Pinned Tweet
how do we collectively use technology…
to map our future…
with integrity…
mindful of the perspectives of others…
and future generations?
to map our future…
with integrity…
mindful of the perspectives of others…
and future generations?
0
0
50
36,761
Tom Lynch retweeted
Here's some other models you can run to Reverse Engineer Anything.
I recommend using @actualinc for inference and @NousResearch Hermes Agent for everything else.
@huggingface / @thepirateface
8GB: Huihui Qwen3.5-4B abliterated (Q4)
hf.co/huihui-ai/Huih…
16GB: Huihui Qwen3.5-9B abliterated (Q4)
hf.co/huihui-ai/Huih…
32GB: Huihui Qwen3-Coder-30B-A3B abliterated (Q4)
hf.co/huihui-ai/Huih…
64GB: Huihui Qwen3-Coder-Next 80B abliterated (Q4)
hf.co/huihui-ai/Huih…
128GB+: Huihui Qwen3-Coder-Next 80B abliterated (Q8)
hf.co/huihui-ai/Huih…
I recommend using @actualinc for inference and @NousResearch Hermes Agent for everything else.
@huggingface / @thepirateface
8GB: Huihui Qwen3.5-4B abliterated (Q4)
hf.co/huihui-ai/Huih…
16GB: Huihui Qwen3.5-9B abliterated (Q4)
hf.co/huihui-ai/Huih…
32GB: Huihui Qwen3-Coder-30B-A3B abliterated (Q4)
hf.co/huihui-ai/Huih…
64GB: Huihui Qwen3-Coder-Next 80B abliterated (Q4)
hf.co/huihui-ai/Huih…
128GB+: Huihui Qwen3-Coder-Next 80B abliterated (Q8)
hf.co/huihui-ai/Huih…
anthropic just killed the REA
it stops claude from doing it
that’s why local models are important
4
4
46
2,273
Tom Lynch retweeted
Actual is back. We migrated everything to bare metal. Service should be restored and a new update is available for your client which closes any upstream exposure to the massive amount of CVEs floating around here recently. One again, your computers were 100% safe + none of your info was ever public or leaked in any way. Sorry for the downtime but doing it this way was the only way. The team were like organized but feral wolves today
6
5
104
2,763
No need to pick low latency mode with e
We auto adapt to what you want
We already know whats best
We auto adapt to what you want
We already know whats best
Introducing SGLang Low-Latency Mode, built with @NVIDIAAI and @AMD.
On DeepSeek-V4.1-Flash with DSpark, it delivers up to 1.38× P90 per-user token speed on NVIDIA GB300 (671 to 928 tok/s) and 1.56× on AMD MI355X (332 to 516 tok/s) in @SemiAnalysis_'s AgentX trace replay.
overlapping kernels cut decoder layer time from 96.8 to 68.4 µs
- AMD: 2 persistent kernels per layer replace dozens of small ones, with cross-GPU communication overlapped inside them.
- New plugin system: custom kernels plug in from external libraries, with automatic fallback.
- Low-latency kernels for GLM-5.3, K3, and MiniMax M3 are in progress and will plug in next.
0
0
21
928
e now supports
- nemotron NVFP4
- GLM 5.3 NVFP4
for your GB300 station
We took time and care to make sure our NVFP4 was done right
throughput beat the hell out of everyone
your loss if you own a station and dont use actual.
- nemotron NVFP4
- GLM 5.3 NVFP4
for your GB300 station
We took time and care to make sure our NVFP4 was done right
throughput beat the hell out of everyone
your loss if you own a station and dont use actual.
Actual Computer turns your NVIDIA DGX Station into a department-scale AI factory.
Read the full case study:
actual.inc/company/blog/c…
4
9
54
2,555
Get even higher TPS for your station with @actualinc
The highest throughput engine for the dgx station
The highest throughput engine for the dgx station
dario begged me in dms to not do agent swarms at home but whatever, this is nemotron 120B A12B running 64 concurrents at a bit over 100 tok/sec each for an aggregate of almost 7K tok/sec on a single GB300 (Asus DGX Station)
believe it or not, the agents did NOT magically just t tho. weird.
0
5
39
1,966
Tom Lynch retweeted
all models on models.actual.inc (even the decision models) use toks as the tokenizer now
6
7
61
3,186
I think its funny how normies are in love with “ZDR” and crypto people are obsessed with “confidential compute”
But neither has gotten excited about real privacy yet
Own your computer
But neither has gotten excited about real privacy yet
Own your computer
2
7
72
2,120
These things are so fast
RTX Spark is now available for purchase!
The first laptop that can do it all - it's thin, light, has good battery, has lots of memory for AI, it can game, it can create... Simply put, it's the best Windows Laptop.
youtube.com/watch?v=vJvFta…
0
1
21
1,175
woke up and it feels like christmas
for the first time the Actual Computer twitter account has more followers than me
LFGGG
for the first time the Actual Computer twitter account has more followers than me
LFGGG
14
0
108
1,848
running an agent factory reminds me a lot of a glass furnace.
for industrial glass furnaces, they never turn them off, they just run 24/7 hot because once its running, letting it cool is a total mess.
a software factory is often like that.
for industrial glass furnaces, they never turn them off, they just run 24/7 hot because once its running, letting it cool is a total mess.
a software factory is often like that.
1
0
45
1,235
my favorite part about showing e to people
is when i show them tok/s they always think theres gonna be a catch
'oh the ttft is gonna be slow'
'hmm but is its probably lobotomized'
'the slo between tokens will be sloppy'
then you look close and its all somehow better than vllm
is when i show them tok/s they always think theres gonna be a catch
'oh the ttft is gonna be slow'
'hmm but is its probably lobotomized'
'the slo between tokens will be sloppy'
then you look close and its all somehow better than vllm
3
4
49
1,584
When serving real agent traffic
e's p95 for ttft
is near vllm's p50
thats what delivering the machine roofline really looks like
e's p95 for ttft
is near vllm's p50
thats what delivering the machine roofline really looks like
An agent swarm needs more than a fast first request. It needs the whole machine to stay responsive as agents join in.
In a single stream, e serves Nemotron 3.5 Super VL at 455 tokens per second. A fresh 8K-token prompt gets its first token back in 0.58 seconds.
1
3
42
1,884
Normally when you load a model with other leading engines, you're forced to wait minutes before you get a fully functional server.
When you use Actual thats not the case. Cold starts in less than 20 seconds and cached starts in less than 4 seconds!
best engine everrr
When you use Actual thats not the case. Cold starts in less than 20 seconds and cached starts in less than 4 seconds!
best engine everrr
An agent swarm needs more than a fast first request. It needs the whole machine to stay responsive as agents join in.
In a single stream, e serves Nemotron 3.5 Super VL at 455 tokens per second. A fresh 8K-token prompt gets its first token back in 0.58 seconds.
3
3
55
2,217