hopper-g

HopitAI Research / non-commercial (research-and-demo) out 29 days ago

Verified 2026-10-25 Rank #0 in Decision models · score 0 method v1 Sources 3 Downloads 30d 206 · #612 of 619 tracked

Can I use it?Research / non-commercial (research-and-demo)Non-commercial onlylicence file2026-10-24

Do people use it?206downloads, 30 days#612 of 619 we track · below the median+0 likes, downloads +6% this week0 likesHugging Face2026-10-24

Worth reading

Picked by rule from 2 tracked pieces: the lab's own post, the most shared, a field report, the latest.

Most discussed · 441 HN pts

DeepSeek Open Source FlashMLA – MLA Decoding Kernel for Hopper GPUs

FlashMLA is DeepSeek's library of optimized attention kernels powering inference of the DeepSeek-V4.1 model on NVIDIA GPUs and Huawei Ascend NPUs.

The fused norm-RoPE-attn-RoPE-cast kernel requires pre-permuting Q_b and Wv weights to achieve higher TFlops.
CUDA 13.1 decoding kernels fall back to slower widening through FP32 instead of using native bf16 conversions.

github.com 2025-02-24 441 HN pts · 108 comments 2.7k words about the hopper family family

2 best of 2 tracked

The 1 pick above, plus these 1. Top is the radar's score: discussion, source authority and the lab's own posts, fading with age.

ArticleNvidia Hopper Sweeps AI Inference Benchmarks in MLPerf Debutblogs.nvidia.com · 2022-09-08 · 139 HN pts

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: 5 Spaces.

As a service0no API provider listed
On own hardware0quantized builds
Built on00 fine-tunes · 0 adapters · 0 merges
In apps5Hugging Face Spaces

Top Spaces running it: multimodalart/jev-decision-index ♥ 802

The story so far

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

2 dated articles and 6 repo commits so far

Teal is the weights (2026-09-26), amber triangles are 2 dated articles (each links to it), grey ticks are repo commits.

Reference

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

What the card says

2 cited sentences

Summary generated, cited

Hopper (G) is a LoRA adapter for the Qwen3.5-4B base model designed to answer typed decision questions.^[1] The model functions by reading the probability of each option letter and uses a per-kind calibration map.^[2]

“a LoRA adapter for Qwen3.5-4B”
“reading the probability of each option letter, with a per…”

Generated 2026-10-07 15:19 UTC by parasail-gemma-4-26b-a4b-it; each numeral points at the fact it rests on. Not a verdict.

Use-case evidence

2 use cases

What it is good at 2 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

9 fields

Listed in Decision models

Facts

License: Research / non-commercial (research-and-demo) (friction: non-commercial). (license file, observed 2026-10-24)
Created 2026-09-26, last modified 2026-09-28. (HF API, observed 2026-10-24)

Library	peft
MoE	no
License	Research / non-commercial (research-and-demo) (other)
Hub snapshot	2026-10-25

Decision facts

catalog

Decision-model catalog curated

How it was made	LoRA/adapter fine-tune
Lineage bucket	F - fine-tune of an open model (public/synthetic data)
Lineage evidence	board file inspection (adapter/head/weight files) + HF base_model field
Base model	Qwen/Qwen3.5-4B-Base
DI 0.2.1 score	40.77 (independent (DI 0.2.1 board) ; author-submitted run adopted after maintainer check)
ECE (board)	0.0927
Median latency	23.3 ms (DI board, 1x RTX PRO 6000, single process)
Hosted price (input)	self-hosted/M
Source	https://huggingface.co/HopitAI/hopper-g

All variants

lineage

Lineage

Adapter Family: HopitAI/hopper 0 derivatives on the Hub

Base model: Qwen/Qwen3.5-4B

Repository changelog

6 commits

Repository changelog 6 commits

Date	Commit	SHA
2026-09-28	Hopper (G) 1.3: adapter_config task_type null -> CAUSAL_LM (Hub metadata only; weights, map and outputs unchanged); card with the pinned revision	387b2b3
2026-09-28	Hopper (G) 1.3: new adapter weights (continued from 1.2); calibration map and serving unchanged; checksums and card	8b4cd7c
2026-09-27	Card: official leaderboard results on JevBench and the Jev Decision Index, and links between Hopper and Hopper (G); weights unchanged	71d991f
2026-09-27	adapter_config.json: task_type CAUSAL_LM (Hub metadata only; weights unchanged); checksums and card note	060b1bd
2026-09-26	Hopper (G) 1.2: adapter, calibration map, checksums and model card	d60a1d6
2026-09-26	initial commit	eb3dccf

Articles and coverage

2 pieces

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

Nvidia Hopper Sweeps AI Inference Benchmarks in MLPerf Debut

blogs.nvidia.com2022-09-08HN front pageabout the hopper familyHN 139 pts · 81 comments793 words

NVIDIA H100 GPUs set world records in MLPerf inference benchmarks, delivering up to 4.5x more performance than previous-generation GPUs. The Hopper architecture, enabled by its Transformer Engine, excelled on the BERT model and raised the bar across all six neural networks in the round. These H100 GPUs will be available later this year and will participate in future MLPerf rounds for training.written, reviewed, cited

NVIDIA A100 GPUs advanced their performance by 6x since their July 2020 debut on MLPerf.
NVIDIA Orin showed up to a 50% gain in energy efficiency compared to its debut on MLPerf in April.
Every submission in the round used NVIDIA TensorRT to optimize AI inference.

Analysis1

DeepSeek Open Source FlashMLA – MLA Decoding Kernel for Hopper GPUs

github.com2025-02-24authorityHN front pageabout the hopper familyHN 441 pts · 108 comments2,705 words

FlashMLA is DeepSeek's library of optimized attention kernels powering inference of the DeepSeek-V4.1 model on NVIDIA GPUs and Huawei Ascend NPUs. The 2026.09.30 release removes support for the Hopper architecture and earlier models including DeepSeek V3 / V3.2 / V4.0 while changing the FP8 / FP4 KV cache format. On the Huawei Ascend 950 NPU, sparse attention kernels achieve up to 410 TFlops during prefill and 360 TFlops during decoding, while on B200 with CUDA 13.3, sparse decoding reaches up to 1024 TFlops and fused kernels hit up to 1460 TFlops during prefill.written, reviewed, cited

The fused norm-RoPE-attn-RoPE-cast kernel requires pre-permuting Q_b and Wv weights to achieve higher TFlops.
CUDA 13.1 decoding kernels fall back to slower widening through FP32 instead of using native bf16 conversions.

Similar models

3 models

Similar models

Bespoke-Nimble-9B-v3
same task (None)
system-one-qwen3.5-4b-scorer-v2b
same task (None)
NanoJev
same task (None)

Sources

3 sources

Sources

Hub snapshot: 2026-10-25 · HF API: 2026-10-24 21:44 UTC · Providers: 2026-10-24 21:44 UTC · Articles: 2026-10-24 21:46 UTC · Summary: 2026-10-07 15:19 UTC · Last verified 2026-10-25

Follow hopper-g

Watch this model

/feeds/models/HopitAI--hopper-g.xml(RSS 2.0) · every change to hopper-g
/feeds/families/HopitAI--hopper.xml(RSS 2.0) · every change across the hopper family

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