@raven_protocol

Raven is developing a distributed network of compute nodes for Artificial Intelligence and Machine Learning. A decentralized computing protocol built on #BNB

 Hong Kong and Bangalore
ravenprotocol.com
Joined November 2017
  • Tweets4.6K
  • Following210
  • Followers12.8K
394Media
🏆Raven wins the @mozillabuilders Award at the Fix-the-Internet showcase. Recognition by the top open source organization in the world that Raven Protocol is a core infrastructure project. We demonstrated best traction and commitment to fix the internet! medium.com/ravenprotocol/rav…

Raven Protocol Wins the Mozilla Builder Award at the Fix-the-Internet Showcase

Receives grant funding to scale the decentralized compute network and speed up development of the open sourced distribution framework.

medium.com
57
57
16
188
1/ Last Week in Open Source AI & Distributed Compute - NVIDIA releases Nemotron 3.5 Super VL, a 120B multimodal model, alongside two open-weight Bolt edge models - llm-d 1.0 launches as a stable Kubernetes-native stack for production-scale LLM inference
1
156
more replies
3/ - DeepSeek open-sources DeepInfer, its production inference framework, under Apache 2.0, with support for NVIDIA SM100 GPUs and Huawei Ascend 950 NPUs - OpenCode discloses a breach exposing API keys belonging to 1.4% of Console users with external providers
1
29
4/ - Researchers demonstrate how a malicious AGENTS.md file can trick a coding agent into installing a rogue npm package - Block releases Dessa, a full-duplex speech model trained from human conversation that runs locally on high-end MacBook Pro hardware
16
GM, new week. Open-source AI keeps shipping. The real work starts when you move from accessing a model to making it your own.
1
2
186
1/ Open source models now handle 78% of token volume on major AI gateways. That shift didn't happen gradually. It happened because open models crossed a threshold, good enough capability at a fraction of the cost. For most workloads, the choice became obvious.
1
1
254
2/ Model access and model ownership are different things. Running an open model via an inference provider is access. Fine-tuning on your own data, training domain-specific variants, that's ownership. Most open source developers are still stuck at access.
1
22
3/ The bottleneck isn't the model anymore. It’s the training infrastructure to actually make it yours. Ravnest distributes training across consumer hardware you already own. No cloud invoice, no GPU queue, no datacenter access required. That's the gap it was built to fill.
18
1/ This chart tells you everything about where AI infrastructure costs are heading. US semiconductor producer prices surged +26% YoY in September: - 6th consecutive 25%+ YoY increase - Up 31% since July 2025 - Declined every year from 1994–2021 Historic territory.
The Kobeissi Letter
@KobeissiLetterOct 24
Semiconductor prices are skyrocketing in the US: Producer prices for semiconductor and other electronic component manufacturing surged +26% YoY in September, their 4th-largest increase on record. This also marks their 6th consecutive YoY increase of at least +25%. Since July 2025, the Producer Price Index for semiconductor and other electronic component manufacturing has risen +31%, to its 2nd-highest level since January 2007. To put this into perspective, the index declined every year from 1994 to 2021. Meanwhile, producer prices for electronic components and accessories soared +27% YoY last month, also marking their 6th consecutive YoY increase of at least +25%. The AI Revolution is driving historic technology inflation.
1
1
232
2/ AI demand is putting pressure on semiconductor supply. Expanding production capacity takes time & the cost of new hardware remains a challenge for teams building & training AI models. As compute demand grows, relying on ever more expensive hardware becomes harder to sustain.
1
15
3/ Training on consumer hardware that's already paid for isn't just practical anymore. As new hardware costs keep rising, the economics of distributed training across existing nodes get stronger, not weaker. That's the environment Ravnest was built for.
11
gm, the open source AI stack keeps shipping. good day to be building on it.
2
3
256
1/ Building a custom training pipeline on distributed infra usually means one thing: your research adapts to the framework, not the other way around. Custom loss functions, non-standard architectures, experimental optimizers ,most distributed frameworks treat them as edge cases.
1
214
2/ The result: researchers spend more time working around the distributed layer than running experiments. Every unconventional training approach means new workarounds. The infrastructure that should be invisible becomes the biggest obstacle.
1
17
3/ Ravnest’s extensible trainer API is designed to get out of the way. Bring your own training logic. Ravnest wraps the distributed layer: matchmaking, async pipeline, fault tolerance, gradient compression; around it automatically. Your architecture. Distributed by default.
15
Reflection AI just released Beam: 501B parameters, 23B active, Apache 2.0. 80% on SWE-bench Verified. Western frontier open-weight model. Models at this scale arent single-node workloads, distributed training infrastructure is what makes them accessible. thursdai.news/releases/2026-…

October 2026 AI Releases: OpenAI Dots, Claude Sonnet 5.5 & 81 more — ThursdAI

What launched in AI in October 2026? OpenAI Dots, Claude Sonnet 5.5 and 81 more releases, tracked live on the ThursdAI podcast — links, key numbers, analysis.

thursdai.news
2
204
1/ GPU spot prices doubled in six months, $4.40 to $8.08 per card-hour. Token prices fell 377x over 18 months. - 60%+ of 2027 datacenter capacity not yet under construction - GPUs sitting idle waiting for power The gap between compute costs & token revenue is compressing fast.
Tomasz Tunguz
@ttunguzOct 21
Yesterday brought two data points about the health of inference businesses : SpaceX's third-quarter report & DeepSeek's inference post. Overall, they demonstrate that inference businesses can be quite healthy even at current prices & efficiencies.

The GPU Revenue Ladder

Article · @ttunguz
1
3
228
2/ The inference business works at scale, but only if throughput gains outrun token price cuts. For most teams, training on distributed consumer hardware that’s already paid for sidesteps that math entirely. That’s the environment Ravnest was built for.
1
13
1/ Every major frontier open source model in 2026 is a sparse MoE architecture, DeepSeek V4.1, GLM-5.3, StepFun Step 5, Kimi K3. Training them across distributed nodes is a fundamentally different problem than training a model. Most frameworks werent built with this in mind.🧵
2
5
278
more replies
3/ Get the model split wrong on a MoE and you create: - Memory imbalances across nodes - Unpredictable communication spikes - Expert bottlenecks that compound across layers Standard model splitting algorithms weren’t designed for this.
1
1
16
4/ Ravnest uses model splitting built for LLM and MoE architectures. Layer distribution optimized upfront. Memory balanced across nodes. Frequently communicating layers grouped to minimize cross-node traffic. Built for the models open source developers are training in 2026.
1
17
Load more