openhubresearch

ATLAS-OLMo-3-7B-Think-v4

README

License: apache-2.0

Provenance & attribution

Upstream: allenai/Olmo-3-7B-Think · Allen Institute for AI · Apache-2.0 Relationship: mirror / verbatim redistribution. Not a derivative work of the weights.

Verify it yourself — every weight shard in this repo has the same SHA-256 as the corresponding shard in Ai2's repo (checked 2026-08-03 against the Hub's paths-info API):

Table with columns: File, Size (bytes), SHA-256, Identical to Ai2?
File	Size (bytes)	SHA-256	Identical to Ai2?
`model-00001-of-00003.safetensors`	4,969,984,976	`611c266f99a1e0575ba9f595e2f44e103aaa03a0c83b244725d6598aee934a7d`	✅ yes
`model-00002-of-00003.safetensors`	4,981,161,496	`a032a27f1768184ad0d955752e3564c14b7d1c673d1630f0df7429a439da4f41`	✅ yes
`model-00003-of-00003.safetensors`	4,644,917,240	`50ba91f88796c1ed04cacdb9271fa02d213722734dffe05d399eb27bcff7129b`	✅ yes
`config.json`	1,631	(git blob `360f3bb9…`)	✅ yes

bash
# reproduce
huggingface-cli download openhubresearch/ATLAS-OLMo-3-7B-Think-v4 model-00001-of-00003.safetensors
huggingface-cli download allenai/Olmo-3-7B-Think            model-00001-of-00003.safetensors
sha256sum ~/.cache/huggingface/hub/models--*/snapshots/*/model-00001-of-00003.safetensors

Honest small print — where this mirror is not byte-identical:

tokenizer.json is functionally identical but not byte-identical: same 100,278-entry vocabulary, same 100,000 merges, same added tokens — it was simply re-serialized by a different tokenizers version (legacy "a b" merge strings vs. Ai2's ["a","b"] pairs), which is why the file is 3.0 MB here and 7.1 MB upstream. Verified equal after normalisation.
This mirror omits several files Ai2 ships: merges.txt, vocab.json, chat_template.jinja, olmo-think.png. If you need those, use the upstream repo.

Metadata note (2026-08-03). This repo previously declared base_model: allenai/Olmo-3-7B-Think in its card metadata, which caused the Hub to auto-derive the tag base_model:finetune:allenai/Olmo-3-7B-Think and render "Finetuned from allenai/Olmo-3-7B-Think" on this page. That was wrong — nothing here is fine-tuned. The Hub's base_model_relation field currently accepts only adapter, finetune, merge and quantized (Hub docs) — there is no value meaning "verbatim mirror" — so the base_model field has been removed rather than left carrying a claim we cannot support, and attribution is stated explicitly in prose instead. If you are looking for something we actually trained, it is the LoRA adapter at openhubresearch/ATLAS-Taurus-Expert-7B-poc-v1 (a proof-of-concept: +0.021 keyword-overlap on 12 held-out questions — a small, noisy result, honestly reported there).

Licence & credit

These weights are © Allen Institute for AI, released under the Apache License 2.0, and are redistributed here unmodified under that licence. Ai2's model card, licence terms and intended-use guidance for the original are authoritative: https://huggingface.co/allenai/Olmo-3-7B-Think. For errors in Ai2's model card, Ai2 asks that you contact olmo@allenai.org — please do not contact Ai2 about this mirror; open a discussion here instead.

bibtex
@misc{olmo2025olmo3,
  title  = {Olmo 3},
  author = {Team Olmo and others},
  year   = {2025},
  eprint = {2512.13961},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url    = {https://arxiv.org/abs/2512.13961}
}

What ATLAS adds (the serving stack — not the weights)

Production endpoint — what it actually serves today. ATLAS runs an OpenAI-compatible API at https://atlas.thebeastagi.com/v1 (bearer-key auth; GET /v1/models and GET /health open). ⚠️ As of 2026-08-03 that endpoint serves only olmo3-32b — it does not currently serve this 7B model. GET /v1/models returns exactly one entry, olmo3-32b; check it yourself before assuming otherwise. An earlier version of this card said the endpoint served "this model", which was true when the 7B was deployed and became false after the 32B cutover. To run the 7B, build ATLAS and serve these weights locally — see Quick Start with ATLAS below. An OpenRouter provider application is in progress.

Latest Release — v4.2.0 (2026-07-06): Reference Fidelity

v4.2.0 — HF-Reference Fidelity + Reasoning API · 627 tests · CI green

Differential testing against HuggingFace transformers (token-level logit comparison) uncovered and fixed three silent inference-quality bugs that had been degrading output since the initial port:

YaRN correction range — missing 2π factor in the wavelength calculation, incorrect interpolation ramp, and attn_factor must be applied squared
Layer-type RoPE split — YaRN scaling applies only to the 8 full-attention layers; the 24 sliding-window layers use plain θ=500,000 (per configuration_olmo3.py)
QK-norm scope — RMSNorm is computed over the full 4096-dim Q/K projection, not per-head — the per-head variant silently destroyed retrieval beyond ~128 tokens

ATLAS output is now byte-identical to the HF reference implementation. To our knowledge, ATLAS is the only from-scratch (non-transformers, zero-dependency Rust) OLMo-3-7B-Think implementation serving at full reference quality.

Serving upgrades in v4.2.0:

Context window raised to 16,384 tokens served (model supports up to 64K via YaRN), max output 3,584 tokens
Official OLMo chat template with server-side <think> primer and the Ai2 reference system prompt
Reasoning API: chain-of-thought surfaced as message.reasoning and streaming delta.reasoning (OpenRouter convention) — final answers stay clean, reasoning stays inspectable
Sampling defaults aligned with the Ai2 reference (repetition_penalty=1.0 — penalties make thinking models ramble)
compute-sanitizer memcheck clean; exact CPU/GPU parity verified on long sequences

Benchmarks — these measure the ATLAS serving stack running Ai2's unmodified weights

These numbers are not a claim about model quality: they show ATLAS's own inference bugs being fixed until it reproduced Ai2's reference behaviour. Credit for the model's capability belongs to Ai2.

(measured on the live A100 endpoint, before → after the ATLAS fidelity fixes)

Table with columns: Benchmark, Before fixes, After fixes (v4.2.0), Reference
Benchmark	Before fixes	After fixes (v4.2.0)	Reference
GSM8K (25 problems)	12%	88%	Ai2 reported: 85–91% ✅
MMLU (100 questions, direct)	22%	54%	—
HumanEval (15 tasks, execution-verified)	—	73.3% pass@1	—
Needle-in-a-haystack	failed >128 tok

All remaining misses were output-length truncations; accuracy was 100% on items where reasoning completed within budget.

4× throughput: 61.7 tok/s BF16 on A100 (up from 15.4 tok/s) — zero intra-layer PCIe transfers during decode
Full GPU attention path: custom CUDA kernels handle the entire decode step on-device
New CUDA kernels: decode_attention_kernel, qk_norm_inplace_kernel, rope_precomputed_kernel, kv_cache_write_kernel, atlas_gpu_argmax
New crate atlas-infer: exposes the StigmergicHook trait — a GraphPalace bridge that lets inference hooks read/write stigmergic memory in real time during token generation

Model Details

Table with columns: Property, Value
Property	Value
Base model	allenai/Olmo-3-7B-Think
Architecture	Olmo3ForCausalLM — post-norm + QK-norm, SWA + YaRN RoPE
Parameters	7.3B
Precision	BF16 (bfloat16)
Context	65,536 tokens max (YaRN factor=8); served at 16,384 in v4.2.0
Vocab	100,278 tokens
License	Apache 2.0

Note on model weights: see Provenance & attribution above — these are Ai2's unmodified weights; ATLAS is the inference + memory-palace engine, not a fine-tune.

ATLAS Inference Engine

This model is verified to run correctly with ATLAS, a pure-Rust LLM inference framework with zero external crate dependencies. ATLAS implements the full OLMo-2/3 architecture from scratch:

Post-norm layer ordering — x = residual + rmsnorm(output) matching the HuggingFace Olmo2DecoderLayer reference
QK-norm — RMSNorm over the full 4096-dim Q/K projections before RoPE (v4.2.0 fix)
Sliding Window Attention — 24/32 layers with window=4,096, plain θ=500K RoPE; YaRN on the 8 full-attention layers only (v4.2.0 fix)
YaRN RoPE — factor=8, original_max_seq_len=8,192, corrected 2π wavelength range, attn_factor² (v4.2.0 fix)
BF16 W16A32 — weights in BF16 (14 GB VRAM), activations in f32
Official OLMo chat template — with <think> primer and Ai2 system prompt (v4.2.0)
Reasoning API — message.reasoning / streaming delta.reasoning (v4.2.0)
— repetition penalty, temperature, top-p, top-k, min-p, frequency/presence penalty

Performance (A100-SXM4-40GB, production serving)

Table with columns: Metric, Value
Metric	Value
Decode throughput	~50 tok/s sustained (BF16, live endpoint); 61.7 tok/s peak micro-bench
TTFT	~1.5 s (short prompts)
VRAM	~14 GB
CPU/GPU parity	exact match (long-sequence test)
Model load time	~108 s (3 shards, 14 GB)

Known gaps being tracked for the next cycle: batched prefill, GEMV occupancy, 32K context via BF16 KV cache.

Quick Start with ATLAS

bash
# Build ATLAS from source
git clone https://github.com/web3guru888/ATLAS.git
cd ATLAS
cargo build --release -p atlas-cli

# Download model weights
# (or use huggingface-cli: hf download openhubresearch/ATLAS-OLMo-3-7B-Think-v4)

# Start OpenAI-compatible API server
./target/release/atlas api serve \
  --weights /path/to/ATLAS-OLMo-3-7B-Think-v4 \
  --model olmo3-7b \
  --port 8080

# Query the API (reasoning is returned in message.reasoning)
curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "olmo3-7b",
    "messages": [{"role": "user", "content": "What is the capital of France?"}],
    "max_tokens": 1500,
    "temperature": 0.6,
    "top_p": 0.95
  }'

Tip: OLMo-3-Think reasons before answering — give it max_tokens ≥ 1500 so the chain of thought can complete. Leave repetition_penalty at 1.0 (Ai2 reference).

Quick Start with HuggingFace Transformers

python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "openhubresearch/ATLAS-OLMo-3-7B-Think-v4",
    torch_dtype="bfloat16",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(
    "openhubresearch/ATLAS-OLMo-3-7B-Think-v4"
)

messages = [{"role": "user", "content": "What is the capital of France?"}]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True,
    return_tensors="pt", return_dict=True
).to(model.device)

output = model.generate(
    **inputs, max_new_tokens=1500,
    temperature=0.6, top_p=0.95, do_sample=True,
)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

What's Next: the 32B chapter

With the 7B substrate at reference fidelity, work is beginning on serving OLMo-3-32B-Think on the same A100-40GB via 4-bit weight quantization — tracked in web3guru888/ATLAS#26.

About ATLAS

ATLAS (Active-inference Training with Learned Adaptive Stigmergy) is a next-generation LLM framework built in pure Rust with zero external crate dependencies — the SQLite principle applied to AI infrastructure. It fuses:

GraphPalace — Stigmergic memory palace with pheromone-guided navigation
ASTRA — Live discovery engine hitting NASA, WHO, World Bank APIs
TRM-CausalValidator — 7M-param recursive validator
Champagnat n-Morphic Framework — biologically-grounded training dynamics

22 crates. 627 tests. One coherent system. Zero external Rust dependencies.

Website: atlasagi.org · Observatory: Interactive Demo · Live API (32B only, not this model): atlas.thebeastagi.com · Organization: OpenHub Research · Author: Robin Dey

Citing

If you use the model, cite Ai2's Olmo 3 (see Licence & credit above) — not this repo. The BibTeX below is for the ATLAS inference engine only.

bibtex
@software{atlas2026,
  title       = {ATLAS: Active-inference Training with Learned Adaptive Stigmergy},
  author      = {Robin Dey},
  year        = {2026},
  institution = {OpenHub Research, Thailand},
  url         = {https://github.com/web3guru888/ATLAS},
  note        = {Pure Rust LLM framework. v4.2.0: 22 crates, 627 tests,
                 OLMo-3-7B-Think at HF-reference fidelity on A100 (BF16,
                 full GPU attention, reasoning API). GSM8K 88%, MMLU 54%,
                 HumanEval 73.3% pass@1 through the ATLAS serving stack.}
}

Available on FriendliAI

Dedicated Endpoints

Run this model inference on single tenant GPU with unmatched speed and reliability at scale.

Learn more

Model Details

Model Provider

openhubresearch

Model Tree

Base

this model

Input Modalities

Text

Output Modalities

Text

Supported Functionality

Dedicated Endpoints

Explore FriendliAI today

Get started Talk to an engineer

README

License: apache-2.0

Provenance & attribution

Upstream: allenai/Olmo-3-7B-Think · Allen Institute for AI · Apache-2.0 Relationship: mirror / verbatim redistribution. Not a derivative work of the weights.

Verify it yourself — every weight shard in this repo has the same SHA-256 as the corresponding shard in Ai2's repo (checked 2026-08-03 against the Hub's paths-info API):

Table with columns: File, Size (bytes), SHA-256, Identical to Ai2?
File	Size (bytes)	SHA-256	Identical to Ai2?
`model-00001-of-00003.safetensors`	4,969,984,976	`611c266f99a1e0575ba9f595e2f44e103aaa03a0c83b244725d6598aee934a7d`	✅ yes
`model-00002-of-00003.safetensors`	4,981,161,496	`a032a27f1768184ad0d955752e3564c14b7d1c673d1630f0df7429a439da4f41`	✅ yes
`model-00003-of-00003.safetensors`	4,644,917,240	`50ba91f88796c1ed04cacdb9271fa02d213722734dffe05d399eb27bcff7129b`	✅ yes
`config.json`	1,631	(git blob `360f3bb9…`)	✅ yes

bash
# reproduce
huggingface-cli download openhubresearch/ATLAS-OLMo-3-7B-Think-v4 model-00001-of-00003.safetensors
huggingface-cli download allenai/Olmo-3-7B-Think            model-00001-of-00003.safetensors
sha256sum ~/.cache/huggingface/hub/models--*/snapshots/*/model-00001-of-00003.safetensors

Honest small print — where this mirror is not byte-identical:

tokenizer.json is functionally identical but not byte-identical: same 100,278-entry vocabulary, same 100,000 merges, same added tokens — it was simply re-serialized by a different tokenizers version (legacy "a b" merge strings vs. Ai2's ["a","b"] pairs), which is why the file is 3.0 MB here and 7.1 MB upstream. Verified equal after normalisation.
This mirror omits several files Ai2 ships: merges.txt, vocab.json, chat_template.jinja, olmo-think.png. If you need those, use the upstream repo.

Licence & credit

bibtex
@misc{olmo2025olmo3,
  title  = {Olmo 3},
  author = {Team Olmo and others},
  year   = {2025},
  eprint = {2512.13961},
  archivePrefix = {arXiv},
  primaryClass  = {cs.CL},
  url    = {https://arxiv.org/abs/2512.13961}
}

What ATLAS adds (the serving stack — not the weights)

Latest Release — v4.2.0 (2026-07-06): Reference Fidelity

v4.2.0 — HF-Reference Fidelity + Reasoning API · 627 tests · CI green

YaRN correction range — missing 2π factor in the wavelength calculation, incorrect interpolation ramp, and attn_factor must be applied squared
Layer-type RoPE split — YaRN scaling applies only to the 8 full-attention layers; the 24 sliding-window layers use plain θ=500,000 (per configuration_olmo3.py)
QK-norm scope — RMSNorm is computed over the full 4096-dim Q/K projection, not per-head — the per-head variant silently destroyed retrieval beyond ~128 tokens

Serving upgrades in v4.2.0:

Context window raised to 16,384 tokens served (model supports up to 64K via YaRN), max output 3,584 tokens
Official OLMo chat template with server-side <think> primer and the Ai2 reference system prompt
Reasoning API: chain-of-thought surfaced as message.reasoning and streaming delta.reasoning (OpenRouter convention) — final answers stay clean, reasoning stays inspectable
Sampling defaults aligned with the Ai2 reference (repetition_penalty=1.0 — penalties make thinking models ramble)
compute-sanitizer memcheck clean; exact CPU/GPU parity verified on long sequences

Benchmarks — these measure the ATLAS serving stack running Ai2's unmodified weights

These numbers are not a claim about model quality: they show ATLAS's own inference bugs being fixed until it reproduced Ai2's reference behaviour. Credit for the model's capability belongs to Ai2.

(measured on the live A100 endpoint, before → after the ATLAS fidelity fixes)

Table with columns: Benchmark, Before fixes, After fixes (v4.2.0), Reference
Benchmark	Before fixes	After fixes (v4.2.0)	Reference
GSM8K (25 problems)	12%	88%	Ai2 reported: 85–91% ✅
MMLU (100 questions, direct)	22%	54%	—
HumanEval (15 tasks, execution-verified)	—	73.3% pass@1	—
Needle-in-a-haystack	failed >128 tok

All remaining misses were output-length truncations; accuracy was 100% on items where reasoning completed within budget.

4× throughput: 61.7 tok/s BF16 on A100 (up from 15.4 tok/s) — zero intra-layer PCIe transfers during decode
Full GPU attention path: custom CUDA kernels handle the entire decode step on-device
New CUDA kernels: decode_attention_kernel, qk_norm_inplace_kernel, rope_precomputed_kernel, kv_cache_write_kernel, atlas_gpu_argmax
New crate atlas-infer: exposes the StigmergicHook trait — a GraphPalace bridge that lets inference hooks read/write stigmergic memory in real time during token generation

Model Details

Table with columns: Property, Value
Property	Value
Base model	allenai/Olmo-3-7B-Think
Architecture	Olmo3ForCausalLM — post-norm + QK-norm, SWA + YaRN RoPE
Parameters	7.3B
Precision	BF16 (bfloat16)
Context	65,536 tokens max (YaRN factor=8); served at 16,384 in v4.2.0
Vocab	100,278 tokens
License	Apache 2.0

Note on model weights: see Provenance & attribution above — these are Ai2's unmodified weights; ATLAS is the inference + memory-palace engine, not a fine-tune.

ATLAS Inference Engine

This model is verified to run correctly with ATLAS, a pure-Rust LLM inference framework with zero external crate dependencies. ATLAS implements the full OLMo-2/3 architecture from scratch:

Post-norm layer ordering — x = residual + rmsnorm(output) matching the HuggingFace Olmo2DecoderLayer reference
QK-norm — RMSNorm over the full 4096-dim Q/K projections before RoPE (v4.2.0 fix)
Sliding Window Attention — 24/32 layers with window=4,096, plain θ=500K RoPE; YaRN on the 8 full-attention layers only (v4.2.0 fix)
YaRN RoPE — factor=8, original_max_seq_len=8,192, corrected 2π wavelength range, attn_factor² (v4.2.0 fix)
BF16 W16A32 — weights in BF16 (14 GB VRAM), activations in f32
Official OLMo chat template — with <think> primer and Ai2 system prompt (v4.2.0)
Reasoning API — message.reasoning / streaming delta.reasoning (v4.2.0)
— repetition penalty, temperature, top-p, top-k, min-p, frequency/presence penalty

Performance (A100-SXM4-40GB, production serving)

Table with columns: Metric, Value
Metric	Value
Decode throughput	~50 tok/s sustained (BF16, live endpoint); 61.7 tok/s peak micro-bench
TTFT	~1.5 s (short prompts)
VRAM	~14 GB
CPU/GPU parity	exact match (long-sequence test)
Model load time	~108 s (3 shards, 14 GB)

Known gaps being tracked for the next cycle: batched prefill, GEMV occupancy, 32K context via BF16 KV cache.

Quick Start with ATLAS

bash
# Build ATLAS from source
git clone https://github.com/web3guru888/ATLAS.git
cd ATLAS
cargo build --release -p atlas-cli

# Download model weights
# (or use huggingface-cli: hf download openhubresearch/ATLAS-OLMo-3-7B-Think-v4)

# Start OpenAI-compatible API server
./target/release/atlas api serve \
  --weights /path/to/ATLAS-OLMo-3-7B-Think-v4 \
  --model olmo3-7b \
  --port 8080

# Query the API (reasoning is returned in message.reasoning)
curl http://localhost:8080/v1/chat/completions \
  -H "Content-Type: application/json" \
  -d '{
    "model": "olmo3-7b",
    "messages": [{"role": "user", "content": "What is the capital of France?"}],
    "max_tokens": 1500,
    "temperature": 0.6,
    "top_p": 0.95
  }'

Tip: OLMo-3-Think reasons before answering — give it max_tokens ≥ 1500 so the chain of thought can complete. Leave repetition_penalty at 1.0 (Ai2 reference).

Quick Start with HuggingFace Transformers

python
from transformers import AutoModelForCausalLM, AutoTokenizer

model = AutoModelForCausalLM.from_pretrained(
    "openhubresearch/ATLAS-OLMo-3-7B-Think-v4",
    torch_dtype="bfloat16",
    device_map="auto",
)
tokenizer = AutoTokenizer.from_pretrained(
    "openhubresearch/ATLAS-OLMo-3-7B-Think-v4"
)

messages = [{"role": "user", "content": "What is the capital of France?"}]
inputs = tokenizer.apply_chat_template(
    messages, add_generation_prompt=True,
    return_tensors="pt", return_dict=True
).to(model.device)

output = model.generate(
    **inputs, max_new_tokens=1500,
    temperature=0.6, top_p=0.95, do_sample=True,
)
print(tokenizer.decode(output[0][inputs["input_ids"].shape[1]:], skip_special_tokens=True))

What's Next: the 32B chapter

With the 7B substrate at reference fidelity, work is beginning on serving OLMo-3-32B-Think on the same A100-40GB via 4-bit weight quantization — tracked in web3guru888/ATLAS#26.

About ATLAS

GraphPalace — Stigmergic memory palace with pheromone-guided navigation
ASTRA — Live discovery engine hitting NASA, WHO, World Bank APIs
TRM-CausalValidator — 7M-param recursive validator
Champagnat n-Morphic Framework — biologically-grounded training dynamics

22 crates. 627 tests. One coherent system. Zero external Rust dependencies.

Website: atlasagi.org · Observatory: Interactive Demo · Live API (32B only, not this model): atlas.thebeastagi.com · Organization: OpenHub Research · Author: Robin Dey

Citing

If you use the model, cite Ai2's Olmo 3 (see Licence & credit above) — not this repo. The BibTeX below is for the ATLAS inference engine only.

bibtex
@software{atlas2026,
  title       = {ATLAS: Active-inference Training with Learned Adaptive Stigmergy},
  author      = {Robin Dey},
  year        = {2026},
  institution = {OpenHub Research, Thailand},
  url         = {https://github.com/web3guru888/ATLAS},
  note        = {Pure Rust LLM framework. v4.2.0: 22 crates, 627 tests,
                 OLMo-3-7B-Think at HF-reference fidelity on A100 (BF16,
                 full GPU attention, reasoning API). GSM8K 88%, MMLU 54%,
                 HumanEval 73.3% pass@1 through the ATLAS serving stack.}
}