Voltropy
@Voltropy

The long-context AI lab. Makers of Vast-10M.

voltropy.com
Joined July 2025
7 Photos and videos
Pinned Tweet
Voltropy
@Voltropy
Sep 29
Introducing Vast-10M from Voltropy: - The first frontier LLM with a ten million token context window - Beats Anthropic's Fable 5.1 at long-context reasoning - Early access opens TODAY
52
57
442
290,487
Voltropy retweeted
Jason Kneen
@jasonkneen
Oct 21
This should be interesting thank you @Voltropy
1
4
553
VAST by @Voltropy has a 10M-token context window, which is hard to picture. So here's a picture of it: Aristotle, the Mahabharata, the Bible, the Ramayana, Homer, Plato, the Qur’an, the Gita, Gilgamesh, Confucius, Sun Tzu, Lao Tzu. All of it fits, with 2.5M tokens to spare.
2
3
688
Trying out VAST from @Voltropy inside codex
2
5
510
At ten million tokens, Vast-10M-Flash exceeds the performance of its base model at 1M tokens. It matches the performance of GLM 5.2 at 500k tokens. In other words, the models perform the same when Vast-10M-Flash has to process 20x more context.
1
4
513
Vast-10M-Flash is our strongest model, but our experiments validated VSA in all three models. VSA's scaling properties boost performance at shorter contexts and retain it all the way to 10M tokens.
1
5
475
We aren't yet ready to publicly share the details of how VSA works. But we invite you to try it for yourself in early access at voltropy.com. Our blog post there goes into more detail about why we created this technology and how it can make AI smarter and safer.

Vast-10M · Voltropy

The first frontier LLM with ten million tokens of native context.

voltropy.com
13
448
Vast-10M-Flash doesn't just beat its base model, DeepSeek v4.0 Flash. It beats its base model's replacement. We unexpectedly unlocked more capability by adding VSA than DeepSeek did by training a new model.
2
9
564
EXPLAINER: How Voltropy built Vast-10M, the best long-context LLM in the world. We invented a new form of scalable attention (VSA) and retrofitted it to legacy DeepSeek weights. Suddenly the old weights could compete head-to-head with OpenAI and Anthropic's strongest models.
6
6
115
9,430
We tested VSA on three model sizes spanning two model families. Every VSA-equipped model beat every DeepSeek and GLM baseline at every context length. It was a 9-0 victory for VSA.
5
14
908
Introducing Vast-10M from Voltropy: - The first frontier LLM with a ten million token context window - Beats Anthropic's Fable 5.1 at long-context reasoning - Early access opens TODAY
52
57
442
290,487
Early access to Vast-10M is now open at voltropy.com The technical paper is also live. Come read how Vast-10M measures up to Fable 5.1 and GPT-6 Astra.

Vast-10M · Voltropy

The first frontier LLM with ten million tokens of native context.

voltropy.com
2
22
2,459
Voltropy retweeted
Dave Font
@davefontenot
May 22
Some founders are fighting for b2b saas. And others are fighting to win it all. This conversation is the moment I knew Ted and @ClintEhrlich were taking on Dario.
HFØ
@hf0
May 22
5 of 10 teams at demo day broke $10M annualized HF0 is not just for early stage companies Interviews started yesterday.
4
27
8,930
Voltropy retweeted
HFØ
@hf0
May 22
5 of 10 teams at demo day broke $10M annualized HF0 is not just for early stage companies Interviews started yesterday.
36
44
560
282,656
Voltropy retweeted
Josh Lehman
@jlehman_
May 18
lossless-claw 0.11.1 — the focus mode release 🎯 /lossless focus curates your context ↩️ /lossless unfocus brings the normal context view back 🖼️ image externalization now works across roles 📦 installs stop pulling a second OpenClaw
13
6
112
31,539
"LCM: Lossless Context Management" (Ehrlich and Blackman, 2026) is now available on arXiv. This is the paper that caused @openclaw to modify its architecture to support context management. It's a must-read for anyone interested in RLMs vs LCM. arxiv.org/abs/2605.04050

LCM: Lossless Context Management

We introduce Lossless Context Management (LCM), a deterministic architecture for LLM memory that outperforms Claude Code on long-context tasks. When benchmarked using Opus 4.6, our LCM-augmented...

arxiv.org
1
1
14
1,861
Voltropy retweeted
Josh Lehman
@jlehman_
May 3
lossless-claw 0.9.3 — the "please just keep working" release 🧠 cache-aware compaction fires before overflow 🔁 fewer repeated old instructions 🧰 lcm tools load on OpenClaw 2026.5.2+ 🔌 Codex, DeepSeek, Bedrock fixes 🛡️ safer migrations, payloads, and replay
12
15
215
40,280
🚀 hermes-lcm v0.7.1 is live! 🧠 Compression rollover continuity 🔌 Provider-prefixed model routing 🧼 Thinking-tag cleanup Update: git -C ~/.hermes/plugins/hermes-lcm pull --ff-only Bounded context, unbounded memory. Nothing is ever lost. @Voscko @Teknium @NousResearch github.com/stephenschoettler…

Release v0.7.1 · stephenschoettler/hermes-lcm

chore: bump version to 0.7.1 (156391c) fix: continue LCM state across compression boundaries (#76) (015ab3d) fix: route provider-prefixed LCM models (#72) (b41836a) fix: accept valid integer node_i...

github.com
6
8
121
11,098
Voltropy retweeted
Dave Font
@davefontenot
Mar 26
Recursive self-improving AI is here. I had heard people talking about self-improving AIs at some of the frontier labs but wasn't sure how real it was. One of the neo labs in the current @hf0 batch had a breakthrough and now it’s running in our basement. It’s real.
28
18
239
26,226
Lossless Context in OpenClaw feat. Josh Lehman nitter.siddz.com/i/broadcasts/1AKEmONzY…
4
9
632
Voltropy retweeted
Josh Lehman
@jlehman_
Mar 16
Fun fact about lossless-claw: in addition to solving agent amnesia and enabling infinite-length sessions, it's also very token efficient. Lossless summaries are great for prompt caching. I at about a 90-94% cache hit rate. Thanks to incremental compaction, your context rarely grows beyond 80k tokens before truncating back to 30-40k or less. This means that your model is almost always operating faster, smarter and cheaper since it has less overall context to operate on.
9
6
158
10,992
Load more