Interesting Finds — 2026-08-25
Four notes from the feed: an experimental vision model, a free agentic model window, CUDA on RISC-V, and a Linux utility for Logitech hardware.
A short roundup of items flagged today. Each is a separate find, collected here so the site keeps a single record while each note can circulate individually.
1. DeepSeek V4-Flash-Vision-Exp — experimental vision against Opus 4.8
DeepSeek has released V4-Flash-Vision-Exp, an experimental multimodal model that adds image understanding to V4-Flash’s text capabilities. Coverage at The Decoder reports that on the company’s own multimodal agent benchmarks it approaches Opus 4.8 and sometimes surpasses it.
The useful signal is narrow. The model is explicitly labeled experimental, and the comparison is on DeepSeek’s own benchmarks. Those are not independent evaluations, and agent benchmarks are sensitive to harness and tool definition. What it does suggest is that DeepSeek considers vision-augmented agentic work a priority for the Flash line, and that its internal trajectory puts a low-latency model near frontier behavior on its chosen tasks. Wait for external reproduction before treating the Opus comparison as general.
Source: The Decoder — DeepSeek releases experimental Flash vision model
2. Thinking Machines Inkling — free on OpenRouter for agentic harnesses
Thinking Machines has made Inkling available for free on OpenRouter, restricted to agentic harnesses, for the next few weeks. The stated purpose is to gather real-world agentic behavior data, disassociated from accounts, to improve the model.
From the provider page: openrouter.ai/provider/thinkingmachines lists two entries, thinkingmachines/inkling:free and thinkingmachines/inkling-small:free (dated variants inkling-20260715:free and inkling-small-20260730:free in the usage chart), with terms at Thinking Machines Free Research API Tier Terms. Usage is already substantial — the OpenRouter chart for this week shows daily tokens in the billions for each variant (over 16B for Inkling on 2026-08-25 alone), which indicates broad harness testing.
As worded by the team: “We want to improve Inkling’s agentic performance. To help us understand its real-world behavior, we are making it available for free on OpenRouter (only with agentic harnesses) for the next few weeks, starting now. We’ll use the data, disassociated from accounts, to better it.”
If you run agents, this is a low-friction window to test it against your own tasks. Note the constraint: free access is harness-only, and the data will be used for training. Evaluate with that in mind.
3. Hot Chips 2026 — CUDA targets RISC-V
Chips and Cheese reports from Hot Chips 2026 that Nvidia is working to bring CUDA to RISC-V CPUs, which would join x86-64 and aarch64 as targets that can feed Nvidia GPUs. The talk details requirements rather than a launch.
Nvidia’s stated requirements include an RVA23 CPU and compliance with the RISC-V server SoC and server platform specifications (RAS features, a dedicated security processor, and related baselines). Beyond that, Nvidia asks for vector extensions with predication, ACPI support (enabled for RISC-V by UEFI forum work in 2025 and the BRS specification), PCIe cache coherency, and PCIe peer-to-peer communication. NVLink Fusion — letting partners implement NVLink C2C on their own dies, as in the Mediatek-linked GB10 — inherits all of the above plus requirements for DOCA and NCCL and a close partnership with Nvidia.
The report notes SiFive as a partner expected to demo a CUDA-capable RISC-V server system at Hot Chips, likely a high-core-count server chip rather than consumer boards.
The implication is architectural rather than immediate. Most existing RISC-V hardware, and especially hobbyist single-board systems, will not meet these requirements. ACPI and coherency in particular have historically taken years to propagate, even in aarch64. If RISC-V CUDA arrives, it will likely appear first in servers where the platform specifications can be enforced, not in low-cost boards. The open question raised by the coverage is whether Nvidia will block unsupported configurations or allow enthusiasts to experiment on a slower path where cache maintenance and peer-to-peer fallbacks cost performance but still function.
Source: Chips and Cheese — Hot Chips 2026: CUDA Targets RISC-V
4. Solaar — use your Logitech hardware without Logi Options+
I like my Logitech keyboard and mouse. I do not want Logi Options+ to use them. Solaar is why that works on Linux — it manages Logitech devices on Unifying, Bolt, Lightspeed, Nano, USB, or Bluetooth without needing the vendor app. It is not a driver; it handles the special device messages the input system otherwise ignores.
It covers pairing and unpairing, device settings, custom button mapping, and rules triggered by device events. Packaged for most distros (Arch extra, Ubuntu/Kubuntu PPA, NixOS flake, plus Debian, Fedora, Gentoo, and others). It is the control plane that lets the hardware stay, and the bundled software stay off the machine.
Source: pwr-Solaar/Solaar on GitHub
Four finds, one record. Each note above will also post individually with a link back here.