Datasets:

witcheer
/

rtx-5090-benchmarks

Rig: one RTX 5090 32GB · transformers 5.11 (bf16, SDPA) · greedy, temp 0 Model: sapientinc/HRM-Text-1B (Apache 2.0, 1.18B, bf16 safetensors) — the Hierarchical Reasoning Model line (arXiv 2605.20613) scaled from the 27M ARC-AGI puzzle solver to a text base model. #1 HF text-gen trending the week of 2026-06-10, with no published decode-speed numbers anywhere. These appear to be the first.

Setup: GSM8K test, 200 questions (seed 42), generative + greedy, answers extracted with the rig's first-committed-answer rule, raw generations captured for offline re-scoring. The model's documented protocol: prompt wrapped as <|im_start|>{condition}{question}<|im_end|>, condition tokens <|quad_end|><|object_ref_end|> (synth+cot), and token_type_ids=1 over the whole prompt — that last tensor marks the prompt as a bidirectional PrefixLM prefix. Run twice: once correct, once the way every standard harness would run it (no token_type_ids).

The numbers

protocol	GSM8K (n=200, 95% CI)	decode	peak VRAM
paper claim	84.5 (protocol n/a)	—	—
correct (prefixlm mask)	79.5% ± 5.6	42.9 tok/s	2.41 GB
default harness (no mask)	53.5% ± 6.9	42.9 tok/s	2.41 GB

The claim holds. 79.5% at n=200 puts the paper's 84.5 just inside the confidence interval — a slight shortfall, not a refutation, and protocol differences (their condition tag, extraction rule) could absorb most of it. A 1B base model with a ~$1,500 training budget scoring ~80% generative GSM8K is the real story; it also answers the bat-and-ball trap correctly in 5 tokens in direct mode.

The trap: −26 points from one missing tensor

HRM-Text is trained with PrefixLM masking: the prompt attends bidirectionally, generation is causal. At inference that mask is requested via token_type_ids=1 — a tensor no standard eval harness passes (lm-eval-harness, lighthouse-style loglikelihood runners, naive generate() wrappers all omit it). Omitting it doesn't error. The model just runs causal-only and quietly loses 26 points (79.5 → 53.5).

Two practical consequences:

Community numbers for this model will disagree wildly, and the low ones will be measuring their harness, not the model. (The rig's oldest lesson — a capable model scoring absurdly low is almost always the harness — now has its cleanest quantified example.)
Condition tokens matter too: in smoke tests, synth+cot produced correct Rayleigh-scattering physics where direct/cot confabulated on the same prompt. The tags select training-data distributions, not output formats.

The recurrence bill

The architecture runs two 16-layer stacks in a nested recurrence (2 H-cycles × 3 L-cycles + 1 update per H — 128 layer invocations per forward, "latent reasoning" instead of emitted CoT). Measured costs:

42.9 tok/s decode (bf16, batch 1, dead-stable across runs) — a 1.2B that decodes like a ~5B dense.
128 KV-cache slots, 0.88 MB/token measured — 4× the cache of a normal 32-layer model (~3.6 GB at the full 4K context).
Weights are only 2.2 GB and peak VRAM 2.41 GB — it ran alongside a live 27B llama-server the whole time.
No llama.cpp path (arch unsupported upstream, discussion #23415), so no GGUF quant rescue: transformers is the deployment story today, and these are the speeds.

Honest caveats

n=200 sample, ±5.6 pts CI — fine for verify-or-refute, not for leaderboard precision.
The hierarchy attribution is untested here. The authors dropped ACT/halting (the component the ARC Prize post-hoc analysis found most load-bearing in the 2025 HRM), and the paper's own FLOPs-matched ablations credit the task-completion training objective for the first chunk of the gains, before PrefixLM and the HRM recurrence add theirs. What's verified is the package, not why.
Context is 4096 — this is a proof-of-concept base model, not a daily driver.

Reproduce

scripts/hrm_probe.py (smoke / decode bench / KV probe / mask ablation) · scripts/hrm_gsm8k.py --n 200 [--no-prefix-mask] · raw generations: raw/hrm-gsm8k-synth-cot-n200.jsonl, raw/hrm-gsm8k-synth-cot-nomask-n200.jsonl (this dataset).