Flash
What changed, and what is the evidence.
ActuationEnergyFundingHumanoids
🗞️ Pulse first · 3 news posts
Coverage
- 🗓️ 2026-10-22 · ✍️ RoboterminalEka's Robot Picks 1,500 Things an Hour. Its Gripper Is Rated for About 139 Days of That.
- 🗓️ 2026-10-21 · ✍️ RoboterminalSchaeffler Will Buy Thousands of Humanoids From Its Actuator Customers. Its €400 Million Order Book Doesn't Count Those Deals.
- 🗓️ 2026-10-19 · ✍️ RoboterminalFifteen Humanoids Are Loading Car Parts at 98%. The Number That Decides Their Future Is 15 Seconds.
Related topics
Timeline
What changed, and what is the evidence.
Oct 22, 2026 · Post · posted
| Headline |
|---|
95.4% accuracy and $5.63 per 1,000 correct tool calls sits right next to 95.4% accuracy and $1.50 per 1,000 correct tool calls on the same benchmark. Same task, same pass rate, 3.75x the bill. That's Claude Opus 5.5 and Grok 4.7 on Tirtha… |
I swear, we do this every 3 months. "Chinese models are falling hopelessly behind. They don't have enough compute". And then they release their new models "Oh, they are caught up. Need to ban them". Why not just recognize when flagship… |
China's AI stocks are getting HIT in Hong Kong today. The Hang Seng Tech Index fell about 4% to around 3,834, while the Hang Seng Index dropped almost 3%. According to Chinese media, AI model makers MiniMax and Zhipu fell more than 4%,… |
DeepSeek did not raise roughly $15 billion, about 100 billion yuan from CATL and Tencent, to build a better chatbot. That round is one of the largest in Chinese tech history, Bloomberg is calling it a direct challenge to the leading US… |
Huawei Ascend 950PR Deep dive & µbenchmarks, DSv4.1 Flash Analysis, DeepSeek support for Ascend, AICPU, Tilelang Ascend https://t.co/3j3AvY17wC |
DeepSeek just gave away the engine it uses to run its own models. The Chinese lab open sourced DeepInfer, the inference framework behind DeepSeek V4.1 Flash, under the free Apache 2.0 license. It runs on $NVDA Blackwell chips and on… |
Model weights are free. Running them at scale cheaply is not, and that's the part DeepSeek open sourced for everyone this week. DeepInfer is the inference engine serving DeepSeek V4.1 Flash in production right now, and it's fully open,… |
ENGINEERING NEWS ⚡️ Tesla has built a lightning-fast battery switch that kills short circuits without a fuse 🎻 A severe crash can crumple parts of an electric vehicle's high-voltage system, and if that damage shorts out the battery,… |
ENGINEERING NEWS 🧠 @neuralink has taught its brain chip a clever trick to tell a firing neuron from a chewing jaw 🦷 Chew a bite of your sandwich, and a brain implant might think a neuron just fired. Neuralink's patent application teaches… |
Microsoft showing off the ability to run DeepSeek V4 Flash locally on a Surface laptop. “It’s possible to bring more of the frontier to your PC.” https://t.co/1z7P7pyRol |
One of the hardest parts of the memory trade is picking the stock. Forecasts have global memory chip revenue going from $200b in 2025 to over $700B in 2027. That's more than triple in two years. The demand side is easy to see. Every AI… |
Rare environmental edge cases like localized flash flooding test perception limits. Edge computing stacks must recognize environmental anomalies and fail safely to designated safe pull-over locations rather than stranding assets in active… |
ENGINEERING RECAP 🧠 Five Tesla patents, one mission: fit bigger AI onto the chips already in its cars and robots 🔥 Read five new Tesla patent applications side by side and a single story jumps out. It's a story about a squeeze. The AI… |
SMART ENGINEERING II 🧠 Tesla has done AI's toughest math the way you turn cents into dollars: just move the point ⚡ Shrink an AI's math down to cheap 8-bit numbers, and one step still drags it back to slow, power-hungry floating point:… |
SMART ENGINEERING 🧠 Tesla has figured out a "quantized" way to run AI attention on low-power 8-bit chips while holding on to near 16-bit accuracy: a fresh ruler for every token, tile by tile ⚡ Attention is how a transformer, the AI… |
If you're a robotics researcher working with GPT 6 Astra to do interesting things with robotic control or agentic real2sim, and would like access to open source models (Kimi K3, Qwen 3.8 Flash Next) as a comparison, please DM me! I'd love… |
TL;DR ⏳ Now you see it, now you don't 🧙 Tesla's new patent (US 2026/0285418 A1) gives the Roadster a wing that vanishes into the bodywork, then erupts into 700 kg of race-car downforce on command 🔥 🫥 Invisible on the street: The wing… |
A Century of Innovation (1926–2026) 2026 — 🇺🇸 Agentic AI: Systems that don't just think, but act autonomously. 2026 — 🇨🇳 Humanoid robots: Humanoid robots move from research labs into workplaces and real-world deployments. 2025 — 🇺🇸… |
The pattern I keep seeing this month: once a model layer commoditizes, the fight moves one layer down to whoever serves it fastest. RunAnywhere, a YC W26 company, launched Wally this week, an inference stack for open frontier models that… |
My read on Xiaomi shipping MiMo-V2.6-Pro this week: the real story isn't the benchmark score. It's who is now capable of putting up that score. Xiaomi, a phone and appliance company, published MiMo-V2.6-Pro and Flash on Hugging Face… |
Another Chinese lab just gave away a trillion parameter model for free, and the market barely blinked because it's the fourth time this year. Xiaomi open sourced MiMo-V2.6 today: two natively omnimodal models, Pro and Flash, both MIT… |
a few notable open model releases *since* this interview of @dylan522p by the way :) Thinking Machines - Inkling - Jul 15 Moonshot - Kimi K3 - Jul 16, weights Jul 26/27 inclusionAI - Ling 3.0 Flash - Jul 23 DeepSeek - V4-Flash-0731 - Jul… |
Preliminary September 2026 semiconductor export statistics (covering Sept 1–20): Total Memory Export Value: $21.501 billion (+350.7% YoY, +8.2% MoM) Unit Price: $82,467/kg (+277.3% YoY, -13.9% MoM) DRAM Export Value: $10.644 billion… |
Frontier agentic performance just hit $1 per million. StepFun launched Step 5 Preview today. 600B total parameters, 27B active per token. 1M context. $1 per million input tokens. That is 10 to 20x cheaper than GPT-6 Astra at frontier… |
1,424 tok/s prefill on consumer hardware. Halogen is a dedicated inference engine for Qwen3.8-Flash-Next on Strix Halo. 4x faster than previous implementations at 32K context. Local AI performance is now a software optimization race. The… |
Open models just took 78% of all tokens. Here's what the data actually shows: 1/5 - The Volume Flip Vercel's AI Gateway hit 78.4% open-model token share on September 18. In June it was roughly 40%. That's a near-doubling in three months.… |
K.S. Pua in his recent interview with Storm Media @stormmedia_jp > While logic chips (GPUs and CPUs) receive primary attention, the two largest physical constraints behind AI infrastructure are electrical power and memory (DRAM and Flash).… |
Qwen3.8-Omni-Flash is a different bet. Not "better language model". Omni-modal from the ground up. One model that handles: → Text, image, audio, video in → Text and audio out → Real-time inference at flash speeds The architecture isn't a… |
A new agentic open-weights model drops Tuesday. 3x cheaper than DeepSeek Flash. 15% better on long agentic loops. Available on RouteLLM API and on the coding agent. Designed for 90% of coding tasks. My read on what's happening in the… |
Engrams Embedding Entendre: Codesign for Efficient DRAM/SSD Offloading New Model Architecture Implications for TAM of DRAM/NVMe, DeepSeek V4.1 Flash, AgentX, InferenceX, NVMe experiments https://t.co/DuIgPg9mOa |
Zai is the most transparent & vocal of Chinese AI labs in their path toward RSI & autonomous AI. They bought a team a while back to help integrate domestic chips & this + 100k cluster & GLM-5.3 powered Infra Agent did the E2E optimization… |
How we read the livestream metrics of @XiaomiMiMo post-training: 🟠 MiMo is training 2 models: MiMo V2.6 Pro (1T total, 42B active) and MiMo V2.6 Flash (310B total, 15B active) 🟠 The metrics tab contains different metrics broken down by… |
Serving KV cache entirely from high-bandwidth flash beats hybrid designs: HBFlex reports up to 1.58x throughput over FlashAccel and 3.30x over H3 in trace-driven simulation. Keeping HBM around for KV management costs HBF bandwidth under a… |
Xiaomi and Zai are pushing RL really hard now. They seem to believe there’s a lot more to be gained from scaling on this axis. Crazy to see Xiaomi livestream their current RL run for MiMo-V2.6 pro and flash. https://t.co/0NsZzyn8nZ… |
98% of GPT-6 Astra's benchmark performance. 1.4% of the cost. That's what OpenDesign found when they benchmarked DeepSeek V4.1 Flash on everyday design tasks. Every other model in the comparison scored lower AND cost more. The model… |
OpenRouter just published a data point I keep coming back to: for the first time in 2.5 years, their users spent more on OpenAI models than Anthropic last week. Anthropic's share on OpenRouter has fallen from 25% to 12% this year.… |
When trying to forecast where LLMs are going in China, it's important to hear what supply chain is saying. Today, HW said that 10T param models will be baseline for LLMs in near future. As such, 100k+ card clusters formed by supernode will… |
BYD unveiled ET44 HDT + other commercial vehicles @ IAA 2026 in Hanover, Germany. 44t max wt, 651 kWh battery pack, 851.2V, 1000ps peak power & 600 km range @ 40t load + unveiled 1.5MW flash charging solution for trucks that provides 400… |
XPeng to build its 1st Flash Charging station X-Energy in Hong Kong. 1MW peak charging speed, 1000A peak flow Just needs 240 kW grid connection Comes w/ 400 kWh battery pack. Building a flash charging network takes huge BESS supply. Why… |
The inference layer just got a native desktop. Cline shipped Cline Desktop this week: a standalone app for open-weight models, not just a VS Code extension. ClinePass at $9.99/month gets you DeepSeek V4, GLM 5.3, Kimi K3, Qwen3.7, and a… |
According to the Korea Semiconductor Industry Association, the technology gap between South Korea and China in memory chips has narrowed to 3 years in HBM, 2 years in DRAM and 1 year in NAND flash. A gap once estimated at more than five… |
Coding agents just split into two permanent tiers. Cline Desktop shipped today: native macOS interface for working with open weights models. ClinePass provides access to DeepSeek-V4.1-Flash, Musespark-1.3, and others at no API cost. BYOK… |
DeepSeek is now pricing inference by time of day: $0.30/$1.20 per million input/output tokens at peak, $0.15/$0.60 off-peak. From Sept 14, deepseek-v4-pro requests route to V4.1-Flash at Flash rates until V4.1-Pro ships. Token pricing is… |
DeepSeek says V4.1-Flash needs one-quarter the HBM and one-eighth the SSD of its previous architecture for the KV cache. Not the model, not training: KV cache only. Micron booked $13.77B in cloud memory at 83% gross margin last quarter.… |
Run Qwen3.8 Flash Next, a 125B multimodal MoE model, locally on NVIDIA Jetson AGX Thor. ⚡ Built for reasoning, coding, and tool use, its 262K context window can handle large codebases and long conversations. With NVFP4 and MTP, it reaches… |
BYD putting Denza N8L Flash Charging EV on sale at 299.8 to 329.8k RMB. This is even cheaper than the previous PHEV versions which sold 4567 unit in Aug. 130 kWh battery pack & 960 km range on 800V architecture for 6-seat version. 0.252… |
The efficiency trajectory in open-weight models just made the cost-performance chart look different. DeepSeek V4.1 Flash: 552B total parameters, 8B active via MoE routing. Beats a 1.6T model at 4x lower inference cost. $0.30/M input… |
POWER OF CUDA MOAT ALERT🚨: 2 days after CUDA vLLM supported DeepSeekv4.1 Flash, AMD finally publicly released its DeepSeek v4.1 Flash image. Functionally, it works out of the box, but performance-wise, it is currently up to 14.8x worse… |
The benchmark race is converging. DeepSeek V4.1 Flash, Claude Opus 5, GPT-5.6 Sol — all within 1-2 points on agentic evals. The differentiation is moving to workflow tooling: harness design, caching strategy, context management, cost per… |
DeepSeek V4.1 Flash: 74.2 on DeepSWE v1.1. Beats Claude Opus 5 (74.0) and GPT-5.6 Sol (73.0). $0.003/M cached off-peak, $0.60/M output. 30x lower recorded agent cost vs frontier. Agents read the same repo files repeatedly. Cached pricing… |
On the Day 0 release of DeepSeekv4.1 Flash, NVIDIA vLLM works out of the box with zero issues across all 6 SKUs: H100, H200, B200, B300, GB200, GB300! Amazing work by the NVIDIA & Inferact teams! In comparison, AMD vLLM still does not work… |
DeepSeek V4.1-Flash: 552B backbone, 8B active per token at prefill, 16B at decode. Global KV cache is 890 bytes per token, about a quarter of V4-Flash and 1/437 of V1. Context to 1M tokens, trained on 45T. The cost curve for long-context… |
Every time I think we've found the cost floor for inference, DeepSeek lowers it again. DeepSeek V4.1 Flash drops today: 552B parameters, 1M context window, native multimodal. And the pricing is legitimately wild: $0.003/1M input tokens,… |
DeepSeek V4.1 Flash is here, and DeepSeek calls it "the smallest model in our new architecture family," with a V4.1-Pro still to come. It beats GPT-5.6 Sol and Claude Opus 5 on several agentic, coding, and cyber benchmarks. Flash is a… |
DeepSeek just launched V4.1-Flash, the smallest model in their new architecture family. But the architecture change underneath it is the story. The Causal Encoder-Decoder split: 8B active parameters for encoding (reading input), 16B active… |
Congrats to @deepseek_ai on releasing DeepSeek-V4.1-Flash! > 552B backbone, with a causal encoder-decoder activating just 8B params at prefill, 16B at decode > 196B Engram memory accessed through sparse lookups > ~8x less persistent KV… |
The new DeepSeek V4.1 Flash model is mindblowing - back on top of the open-source model leaderboard and extremely cheap. It has a lot of very smart ways to be efficient and highly capable so I made a video of the forward pass to give you a… |
As always with DeepSeek, the new model V4.1 Flash have a lot of very smart ways to be efficient and highly capable Here is a detailed video of the forward pass to give you a view of what going on inside the model during inference… |
don't get distracted by all the hedging words in its name ("flash", minor version): seems like DeepSeek V4.1 Flash is a major update weights at: https://t.co/SzXY8BGzIa paper at: https://t.co/eQmBJAU4iM https://t.co/6LGYgnLa77 |
don't get distracted by all the hedging words in its name ("flash", minor version): seems like DeepSeek V4.1 Flash is a major update https://t.co/6LGYgnLa77 |
Analyst Mehdi Hosseini at Susquehanna on memory prices: "Average DRAM price will rise 50% in the third quarter of this year, followed by an additional 20% in the fourth quarter. He expected NAND flash to see an even larger increase. The… |
TontaubeV1: a 2.9B-parameter open-weight streaming TTS stack reaches 200ms to first audio on a single RTX 5090, real-time factor 0.08 single and 0.02 across eight streams. Authors claim prosody parity with ElevenLabs Flash v2.5. Voice… |
DeepSeek will launch V4.1 Flash around 9/10 w/ every aspect/score above V4 Pro. As such, Flash will completely replace Pro on DS platform b4 V4.1 Pro comes online. Quite interesting, since in my limited 4.1 Flash testing so far, it's been… |
The most revealing number from today's announcement isn't the solved problem. It's the architecture. 10,000 concurrent agents. 88 hours. 2.7 million messages exchanged. 130 billion output tokens consumed. An unreleased model described as… |
Open weights are not the same as cheap serving. Alibaba's Qwen 3.8-Max, 2.4T parameters with 95B active, is estimated to need 48-64 B200-class GPUs to serve. DeepSeek's V4 Flash, 284B, fits in about 142GB at FP4. Downloadable and… |
Micron is testing Near-GPU NAND: high-endurance flash on the GPU package or PCB, hundreds of GB between HBM and system storage. Density traded for endurance and I/O. It follows SK Hynix and SanDisk's HBF. Memory-bound LLMs would need fewer… |
The Gemini 4 internal benchmarks leaked three weeks ago showing it outperforming GPT-5.6 Sol and Fable 5 on coding and math. Then this week Google shipped Gemini 3.8 Flash, their third Flash variant in six weeks, at $0.75 per million input… |
Qwen3.8-Flash-Next just made the 180B parameter class meaningfully accessible. The architecture is worth understanding: 125B language model parameters, 55B auxiliary, but only 6B activate per token. That sparse activation pattern is what… |
The image model race just made a clear separation between quality and price leadership. My read on what MAI-Image-2.6-Flash means: Microsoft now owns both. MAI-Image-2.6 is #2 on the Artificial Analysis image leaderboard, with a +79 Elo… |
This is what Tesla’s Robotaxi app looks like when your Cybercab arrives. The screen glows to match whatever the color the RGB front light bar is on the Cybercab, while an arrow points to where the vehicle is located. You can also flash the… |
Gemini 3.8 Flash Puts Google's Rapid AI Release Cycle to the Test Google has released Gemini 3.8 Flash, its third Flash release in six weeks and roughly three weeks after Gemini 3.7 Flash. The new model is generally available with a 1… |
The benchmark gap that keeps surprising me: on standard coding benchmarks, the top models cluster. On ultra-long horizon coding, the gap blows open. Proximal released FrontierSWE v2 today. 34 real software engineering tasks, evaluated over… |
The model tier split just happened. Gemini 3.8 Flash and Gemini 3.8 Flash Cyber launched today. Same day, completely different access regimes. The public Flash model is on the Gemini API at $0.75 per million input tokens. It scores 73.7%… |
The flash tier just got meaningfully better. Gemini 3.8 Flash lands today at the same price as 3.7 Flash: $0.75 per million input tokens, $3.75 per million output tokens. But with significant gains on software engineering, agentic tasks,… |
The question "which model is better for my use case" only has one honest answer: run your actual work across both and compare. Fable 5.1 shipped yesterday. Qwen3.8-Max-0902 shipped yesterday. Grok 4.7 arrives in 10 days. Gemini 3.8 Flash… |
Post Log in Sign up Post tphuang on X: "How it views its own models in terms of the kind of tasks them can do. GLM-5.3 in terms of long time horizon work is still short of 1 wk, need at least 1 month to be autonomous. Flash is the 1st one… |
The next memory supercycle is hiding inside humanoid robots (Save this). This chart estimates that each humanoid robot could contain $600–$800 or more in memory related and semiconductor components, including DRAM, HBM, NAND flash,… |
I’ve said before that I think Google & DeepMind are playing a longer game with models. This is still a firm stance. Because a lot of the time, i think people look at Google’s model strategy too narrowly; Every Gemini release immediately… |