DFlash
Speculative decoder in AI tools
Speeds up local LLM text generation, reported at 2-3x on some models, but critics say gains can vanish with higher memory use and drops in speed.
Speeds up local LLM text generation, reported at 2-3x on some models, but critics say gains can vanish with higher memory use and drops in speed.
20 people gave an opinion
- recommend it
- 12
- criticize it
- 8
- approval share
- 60%
Recommended in 6 of 13 months
- November 2025: not recommended
- December 2025: not recommended
- January 2026: not recommended
- February 2026: not recommended
- March 2026: not recommended
- April 2026: recommended
- May 2026: recommended
- June 2026: recommended
- July 2026: recommended
- August 2026: recommended
- September 2026: recommended
- October 2026: not recommended
- November 2026: not recommended (month in progress)
Why it doesn’t carry the badge
- 12 people recommend itNeeds 10 or more, met.
- 1.5 recommenders for every criticNeeds 2 to 1 or better, not met.
- Recommended in 6 of 13 monthsNeeds 4 or more, met.
The same three rules apply to every product, and the badge can’t be bought. How the badge works
What people say about it
The reasons people give, with the number of comments that give each one. Open a reason to read what they said.
Why people recommend it
Each mark is one comment
Boosts token generation speed
“There are several Mac implementations of it that show > 2x faster Qwen3.5 already.”
Comment from 14 Apr 2026 in “Introspective Diffusion Language Models” “Z Lab had been publishing dflash addons and boosting their tgen 2-3x”
Comment from 10 Jul 2026 in “Don't discontinue Gemini 2.5 Flash” “DFlash is currently a SOTA speculative decoding method”
Comment from 10 Jun 2026 in “DiffusionGemma: 4x Faster Text Generation” - Comment from 25 Aug 2026 in “Qwen 3.8-Flash-Next releasing tomorrow (125B a6B)”
Enables models with no other path
“Before this repo, there was no path to run this... SGLang's GGUF path for this model is broken.”
Comment from 20 Apr 2026 in “We got 207 tok/s with Qwen3.5-27B on an RTX 3090” “Before this repo, there was no path to run this”
Comment from 20 Apr 2026 in “We got 207 tok/s with Qwen3.5-27B on an RTX 3090”
What people warn about
Each mark is one comment
Speed can drop off sharply
“model performance suddenly drops off a cliff like maybe there is some memory management issue”
Comment from 10 Aug 2026 in “Meta Muse Glimmer – open weights 30B local coding model” “performance suddenly drops off a cliff if I have DFlash enabled”
Comment from 10 Aug 2026 in “Meta Muse Glimmer – open weights 30B local coding model” “it craters to ~9 t/s with it on, miss rate exceeding 50%”
Comment from 10 Aug 2026 in “Meta Muse Glimmer – open weights 30B local coding model” Higher memory use and OOM crashes
“tgen is way up, but keep the number low, acceptance rate is terrible beyond half a dozen and it requires more memory”
Comment from 9 Jul 2026 in “Hy3” “Have had issues with the dflash drafter and vllm, crashes, OOMs, broken thinking”
Comment from 22 Jul 2026 in “Running Laguna S 2.1 locally on Apple Silicon: 52 tok/s with 38.5 GB peak memory” Slow or no gain for some
“Spec decode (DFlash, n-gram) didn't give a good boost”
Comment from 11 Sep 2026 in “So you want to use OpenRouter?” “Dflash2 is really slow for me, like sub 10 TG”
Comment from 18 Sep 2026 in “Shapelearn Qwen 3.8 27B (13.1 GB VRAM)” Broken output or failed tool calls
“Have had issues with the dflash drafter and vllm, crashes, OOMs, broken thinking”
Comment from 22 Jul 2026 in “Running Laguna S 2.1 locally on Apple Silicon: 52 tok/s with 38.5 GB peak memory” “DFlash2's tool call fails on python syntax.”
Comment from 19 Aug 2026 in “DFlash 2: Keep Drafting Parallel”
What people recommend instead
The same counts for each, so you can compare like for like. The tick on each bar marks the 2 to 1 line the badge needs.
| Product | Recommend | Criticize | Approval | Months | Owned | Badge |
|---|---|---|---|---|---|---|
| DFlashThis page | 12 | 8 | 60% | 6 | — | — |
| Orthrus-Qwen3 | 5 | 0 | 100% | 1 | — | — |
What people said
21 of the 22 comments that mention it, newest first. Each links to the original comment.
“help to do one forwards pass and get 6-7 tokens out of it”
“Dflash2 is really slow for me, like sub 10 TG”
“Spec decode (DFlash, n-gram) didn't give a good boost”
“setup llama.cpp with dflash2 for you”
“Dflash (2) will push your TG even further”
“DFlash2's tool call fails on python syntax.”
“has made low memory bandwidth model usage so much nicer”
“I have exactly the same issue on a M1 Max as well.”
“it craters to ~9 t/s with it on, miss rate exceeding 50%”
“model performance suddenly drops off a cliff like maybe there is some memory management issue”
“performance suddenly drops off a cliff if I have DFlash enabled”
“Have had issues with the dflash drafter and vllm, crashes, OOMs, broken thinking”
“Z Lab had been publishing dflash addons and boosting their tgen 2-3x”
“This is actually what I was hoping would happen when the deepspec stuff dropped!”
“tgen is way up, but keep the number low, acceptance rate is terrible beyond half a dozen and it requires more memory”
“DFlash is currently a SOTA speculative decoding method”
“you should be using dflash with that model, look it up”
“Before this repo, there was no path to run this”
“Q4 is a non-starter for real work. The fact that you're still trying to defend it tells me you haven't used this for anything other than token/sec racing.”
“Before this repo, there was no path to run this... SGLang's GGUF path for this model is broken.”
“There are several Mac implementations of it that show > 2x faster Qwen3.5 already.”
Counts come from 1,795,640 Hacker News comments, November 2025 to November 2026. How we count. Report a mistake.