-generation", model="protoLabsAI/Ornith-1.0-9B-MTP-GGUF")out = pipe("Explain retrieval-augmented generation in one sentence.",           max_new_tokens=128)print(out[0]["generated_text"])

Swap in vLLM or Ollama for production-grade serving. DEV.co can stand up the inference stack.

Deployment

How you'd run it

A typical self-hosted path — open weights, an inference server, your application.

DEV.co builds each layer — from GPU infrastructure to the application.

Best use cases

Local/Self-Hosted Code Generation

9B model with MTP running on single Ampere/Blackwell GPU (5.8–6.6 GB VRAM). Benchmarks show ~145 tok/s on Q4_K_M + MTP (RTX A6000). Standalone, no external API; suitable for IDE plugins, CLI tools, or air-gapped environments.

High-Throughput Inference on Constrained VRAM

IQ2_M variant (3.9 GB, ~0.81 acceptance) serves from 5 GB total VRAM with MTP active. Trade-off: lower precision for lower footprint and acceptable speculative gains. Practical for edge servers, embedded deployments, or multi-model hosting on single node.

Throughput Optimization for Blackwell Infrastructure

NVFP4 rung (6.6 GB, 306 tok/s with MTP) compounds tensor-core GEMM efficiency with MTP's parallel verify. Blackwell-specific tuning; 50% faster than Q4_K_M baseline on same hardware. Ideal for latency-sensitive batch inference or real-time API serving.

Running & fine-tuning it

**Minimum VRAM:** 3.9 GB (IQ2_M); **balanced:** 5.8 GB (Q4_K_M, Ampere) or 6.6 GB (NVFP4, Blackwell); **reference:** 9.8 GB (Q8_0); **full precision:** 18.4 GB (BF16). **GPU:** NVIDIA, CUDA 11.0+. **CPU:** Intel/AMD x86_64; MTP speculative decoding requires GPU acceleration (CPU inference untested/unsupported). **Host RAM:** ~4 GB additional for llama.cpp runtime + context buffer. **Precision note:** IQ variants trade ~1–3% accuracy for 15–20% VRAM savings; k-quants (Q4–Q8) are lossless-adjacent.

Card does not document LoRA, QLoRA, or full fine-tuning feasibility. Base model (deepreinforce-ai/Ornith-1.0-9B) is a Qwen3.5 hybrid; compatibility with peft/bitsandbytes unknown. The MTP head is KL-distilled post hoc and frozen into GGUF (no separate training documented). For continued adaptation: (1) fine-tune the base Qwen3.5-9B model separately, then re-graft/re-quantize; (2) or use non-GGUF format (HF checkpoint) and handle MTP head manually. Requires review with protoLabsAI or upstream Qwen3.5 documentation.

When to avoid it — and what to weigh

  • Bitwise Reproducibility Required — MTP's batched verification computes logits in different floating-point reduction order than sequential decode. Output distribution is identical, but greedy/temp-0 sequences may diverge. Card states this is expected, not a defect—but if your application requires deterministic token-by-token replay, use base model without speculative decoding.
  • Context Length >8192 or Real-Time Streaming at Sub-100ms Latency — Card does not publish tested context lengths beyond 8192. MTP adds per-step parallel verify cost; while net throughput improves, first-token latency is not optimized. If generation must start within 50–100ms cold, profile before committing.
  • CUDA Compute Capability <6.1 or Non-GPU Deployment — GGUF quantization and MTP speculative decoding are GPU-accelerated paths. CPU inference possible but slow (no benchmarks provided). Older GPUs (pre-Ampere) lack tensor cores; NVFP4 requires Blackwell. Verify GPU compatibility before deployment.
  • Production Quality Assurance Without Acceptance Rate Monitoring — Card reports acceptance at ~0.65–0.84 depending on quant/prompt. Lower acceptance (creative prose: 287 tok/s vs. code: 330 tok/s) means MTP gains vary per workload. Requires profiling on representative data to validate SLA.

License & commercial use

MIT license. Base (Ornith-1.0-9B by deepreinforce-ai) is MIT. MTP head (protoLabsAI/Ornith-1.0-9B-MTP) is MIT. Derived GGUF quantizations are MIT. License is OSI-approved permissive.

MIT license permits commercial use, modification, and distribution. No restrictions stated. However, verify: (1) base Qwen3.5-9B upstream license (Alibaba; typically permissive but requires confirmation); (2) whether your intended commercial use (e.g., embedded model in a paid SaaS) requires trademark/attribution compliance under MIT terms (attribution required, not warranty). Recommend legal review of downstream Qwen license and MIT attribution chain before large-scale production deployment.

DEV.co evaluation signals

Editorial assessment — not user reviews. Directional, with an explicit confidence level.

SignalAssessment
MaintenanceActive
DocumentationStrong
License clarityClear
Deployment complexityLow
DEV.co fitStrong
Assessment confidenceHigh
Security considerations

Standard LLM model considerations: (1) GGUF format is a binary serialization; verify source integrity (HuggingFace CDN + git commit hash). (2) No adversarial robustness claims stated; model inherits Qwen3.5-9B base properties (unknown). (3) Speculative decoding (MTP) does not introduce new attack surface (verification is deterministic). (4) No data privacy/training data documentation provided. (5) Quantization to GGUF may compress adversarial gradients; security implications unknown. For production: audit base Qwen3.5-9B security posture and consider input sanitization/rate-limiting on public-facing inference endpoints.

Alternatives to consider

Llama-2-13B GGUF (Meta/Hugging Face)

Larger, well-established baseline. No built-in MTP; requires separate draft model for speculative decoding. Simpler setup but slower throughput (~70–100 tok/s on comparable hardware without MTP).

Mistral-7B GGUF (Mistral AI)

Smaller footprint (3.3–6 GB), similar latency profile. No MTP head; wider community support and integration (Ollama, LM Studio native). Trade-off: lower absolute throughput, but better VRAM scaling.

DeepSeek-Coder-6.7B-GGUF (DeepSeek)

Comparable size, code-optimized training. No speculative decoding. If code-generation quality > throughput, may offer better output despite slower inference.

Software development agency

Ship Ornith-1.0-9B-MTP-GGUF with senior software developers

Download Ornith-1.0-9B-MTP-GGUF and benchmark on your hardware. Start with Q4_K_M (Ampere) or NVFP4 (Blackwell). See build recipes and full benchmark data in the model card. Questions? Open a community discussion on HuggingFace—protoLabs responds within 48h.

Talk to DEV.co

Related open-source tools

Surfaced by semantic similarity across the DEV.co open-source index.

Related on DEV.co

Explore the category and the services that help you build with it.

Ornith-1.0-9B-MTP-GGUF FAQ

Can I use Ornith-1.0-9B-MTP commercially in a SaaS product?
MIT license permits commercial use. However, verify upstream Qwen3.5-9B license (Alibaba) and ensure MIT attribution is included in your product documentation and legal terms. Recommend legal review before launch to confirm compliance with MIT chain of custody.
Which GGUF file should I download for an RTX 4090?
RTX 4090 is Ampere-generation. Use Q4_K_M (5.8 GB, fastest on Ampere). If you need higher quality and have >7.5 GB VRAM, Q5_K_M (6.6 GB) or Q6_K (7.6 GB) are also suitable. Avoid NVFP4; it is slower on Ampere.
What is the difference between 'bundled' and 'standalone head' GGUF files?
Bundled: MTP head is baked into the trunk; load and run directly with `--spec-type draft-mtp`. Standalone head (mtp-head/…): attach to any base Ornith-9B GGUF via `--model-draft`; useful if you have a custom base quantization. Functionally identical; bundled is simpler.
Does MTP speculative decoding guarantee the same output as greedy decode without MTP?
Distribution is identical; output tokens are verified, so no skew. However, not bitwise identical: batched verify computes logits in different floating-point order than sequential decode, which can flip an argmax and fork the sequence. Both are valid; this is expected llama.cpp behavior.

Software developers & web developers for hire

DEV.co is a software development agency delivering custom software development services to companies building on open source. Our software developers and web developers design, integrate, and ship production systems — spanning web development, APIs, AI, data, and cloud. If Ornith-1.0-9B-MTP-GGUF is part of your open-source llms roadmap, our team can implement, customize, migrate, and maintain it.

Ready to Deploy High-Speed Local LLM Inference?

Download Ornith-1.0-9B-MTP-GGUF and benchmark on your hardware. Start with Q4_K_M (Ampere) or NVFP4 (Blackwell). See build recipes and full benchmark data in the model card. Questions? Open a community discussion on HuggingFace—protoLabs responds within 48h.