You can now run a DeepSeek-V4.1-Flash locally 12GB GPU and Offloads consumer hardware. - GSQ-RCO quantizations with Strata inference engine. - RTX 5070 with 64GB/ 105-117 tok/s - Prefill is much faster than decode not bad - More RAM to faster decode - engine uses layer-major chunked reading, - for fun is a great.

Qwen3.8-27B reportedly went from 82 to 381 tokens/sec on a single RTX 3090 with 24GB of VRAM. That is a 4.6× increase in throughput without changing the GPU or the model. Developer TeksEdge achieved this by changing how inference is performed, combining speculative decoding, DFlash2, context lookup and optimized verification. Here is the reported progression: Baseline: 82 tokens/sec Optimized MTP: 114 tokens/sec DFlash2 + context lookup: 138 tokens/sec Optimized lookup verification: 381 tokens/sec The interesting part is that the final improvement did not come from a bigger GPU or a larger model. It came from reducing work the inference engine did not need to perform sequentially or repeatedly. 1. Speculative decoding reduces the cost of generating tokens Most language models generate output autoregressively. Each new token depends on the tokens that came before it, which makes ordinary decoding inherently sequential. Speculative decoding changes the process by using a faster draft mechanism to propose multiple tokens at once. The main model verifies the proposals and can accept several tokens in a single verification step. When the draft predictions are accepted, the system advances through more of the output without repeating the usual sequential decoding process for every token. DFlash2 is designed to improve this speculative decoding workflow. However, the speedup depends on how well the draft predictions match the main model’s output, how many tokens are accepted, and how much computation verification requires. 2. Context lookup takes advantage of text the model has already seen Some generation tasks involve substantial overlap between the input and the expected output. For example, reproducing a document, copying structured text or transforming code may require the model to emit sequences that already exist in its context. Context lookup searches for these reusable sequences and proposes them as candidate output instead of requiring the model to generate every token from scratch. TeksEdge reports processing blocks of up to 16 tokens and accepting roughly 15 tokens per block in its setup. On a 25,000-token document reproduction task, reported throughput increased from approximately 260 to 382 tokens/sec. This is a particularly favorable workload for context lookup because the output contains substantial material that can be reused. Open-ended conversation is different. When the model needs to generate new explanations, ideas or unfamiliar wording, there may be far fewer matching sequences available. That is why the same optimization can produce very different results across workloads. 3. Faster verification helps unlock the remaining performance Finding candidate tokens is only part of the problem. The inference engine must still verify them efficiently. If verification is too expensive, the overhead can erase much of the benefit from speculative decoding and context lookup. TeksEdge’s reported progression from 138 to 381 tokens/sec after optimizing lookup verification suggests that verification efficiency was a major bottleneck in this particular configuration. The result highlights an important engineering principle: improving a draft mechanism is not enough if the rest of the inference pipeline cannot process its proposals efficiently. The reported stages should be viewed as successive configurations, not independent speedups that can simply be multiplied together. 4. Prefix caching solves a different problem: waiting for the first token Fast decoding does not automatically mean fast responses. Before generating an answer, a model must process its input. This stage is called prefill, and it can be expensive when prompts contain long instructions, tool definitions or repository context. Coding agents often reuse the same prefix across multiple requests. Without caching, the engine may repeatedly perform much of the same computation.

Cardano uses Proof of Stake, not Proof of Participation. The final SPO NO ratios speak for themselves: By CIP-1694 standards (including implicit NO): ~101:1 Even by explicitly voting stake: 9.06:1 Only 0.98% of eligible SPO stake explicitly supported the proposal. No matter how you look at it, the sentiment is clear.

The Cardano k=1000 governance Info Action has concluded, and the outcome could hardly be clearer. An overwhelming NO, regardless of which voting metric you look at. What particularly impresses me: 1.91 billion ADA in stake explicitly voted NO by SPOs, even though they didn't need to vote for the proposal to fail. SPOs aren't too lazy to participate in governance. They vote when it matters to them! A big thank you to all SPOs and DReps who participated, regardless of how you voted. This is governance in action! 💪

First time getting the entire cut out in one piece! It's also my first shaded pumpkin carving.

Intersect Budget Administration Services Update 📊 On October 20 and October 21 Intersect executed the SweepTreasury function on its 2025 Treasury Reserve Smart Contract (TRSC) for the first time on Mainnet, returning 115,172 ada to the Cardano treasury from the 2025 Cardano Product Committee budget. Tx 1 - ₳55,172: cexplorer.io/tx/3a1433f2c412… Tx 2 - ₳60,000: cexplorer.io/tx/81924bac490d… Original proposal: gov.tools/outcomes/governanc…

How does post-quantum cryptography fit into Cardano’s roadmap for 2030 and beyond? Join us on Wednesday November 4th for a TSC discussion with Input Output Research (IOR) researchers. All welcome - please share in your networks! luma.com/intersect-af5p

A year ago, I was scared LLMs would never be as good at bleeding edges than they were at industry standards It turns out it was the opposite: The cost of using Rust, Nix, Formal proofs, etc is converging to 0 while the pros are only improving I'm glad I was so wrong about this

“Don't feel bad if people remember you only when they need you. Feel privileged that you are like a candle that comes to their mind when there is darkness." St. Nikolai Velimirovic

Thanks for your work Chris! Sad to see you retiring as a DRep! To delegators, if you wanna redelegate, please checkout drep.me/martinlang To take a look at my past voting decisions, thx!

Good evening. Daedalus 11.6.0 is out. This mainnet maintenance release brings blockchain recovery tools, reliability improvements and dependency security updates. We recommend all Daedalus users upgrade. Here’s what’s new. 1/4🧵

Greek Orthodox Church in Jordan attacked by terrorists. 7 killed, 34 injured. Jordan's King Abdullah visited Karak hours after the attack and said the perpetrators would be punished "with an iron fist". May it be so, and kyrie eleison. 🥺🙏🕯☦️

See donations to treasury on cexplorer.io/treasury/donati…

CIP-23 (minPoolMargin) は、近日実施予定の Dijkstra ハードフォークで完全に実装されます！

カルダノ待望の"Leios"メインネットまでの日本語解説です！ これは必見です👇