Model Follower.

Up to date

Rebuilt
2026-10-25 06:54 UTC
Release radar
117 on watch
Hub snapshot
2026-10-25
Re-checked
daily · every number sourced and dated

Models › Ornith-1.5 family

Ornith-1.5-35B-A3B

Hugging Face ↗ Follow changes

ornith-ai MIT out 2 months ago

36.0B parameters, 3B active (8 of 256 experts per token), text in, text out, 256K tokens of context.

Verified 2026-10-25 Rank #651 in LLMs · score 73.1 method v1 Sources 3 Downloads 30d 263.8k · #227 of 619 tracked

Changed this week: 11 new quantized builds on the Hub (202 in all).2026-10-24 · 10 new quantized builds on the Hub (201 in all).2026-10-24

Can I run it?22GB at INT4Estimatefits a 24 GB cardmethod
Can I use it?MITPermissivelicence file2026-10-24
Do people use it?263.8kdownloads, 30 days#227 of 619 we track · top 37%+15 likes, downloads +5% this week750 likesHugging Face2026-10-24

Worth reading

Picked by rule from 1 tracked piece: the lab's own post, the most shared, a field report, the latest.

Latest · 2 months ago

Ornith-1.5: From Self-Scaffolding to Self-Improvement

Ornith-1.5 extends its self-scaffolding framework into a complete self-improvement loop that jointly optimizes task generation, scaffold construction, and solution rollouts.

  • The 9B model can be deployed on iPhone and Android devices while outperforming larger models like Gemma 4-31B.
  • The system proposes progressively harder tasks that go beyond what the model has already solved to expose capability gaps.
ornith.ai 2026-08-19 216 HN pts · 76 comments 2.3k words about the Ornith 1.5 family family

Where it's used

Hugging Face does not say who downloads a model or with what. These are the proxies: API providers listing it, quantized builds, fine-tunes/adapters/merges, and Spaces that load it.

No API provider serves it, so its downloads are people running it themselves: 202 quantized builds and 1 Spaces.

  • As a service0no API provider listed
  • On own hardware202quantized builds
  • Built on4335 fine-tunes · 0 adapters · 8 merges
  • In apps1Hugging Face Spaces

Will it fit?

Weights only, an estimate: leave room for the KV cache. Method

Fits a 64 GB Mac (unified memory) at INT4

About 22 GB at INT4, 43 GB at FP8, 86 GB at BF16: weights only, so leave room for the KV cache. Every expert must be resident, so total parameters count.

16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 22 GB of weights (estimate; KV cache and activations extra)INT422 GBFP8: about 43 GB of weights (estimate; KV cache and activations extra)FP843 GBBF16: about 86 GB of weights (estimate; KV cache and activations extra)BF1686 GB
16 GB card64 GB Mac2×80 GB8×80 GBINT4: about 22 GB of weights (estimate; KV cache and activations extra)INT422 GBFP8: about 43 GB of weights (estimate; KV cache and activations extra)FP843 GBBF16: about 86 GB of weights (estimate; KV cache and activations extra)BF1686 GB

Estimate params × bytes per param × 1.2 · how we estimate

The story so far

Weights, every dated article, and repo commits on one rail.

1 dated article and 13 repo commits so far

Teal is the weights (2026-08-18), amber triangles are 1 dated article (each links to it), grey ticks are repo commits.

SepOct2026-08-23: 1 repo commit2026-08-20: 1 repo commit2026-08-19: 9 repo commits2026-08-18: 2 repo commitsOrnith-1.5: From Self-Scaffolding to Self-Improvement · 2026-08-19Weights 2026-08-18 · 2026-08-18Weights 2026-08-18now
SepOct2026-08-23: 1 repo commit2026-08-20: 1 repo commit2026-08-19: 9 repo commits2026-08-18: 2 repo commitsOrnith-1.5: From Self-Scaffolding to Self-Improvement · 2026-08-19Weights 2026-08-18 · 2026-08-18Weights 2026-08-18now

What people built on it

245 derivatives on the Hub. Which build to run, by runtime: the most downloaded in each class.

245 derivatives on the Hub, most of them quantized

From the Hugging Face model tree; quantized builds in teal, fine-tunes, merges and adapters in greys.

202 quantized (Hugging Face model tree)35 fine-tunes (Hugging Face model tree)8 merges (Hugging Face model tree)202 quantized35 fine-tunes8 merges
202 quantized (Hugging Face model tree)35 fine-tunes (Hugging Face model tree)8 merges (Hugging Face model tree)202 quantized35 fine-tunes8 merges
GGUF peculiar-ragdoll/Tiel-Coder-35B-A3B-GGUF-MTP for llama.cpp, Ollama, LM Studio 3.4M downloads MLX PocketAiHub/Ornith-1.5-35B-A3B-Abliterated-MLX-4bit for Apple silicon 10.1k downloads AWQ cyankiwi/Ornith-1.5-35B-A3B-AWQ-INT4 for vLLM, SGLang 29.5k downloads

All variants →

Reference

Everything else, folded: the full tables, every source.

What the card says

1 cited sentence

Summary generated, cited

The model contains 36.0B total parameters and activates approximately 3B parameters per token.[1][2]

  1. parameters human
  2. “activates only ~3B parameters per token”

Generated 2026-10-07 15:13 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

6 use cases

What it is good at 6 use cases

Labels are evidence, not verdicts: measured a third-party leaderboard row for this model; author-reported a number the model author published in the card or eval files; structural a deterministic signal from the name, config or chat template, or a decision-model estimate. Each chip links to its source. All use cases.

Facts

23 fields

Listed in LLMs · MoE

Facts

Pipeline tagtext-generation
Librarytransformers
ArchitecturesQwen3_5MoeForConditionalGeneration
Model typeqwen3_5_moe
Dtypebfloat16
Params by dtypeBF16: 35,951,822,704
Size bandXL
MoEyes
Experts256
Experts per token8
Context length262,144 (config.json max_position_embeddings)
LicenseMIT (mit)
Hub snapshot2026-10-25

Servability

vLLM · SGLang

Servability

vLLM: supported · Qwen3_5MoeForConditionalGeneration · 7050ef881ff0835f6890a47db1859a24c65481d6SGLang: supported · Qwen3_5MoeForConditionalGeneration · 773cb938fe9484a8f7e7e46fb1b1fe96412c59ed

vLLM registry (7050ef881ff0835f6890a47db1859a24c65481d6) · SGLang registry (773cb938fe9484a8f7e7e46fb1b1fe96412c59ed) · Matched by architecture class name against each project's registry on main; 'supported' means the architecture is registered, not that this checkpoint was tested.

Hardware fit

3 estimates

Hardware fit ESTIMATE

Weights-only memory for Ornith-1.5-35B-A3B: params × bytes/param × 1.2 overhead. KV cache and activations are extra. How this is computed.

BF16 (2 bytes/param)86 GB · Multi-GPU node (<640 GB)
FP8 (1 byte/param)43 GB · Single high-end GPU (<80 GB)
INT4 (0.5 bytes/param)22 GB · Single GPU (<24 GB)

Weights only (params × bytes/param × 1.2 overhead); KV cache and activations extra. MoE: all experts must be resident, so total parameters count.

All variants

40 variants

Lineage

Original model Family: ornith-ai/Ornith-1.5 245 derivatives on the Hub

quantized: 202finetune: 35merge: 8

Top variants (40 shown, ranked by downloads)

Relation
Quant
Size
40 rows
Variant Relation Quant Params Band License Author Likes Downloads
peculiar-ragdoll/Tiel-Coder-35B-A3B-GGUF-MTP quantized GGUF — — mit peculiar-ragdoll 306 3.4M
peculiar-ragdoll/Tiel-Coder-35B-A3B-GGUF quantized GGUF — — mit peculiar-ragdoll 398 814.3k
mudler/Ornith-1.5-35B-A3B-APEX-MTP-GGUF quantized GGUF — — apache-2.0 mudler 103 762.1k
mudler/Ornith-1.5-35B-A3B-APEX-GGUF quantized GGUF — — apache-2.0 mudler 38 490k
ornith-ai/Ornith-1.5-35B-A3B-DFlash finetune — 386M S mit ornith-ai 34 458.4k
bartowski/Ornith-1.5-35B-A3B-GGUF quantized GGUF — — mit bartowski 49 178.9k
AtomicChat/Ornith-1.5-35B-A3B-GGUF quantized GGUF — — mit AtomicChat 63 122.3k
gbuzhf/Ornith-1.5-35B-A3B-Abliterated-MTP-UD-APEX-GGUF quantized GGUF — — mit gbuzhf 40 113.1k
dealignai/Ornith-1.5-35B-A3B-UNCENSORED-GGUF quantized GGUF — — mit dealignai 41 97.2k
PocketAiHub/Ornith-1.5-35B-A3B-Abliterated-GGUF quantized GGUF — — mit PocketAiHub 41 71.4k
SC117/Ornith-1.5-35B-A3B-Heretic-MTP-APEX-GGUF quantized GGUF — — mit SC117 48 52.8k
chenyumo/moziAI-35B-A3B-MOE-MTP finetune GGUF — — apache-2.0 chenyumo 23 28.7k
peculiar-ragdoll/Unsloth-Ornith-1.5-35B-A3B quantized GGUF — — mit peculiar-ragdoll 24 28.5k
cyankiwi/Ornith-1.5-35B-A3B-AWQ-INT4 quantized AWQ 36.0B XL mit cyankiwi 4 29.5k
gbuzhf/Ornith-1.5-35B-A3B-Huihui-Sangreal-MTP-ICE-GGUF quantized GGUF — — mit gbuzhf 6 22.3k
SC117/Ornith-1.5-35B-A3B-MTP-APEX-GGUF quantized GGUF — — mit SC117 12 21.8k
jashepp/Ornith-1.5-35B-A3B-MXFP4_MOE_Hybrid-Imatrix-GGUF quantized Imatrix — — mit jashepp 9 18.5k
LuffyTheFox/Tiel-Coder-35B-A3B-Genesis-Hermes-GGUF quantized GGUF — — mit LuffyTheFox 29 17k
gbuzhf/Ornith-1.5-35B-MTP-UD-APEX-GGUF quantized GGUF — — mit gbuzhf 5 15.1k
pfeifferj/Ornith-1.5-35B-A3B-GSQ-RCO-GGUF quantized GGUF — — mit pfeifferj 25 12.6k
pragmaticcs/Qwen-35B-A3B-SignOfFour-Coder-GGUF merge GGUF — — apache-2.0 pragmaticcs 8 12.7k
cyankiwi/Ornith-1.5-35B-A3B-AWQ-FP8 quantized AWQ 36.0B XL mit cyankiwi 1 11.4k
IsValorum/Ornith-1.5-35B-A3B-APEX-I-MiniPlus-V2.1-GGUF quantized GGUF — — mit IsValorum 11 10.7k
tiyuvta/Ornith-1.5-35B-A3B-NVFP4-MTP-GGUF quantized NVFP4 — — mit tiyuvta 6 10.6k
LuffyTheFox/Ornith1.5-35B-A3B-Genesis-Hermes-GGUF quantized GGUF — — mit LuffyTheFox 27 9.5k
biMEMO/Ornith-1.5-35B-A3B-int4-AutoRound-MTP quantized int4 5.98B M mit biMEMO 6 10.5k
PocketAiHub/Ornith-1.5-35B-A3B-Abliterated-MLX-4bit quantized MLX 35.1B XL mit PocketAiHub 8 10.1k
IsValorum/Tiel-Coder-35B-A3B-APEX-I-MiniPlus-V2.1-GGUF quantized GGUF — — mit IsValorum 23 8.9k
ulkaa/Ornith-1.5-35B-A3B-AWQ-INT4 quantized AWQ 8.05B M mit ulkaa 2 9.9k
offmonreal/Ornith-1.5-35B-MaxQuality-MTP-GGUF quantized GGUF — — mit offmonreal 8 8.6k
junafinity/Ornith-1.5-35B-A3B-uncensored-GGUF-8bit quantized GGUF — — apache-2.0 junafinity 4 8.2k
mradermacher/Ornith-1.5-35B-A3B-i1-GGUF quantized GGUF — — mit mradermacher 3 7.9k
gbuzhf/Ornith-1.5-35B-A3B-TIEL-Calibrated-MTPv2-ICE-GGUF quantized GGUF — — mit gbuzhf 9 7.6k
ArtomYuan/Ornith-1.5-35B-A3B-ROCmFPX quantized GGUF — — apache-2.0 ArtomYuan 0 7.5k
huihui-ai/Huihui-Ornith-1.5-35B-A3B-abliterated finetune — 36.0B XL mit huihui-ai 29 6k
mlx-community/Ornith-1.5-35B-A3B-OptiQ-4bit quantized -4bit 34.7B L apache-2.0 mlx-community 4 7k
rdtand/Ornith-1.5-35B-A3B-PrismaQuant-AQUA-25GB-vllm quantized — 20.2B L mit rdtand 1 7k
ressl/Ornith-1.5-35B-A3B-uncensored-GGUF finetune GGUF — — mit ressl 3 6.7k
peculiar-ragdoll/Tiel-Coder-35B-A3B-MLX-oQ4e-MTP quantized MLX 36.0B XL mit peculiar-ragdoll 21 5.3k
IsValorum/Tiel-Coder-35B-A3B-APEX-I-NanoPlus-GGUF quantized GGUF — — mit IsValorum 10 5.7k

Repository changelog

13 commits
changed in the last 7 daysOther: 2026-10-24, 2026-10-24all changes

Repository changelog 13 commits

DateCommitSHA
2026-08-23 Upload model-00016-of-00016.safetensors with huggingface_hub 10fbf86
2026-08-20 Update README.md e4dfb35
2026-08-19 Update README.md fbb995a
2026-08-19 Upload ornith_35b_eval.png e0e1301
2026-08-19 Delete assets/ornith_35b_eval.png c53249d
2026-08-19 Update README.md 6e89a57
2026-08-19 Update README.md 9d04d53
2026-08-19 Upload ornith_35b_eval.png 68a1bbc
2026-08-19 Delete assets/ornith_9b_eval.png 6f751c2
2026-08-19 Upload 2 files c56d43f
2026-08-19 Upload folder using huggingface_hub 97d67d3
2026-08-18 Upload folder using huggingface_hub a111f30
2026-08-18 initial commit 7e15393

Articles and coverage

1 piece

Ranked by source signals (HN, Lobsters, authority domains, official posts) and recency. Summaries are generated from a stored copy of the article text; every sentence cites a span of that text, and nothing is shown that failed the citation check.

Official1

Ornith-1.5: From Self-Scaffolding to Self-Improvement
ornith.ai2026-08-19officialHN front pageabout the Ornith 1.5 familyHN 216 pts · 76 comments2,250 words
Ornith-1.5 extends its self-scaffolding framework into a complete self-improvement loop that jointly optimizes task generation, scaffold construction, and solution rollouts. The system uses GRPO to optimize the loop based on validity, frontier difficulty, and novelty rewards, with a target frontier difficulty of 0.2. The 397B MoE variant scores 86.1 on Terminal-Bench 2.1 and 56.0 on DeepSWE, matching Claude Opus 4.8 while outperforming peers like GLM-5.2.written, reviewed, cited
  • The 9B model can be deployed on iPhone and Android devices while outperforming larger models like Gemma 4-31B.
  • The system proposes progressively harder tasks that go beyond what the model has already solved to expose capability gaps.
  • Validity is treated as a hard gate to prevent malformed tasks from receiving reward simply because they appear difficult.

Similar models

3 models

Similar models

  • GLM-4.7-Flash · 31.2B
    also MoE, same task (text-generation), more providers, permissive license
  • Qwen3-Coder-30B-A3B-Instruct · 30.5B
    also MoE, same task (text-generation), more providers, permissive license
  • Qwen3-30B-A3B · 30.5B
    also MoE, same task (text-generation), more providers, permissive license

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:41 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:13 UTC · Last verified 2026-10-25

Follow Ornith-1.5-35B-A3B

Watch this model

Email alerts are not open yet; the feeds above are the same data. See feeds for the item schema.