DOdot_treo·May 16, 2026
And it also looks like the original authors are working on qwen 3.5 too:
https://github.com/chiennv2000/orthrus/issues/1#issuecomment...
In Orthrus-Qwen3: up to 7.8×tokens/forward on Qwen3, identical output distributionchiennv2000
orthrus
Python2026–2026
chiennv2000 / orthrus
orthrus
[NeurIPS 2026 Spotlight] Fast, lossless LLM inference via dual-view diffusion decoding.
2Distinct people
3Mentions
0Explicit recommenders
482GitHub stars
What to know before exploring
What it is
[NeurIPS 2026 Spotlight] Fast, lossless LLM inference via dual-view diffusion decoding.
GitHub
Where it is used
No documented product use has been verified yet.
What people are saying
No recommendation phrase was found in the indexed comments.
No classified warning is attached to this page. Read the discussions before treating silence as approval.
Also in these discussions
No other repository is repeatedly linked in the same discussions.
This week on Hacker News
Nov 7–Nov 14 UTC · Direct repository links and separately labeled related project stories. Indexed through 11/14/2026, 5:04:57 PM UTC.
No direct repository links were found in the indexed HN discussions over the past seven days. Older discussions are available below.
Mentions over time
HN mentions by hour3 dated mentions
Observed mentions: 2026-05-15 22:38:58 UTC to 2026-05-16 19:45:39 UTC.
Counts come from the repository’s dated source evidence. Index cutoff: 2026-11-02 08:16:59 UTC.
View mention counts and date ranges
| Period | Start (UTC) | End (exclusive, UTC) | Mentions | Submissions | Comments |
|---|---|---|---|---|---|
| May 15, 22:00 | 2026-05-15 22:00:00 UTC | 2026-05-15 23:00:00 UTC | 2 | 1 | 1 |
| May 15, 23:00 | 2026-05-15 23:00:00 UTC | 2026-05-16 00:00:00 UTC | 0 | 0 | 0 |
| May 16, 00:00 | 2026-05-16 00:00:00 UTC | 2026-05-16 01:00:00 UTC | 0 | 0 | 0 |
| May 16, 01:00 | 2026-05-16 01:00:00 UTC | 2026-05-16 02:00:00 UTC | 0 | 0 | 0 |
| May 16, 02:00 | 2026-05-16 02:00:00 UTC | 2026-05-16 03:00:00 UTC | 0 | 0 | 0 |
| May 16, 03:00 | 2026-05-16 03:00:00 UTC | 2026-05-16 04:00:00 UTC | 0 | 0 | 0 |
| May 16, 04:00 | 2026-05-16 04:00:00 UTC | 2026-05-16 05:00:00 UTC | 0 | 0 | 0 |
| May 16, 05:00 | 2026-05-16 05:00:00 UTC | 2026-05-16 06:00:00 UTC | 0 | 0 | 0 |
| May 16, 06:00 | 2026-05-16 06:00:00 UTC | 2026-05-16 07:00:00 UTC | 0 | 0 | 0 |
| May 16, 07:00 | 2026-05-16 07:00:00 UTC | 2026-05-16 08:00:00 UTC | 0 | 0 | 0 |
| May 16, 08:00 | 2026-05-16 08:00:00 UTC | 2026-05-16 09:00:00 UTC | 0 | 0 | 0 |
| May 16, 09:00 | 2026-05-16 09:00:00 UTC | 2026-05-16 10:00:00 UTC | 0 | 0 | 0 |
| May 16, 10:00 | 2026-05-16 10:00:00 UTC | 2026-05-16 11:00:00 UTC | 0 | 0 | 0 |
| May 16, 11:00 | 2026-05-16 11:00:00 UTC | 2026-05-16 12:00:00 UTC | 0 | 0 | 0 |
| May 16, 12:00 | 2026-05-16 12:00:00 UTC | 2026-05-16 13:00:00 UTC | 0 | 0 | 0 |
| May 16, 13:00 | 2026-05-16 13:00:00 UTC | 2026-05-16 14:00:00 UTC | 0 | 0 | 0 |
| May 16, 14:00 | 2026-05-16 14:00:00 UTC | 2026-05-16 15:00:00 UTC | 0 | 0 | 0 |
| May 16, 15:00 | 2026-05-16 15:00:00 UTC | 2026-05-16 16:00:00 UTC | 0 | 0 | 0 |
| May 16, 16:00 | 2026-05-16 16:00:00 UTC | 2026-05-16 17:00:00 UTC | 0 | 0 | 0 |
| May 16, 17:00 | 2026-05-16 17:00:00 UTC | 2026-05-16 18:00:00 UTC | 0 | 0 | 0 |
| May 16, 18:00 | 2026-05-16 18:00:00 UTC | 2026-05-16 19:00:00 UTC | 0 | 0 | 0 |
| May 16, 19:00 | 2026-05-16 19:00:00 UTC | 2026-05-16 20:00:00 UTC | 1 | 0 | 1 |
What Hacker News says
3 sourcesOrthrus-Qwen3: up to 7.8×tokens/forward on Qwen3, identical output distribution
246 points · Open discussionPaper: https://arxiv.org/abs/2605.12825 ; Code+models: https://github.com/chiennv2000/orthrus ; Disclosure: co-author.
Idea: Inject a trainable diffusion attention module into each layer of a frozen AR Transformer. Both heads share one KV cache. Diffusion head projects K=32 tokens in parallel; AR head verifies in a second pass and accepts the longest matching prefix. Output distribution is provably identical to the base model.
Results:
- Up to 7.8x TPF, ~6x wall-clock on MATH-500.
- 16% of params trained, <1B tokens, 24h on 8xH200.
- vs. diffusion LMs (Dream, Fast-dLLM-v2, SDAR, Mercury, Gemini Diffusion): they modify base weights and lose accuracy (Fast-dLLM-v2: -11 pts on MATH-500). Orthrus freezes the backbone; accuracy matches Qwen3-8B exactly.
- vs. Speculative Decoding (EAGLE-3, DFlash): no external drafter, no separate cache, zero
Read the rest
TTFT penalty (no drafter to init/sync). KV overhead is O(1) (~4.5 MiB flat). Acceptance length on MATH-500: 11.7 vs. 7.9 (DFlash) vs. 3.5 (EAGLE-3).
- Single-step denoising beats multi-step (6.35 vs. 3.53 TPF). KL distillation beats CE on acceptance rate.
Limitations: strictly bounded by the frozen base model (inherits its biases, hallucinations, knowledge gaps); Qwen3-only evaluation; greedy + rejection sampling only.