Strata
Hacker ApprovedRanked 2 of 27 inference engines by recommenders
Runs large local LLMs fast on consumer hardware. Caveats: not open source, and quantized models may make more mistakes.
Runs large local LLMs fast on consumer hardware. Caveats: not open source, and quantized models may make more mistakes.
46 people gave an opinion
- recommend it
- 31
- criticize it
- 15
- approval share
- 67%
Recommended in 5 of 13 months
- November 2025: not recommended
- December 2025: recommended
- January 2026: not recommended
- February 2026: not recommended
- March 2026: not recommended
- April 2026: not recommended
- May 2026: not recommended
- June 2026: not recommended
- July 2026: recommended
- August 2026: not recommended
- September 2026: recommended
- October 2026: recommended
- November 2026: recommended (month in progress)
Why it carries the badge
- 31 people recommend itNeeds 10 or more, met.
- 2.1 recommenders for every criticNeeds 2 to 1 or better, met.
- Recommended in 5 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
Runs big models on modest hardware
“It will run reasonably fast if you have 64+ gb ram and use strata or a strata fork”
Comment from 20 Oct 2026 in “MiniMax M3.1 Flash” “You can get something relatively capable and usable with your configuration”
Comment from 19 Oct 2026 in “DeepSeek V4.1 Flash 1017 (Max)” “It is fully possible to run the 180B Owen model on strix halo using strata”
Comment from 9 Oct 2026 in “Step 5 Preview, a 1M-context MoE from StepFun, shows up on OpenRouter” - Comment from 15 Oct 2026 in “LFM2.5-24B-A2B: Built for fast on-device agents”
- Comment from 29 Sep 2026 in “Dots: Always-on agents”
- Comment from 4 Oct 2026 in “Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s”
- Comment from 9 Oct 2026 in “Why isn't the industry freaking out about DeepSeek 4.1 Flash?”
- Comment from 27 Oct 2026 in “New iMac with M6”
Faster token speeds than llama.cpp
“the results are not even close. You can basically ignore everything in that table and use Strata instead”
Comment from 27 Oct 2026 in “New iMac with M6” “I've switched to IQ3_XXS and am running at 60 t/s on Strata vs the 21 t/s I was getting in llama.cpp.”
Comment from 4 Oct 2026 in “Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s” “Strata has really shown how much performance is left on the table”
Comment from 17 Oct 2026 in “Show HN: Llmspeed – feel how fast a local LLM runs before you buy the hardware” - Comment from 5 Oct 2026 in “Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s”
- Comment from 4 Oct 2026 in “Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s”
- Comment from 5 Oct 2026 in “Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s”
Stable with no crashes
“its been running on my system for days without any crash”
Comment from 19 Oct 2026 in “DeepSeek V4.1 Flash 1017 (Max)” “without inference crashing before long”
Comment from 19 Oct 2026 in “DeepSeek V4.1 Flash 1017 (Max)” Opens huge log files quickly
“it takes 8 seconds to open a 50GB file and once you do scrolling is buttery smooth with 60fps”
Comment from 17 Dec 2025 on Hacker News “can open a 50GB log file in 8 seconds and not eat your RAM”
Comment from 17 Dec 2025 in “Ask HN: What Are You Working On? (December 2025)”
What people warn about
Each mark is one comment
Not open source
“though I'd find it much more compelling if it were OSS”
Comment from 30 Sep 2026 in “Show HN: Strata – an expressive semantic layer that can say no to your LLM” “I'd find it much more compelling if it were OSS.”
Comment from 30 Sep 2026 in “Show HN: Strata – an expressive semantic layer that can say no to your LLM” Quantization hurts output quality
“they quantize, worse they quantize KV caches, they approximate expert routing etc.”
Comment from 31 Oct 2026 in “OpenAI raises less than half of $30B target” “perhaps such metric is the most challenging one for a brutally quantized and otherwise lobotomized model”
Comment from 5 Oct 2026 in “Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s” “same model running on Strata made more mistakes and hallucinated more when compare to llama.cpp”
Comment from 5 Oct 2026 in “Run Qwen 3.8 Flash Next (125B) on consumer hardware (RTX 4090) at 100T/s” Dangerous for outdoor safety decisions
“I saw no evidence of trials to test to see how the technology performs against real conditions.”
Comment from 7 Jul 2026 in “Strata – An app that talks me out of dying outdoors” “Llms never took a difficult on field decision, never had to turn gut feeling into calls.”
Comment from 7 Jul 2026 in “Strata – An app that talks me out of dying outdoors” “This is going to kill someone... you can't just make these calls for people.”
Comment from 7 Jul 2026 in “Strata – An app that talks me out of dying outdoors” - Comment from 7 Jul 2026 in “Strata – An app that talks me out of dying outdoors”
- Comment from 7 Jul 2026 in “Strata – An app that talks me out of dying outdoors”
Other inference engines people recommend
The four most recommended other inference engines, with 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 |
|---|---|---|---|---|---|---|
| vLLM | 112 | 56 | 67% | 12 | — | Approved |
| StrataThis page | 31 | 15 | 67% | 5 | — | Approved |
| DwarfStar 4 | 26 | 2 | 93% | 6 | — | Approved |
| SGLang | 19 | 13 | 59% | 8 | — | — |
| ninfer | 16 | 2 | 89% | 3 | — | — |
What people said
40 of the 53 comments that mention it, newest first. Each links to the original comment.
“these models are the most I have ever gotten out of my machine”
“You can already run them locally on a top-shelf gamer PC with decent token throughput”
“they quantize, worse they quantize KV caches, they approximate expert routing etc.”
“the results are not even close. You can basically ignore everything in that table and use Strata instead”
“Yes it runs on all levels of hardware”
“this repo is pretty amazing”
“In fact, I myself use Strata and Opencode, so for advanced users, it recommends those”
“the problem with Strata is that you cannot easily share the server as the performance will collapse outside of a single user”
“at very respectable token rates”
“It will run reasonably fast if you have 64+ gb ram and use strata or a strata fork”
“its been running on my system for days without any crash”
“without inference crashing before long”
“You can get something relatively capable and usable with your configuration”
“thanks to a local Qwen-3.8 running under Strata and Hermes”
“Strata has really shown how much performance is left on the table”
“It runs quite fast on 16+64, you should be able to get 50 tok/s or so.”
“Strata has been such a boon to my own local development”
“Strata already does this with Qwen3.8-Flash-Next”
“It is fully possible to run the 180B Owen model on strix halo using strata”
“it rips with just 64 ram and a 9070xt”
“same model running on Strata made more mistakes and hallucinated more when compare to llama.cpp”
“perhaps such metric is the most challenging one for a brutally quantized and otherwise lobotomized model”
“running the swift iq2_xs quant, ukisai/Swift-1.5-Qwen3.8-Flash-Next-GSQ-RCO-GGUF with strata”
“with Strata I'm doing 60 t/s”
“on Strata I can run it at 60 t/s”
“is amazing running with Strata”
“I've switched to IQ3_XXS and am running at 60 t/s on Strata vs the 21 t/s I was getting in llama.cpp.”
“You can run it on even more modest hardware with Strata too.”
“I'd find it much more compelling if it were OSS.”
“though I'd find it much more compelling if it were OSS”
“I’m a big fan of the strict naming as an abstraction above tables”
“lets me run Qwen 3.8 Flash Next at home on a single 3090 at 90 tok/s”
“More AI hiking crap that will probably kill someone”
“Llms never took a difficult on field decision, never had to turn gut feeling into calls.”
“I saw no evidence of trials to test to see how the technology performs against real conditions.”
“This is going to kill someone... you can't just make these calls for people.”
“Absolutely the wrong application for an LLM.”
“Otherwise if actually keeping files on disc is no obstacle, the web sockets api works just fine”
“it takes 8 seconds to open a 50GB file and once you do scrolling is buttery smooth with 60fps”
“can open a 50GB log file in 8 seconds and not eat your RAM”
Counts come from 1,795,640 Hacker News comments, November 2025 to November 2026. How we count. Report a mistake.