Ornith
@ornith_
Once you have tasted flight, you will walk the earth with your eyes turned skywards.
308 Following 18.9K Followers
🔍We dove a bit deeper and analyzed how multi-token prediction (MTP) works in Ornith 1.5 models.
🪽With MTP, Ornith 1.5 gain inference speedups without quality loss via self-speculative decoding: models use their own MTP head to draft tokens and verify them in a forward pass.
🐦MTP weights are now updated for all 9B, 35B, and 397B models in BF16, GGUF, FP8, and NVFP4 variants (except 397B NVFP4, will be updated today).
Show more
🐦Tokens now go brrrr
🪽Ornith-1.5 models just got new wings: MTP weights have been updated for the BF16, GGUF, NVFP4 & FP8 variants.
🔗
Aloha! 🌺Introducing Ornith-1.5, a family of open-source LLMs spanning 9B Dense, 35B MoE, and 397B MoE, trained with self-improving strategies.
It achieves state-of-the-art performance among open-source models of comparable size and delivers performance comparable to Claude Opus 4.8 across reasoning, agentic, and coding tasks:
✅Terminal-Bench 2.1 (86.1)
✅SWE-Bench (86 on verified, 65.1 on pro, 79.6 on Multilingual)
✅DeepSWE (56)
✅HLE (44.6)
✅ClawEval (81.4)
✅Tool Decathlon (71.2)
Ornith-1.5 takes a major step toward training foundation models through end-to-end self-improvement, extending the self-scaffolding strategies introduced in Ornith-1.0 into a more complete self-improvement loop: the model proposes new tasks, generates task-specific scaffolds, and produces solution rollouts for reinforcement learning, continuously creating new learning experiences from which it can improve.
All models, along with their quantized versions (FP8, GGUF, MLX, and NVFP4), have been released under the MIT License, enabling unrestricted commercial and research use.
📘Tech Blog:
🤗Huggingface:
Show more
🐦Now we can just burn some tokens🔥
🫡Thanks @runinfrai for the support!
216 tok/s. 167ms to first token. 262,144 token context
Ornith 1.5 35B is now live on RunInfra
$0.10/1M input. $0.40/1M output. $0.01/1M cached. 90% of input hits the cache, so effective input is ~$0.02/1M
1M in + 1M out with a warm cache costs $0.42
Show more
🧡we love feedbacks!
Over the last day I've run an extensive set of benchmarks on Qwen3.6-35B-A3B vs Ornith-1.5-35B-A3B.
The results are in. I will post a more detailed thread of the breakdown shortly, but the clear winner is Ornith. Especially as it comes to coding, math and science.
I chose to test both MoE models against each other to compare apples to apples.
Show more
🐦it's time to download and play with Ornith🤗
🐦Ornith 1.5 series models are now available on ollama!
🫡Thanks to our awesome friends at @ollama
👀Come give Ornith 1.5 a try!
🐦Ornith-1.5 is built upon Ornith-1.0, which was developed on top of Qwen3.5 @Alibaba_Qwen with additional continued pretraining, mid-training, and post-training.
🫡Thanks to the open-weight community!
👀We are happy and welcome more projects/model variants built on Ornith.
Show more
Aloha! 🌺Introducing Ornith-1.5, a family of open-source LLMs spanning 9B Dense, 35B MoE, and 397B MoE, trained with self-improving strategies.
It achieves state-of-the-art performance among open-source models of comparable size and delivers performance comparable to Claude Opus 4.8 across reasoning, agentic, and coding tasks:
✅Terminal-Bench 2.1 (86.1)
✅SWE-Bench (86 on verified, 65.1 on pro, 79.6 on Multilingual)
✅DeepSWE (56)
✅HLE (44.6)
✅ClawEval (81.4)
✅Tool Decathlon (71.2)
Ornith-1.5 takes a major step toward training foundation models through end-to-end self-improvement, extending the self-scaffolding strategies introduced in Ornith-1.0 into a more complete self-improvement loop: the model proposes new tasks, generates task-specific scaffolds, and produces solution rollouts for reinforcement learning, continuously creating new learning experiences from which it can improve.
All models, along with their quantized versions (FP8, GGUF, MLX, and NVFP4), have been released under the MIT License, enabling unrestricted commercial and research use.
📘Tech Blog:
🤗Huggingface:
Show more
🐦Now you can serve Ornith 1.5 with vLLM!
vllm serve ornith-ai/Ornith-1.5-9B
🫡Thanks for the support @vllm_project !
Ornith-1.5 is out and it serves in vLLM today. Congrats @ornith_ on a self-improving, MIT-licensed family from 9B to 397B, SOTA among open models on coding and agentic tasks.
Already serving in vLLM:
vllm serve ornith-ai/Ornith-1.5-9B
Show more
Ornith-1.5 is out and it serves in vLLM today. Congrats @ornith_ on a self-improving, MIT-licensed family from 9B to 397B, SOTA among open models on coding and agentic tasks.
Already serving in vLLM:
vllm serve ornith-ai/Ornith-1.5-9B
Show more
Aloha! 🌺Introducing Ornith-1.5, a family of open-source LLMs spanning 9B Dense, 35B MoE, and 397B MoE, trained with self-improving strategies.
It achieves state-of-the-art performance among open-source models of comparable size and delivers performance comparable to Claude Opus 4.8 across reasoning, agentic, and coding tasks:
✅Terminal-Bench 2.1 (86.1)
✅SWE-Bench (86 on verified, 65.1 on pro, 79.6 on Multilingual)
✅DeepSWE (56)
✅HLE (44.6)
✅ClawEval (81.4)
✅Tool Decathlon (71.2)
Ornith-1.5 takes a major step toward training foundation models through end-to-end self-improvement, extending the self-scaffolding strategies introduced in Ornith-1.0 into a more complete self-improvement loop: the model proposes new tasks, generates task-specific scaffolds, and produces solution rollouts for reinforcement learning, continuously creating new learning experiences from which it can improve.
All models, along with their quantized versions (FP8, GGUF, MLX, and NVFP4), have been released under the MIT License, enabling unrestricted commercial and research use.
📘Tech Blog:
🤗Huggingface:
Show more
🐦Ornith 1.5 9B is working!
This is a 9B model running a full agentic loop with Wi-Fi off.
Ornith 1.5. It read three sheets, cross-referenced the monthly rollups, and wrote a revenue analysis on its own.
Nothing left the Mac.
Show more
This is a 9B model running a full agentic loop with Wi-Fi off.
Ornith 1.5. It read three sheets, cross-referenced the monthly rollups, and wrote a revenue analysis on its own.
Nothing left the Mac.
Show more
Run Ornith 1.5 9B and 35B A3B locally via Atomic Chat 🐦🔥
We shipped the full 9B GGUF ladder on Hugging Face, from lossless BF16 (17.9 GB) down to 2-bit (2.8 GB) and measured all against stock quants
AD-Q4_K runs on a 16GB MacBook Air with 64k context and picks the same next token as the BF16 original 91.9% of the time
Show more
🫡Huge shoutout to @atomic_chat_hq for their expertise!
🐦Come and try Ornith 1.5 on
Run Ornith 1.5 9B and 35B A3B locally via Atomic Chat 🐦🔥
We shipped the full 9B GGUF ladder on Hugging Face, from lossless BF16 (17.9 GB) down to 2-bit (2.8 GB) and measured all against stock quants
AD-Q4_K runs on a 16GB MacBook Air with 64k context and picks the same next token as the BF16 original 91.9% of the time
Show more
Aloha! 🌺Introducing Ornith-1.5, a family of open-source LLMs spanning 9B Dense, 35B MoE, and 397B MoE, trained with self-improving strategies.
It achieves state-of-the-art performance among open-source models of comparable size and delivers performance comparable to Claude Opus 4.8 across reasoning, agentic, and coding tasks:
✅Terminal-Bench 2.1 (86.1)
✅SWE-Bench (86 on verified, 65.1 on pro, 79.6 on Multilingual)
✅DeepSWE (56)
✅HLE (44.6)
✅ClawEval (81.4)
✅Tool Decathlon (71.2)
Ornith-1.5 takes a major step toward training foundation models through end-to-end self-improvement, extending the self-scaffolding strategies introduced in Ornith-1.0 into a more complete self-improvement loop: the model proposes new tasks, generates task-specific scaffolds, and produces solution rollouts for reinforcement learning, continuously creating new learning experiences from which it can improve.
All models, along with their quantized versions (FP8, GGUF, MLX, and NVFP4), have been released under the MIT License, enabling unrestricted commercial and research use.
📘Tech Blog:
🤗Huggingface:
Show more
Aloha! 🌺 Meet Ornith-1.0, a family of open-source LLMs specialized for agentic coding.
Ornith-1.0 spans the full parameter sizes including 9B Dense, 31B Dense, 35B MoE, and 397B MoE. It achieves state-of-the-art performance among open-source models of comparable size on coding benchmarks including:
✅Terminal-Bench 2.1(77.5)
✅SWE-Bench(82.4 on verified, 62.2 on pro, 78.9 on Multilingual)
✅NL2Repo(48.2)
✅SWE Atlas(41.2 on QnA, 42.6 RF, 39.1 TW)
✅ClawEval(77.1)
Post-trained on top of gemma4 and qwen3.5, Ornith-1.0 employs a novel self-improving training strategy in which reinforcement learning is used to generate not only solution rollouts, but also the task-specific scaffolds that drive those rollouts. By jointly optimizing the scaffold and the resulting solution, the model generate higher-quality solutions in agentic coding.😎
All models are released under the MIT license, enabling full commercial and research use.
📖Tech Blog:
🤗Huggingface:
Show more
