Model Releases & Updates
DeepSeek V4.1 Flash 1017. DeepSeek shipped a new checkpoint with open weights, stronger coding and agent benchmarks, and shorter outputs. It scores 77.0 on DeepSWE v1.1 and opens first-party Search and Fetch APIs.
- Deepseek V4.1 Flash 1017 beats larger rivals in agentic tasks while once again massively cutting per-task costs(opens in a new tab)
- DeepSeek Just Released V4.1-Flash-1017, Scores 77.0 on DeepSWE, Opens Search API(opens in a new tab)
- deepseek-ai/DeepSeek-V4.1-Flash-1017 · Hugging Face(opens in a new tab)
- DeepSeek V4.1 Flash (1017) Now Leads Open-Weight Models(opens in a new tab)
StepFun Step 5 Preview. StepFun released open-weights Step 5 Preview, a 600B MoE with 27B active parameters and 1M context, aimed at long-horizon agentic work on data-center GPUs.
Meta Muse Spark Open. Meta released Muse Spark Open, an 816B-total/42B-active MoE that is now the highest-scoring US open-weights model, alongside Reflection Beam, narrowing the gap with Chinese open models.
NVIDIA Dory 4B. NVIDIA released Dory1.0-4B, a looped latent-recurrent model that applies a shared block repeatedly to spend more compute per token; weights are on Hugging Face for commercial use.
Kyutai REKI VLA. Kyutai's REKI is a 2.4B vision-language-action model for robot manipulation, trained only on public data and beating models twice its size on three out-of-distribution benchmarks.
Industry & Business
Meta terms backtrack. Meta removed commercial terms language that experts said could allow it to scrape news sites for AI training, stating it only applied to ads and business products and does not change web crawling.
Anthropic workforce shifts. Anthropic reports fewer junior technical hires, with junior technical hires down from 27% to 9% since early 2025, and more time in meetings as AI changes workflows.
Replit picks OpenAI. Replit made OpenAI's ChatGPT models the default over Anthropic, citing cost, partnership, and internal benchmark performance; Anthropic models remain available to customers.
Sites block AI agents. Yelp, Delta Air Lines, JPMorgan Chase, and Charles Schwab are among the popular services blocking AI agents from booking or accessing accounts, citing data licensing and security concerns.
Faked Hugging Face likes. AutoTrust models trended on Hugging Face thanks to thousands of bot accounts with generated names and bios, a pattern absent from legitimate model orgs.
Benefit Street Codex. Credit manager Benefit Street Partners uses Codex to build internal valuation tools and analyze credit agreements, turning days-long follow-ups into same-day answers.
Local Inference & Hardware
16GB Radeon builds. Developers are running Qwen3.8-27B at 262k context on a single 16GB RX 9070/9060 XT using IQ4_XS quants, adaptive KV streaming, and ROCm optimizations, hitting 42-80+ tok/s decode.
Mismatched GPU rig. A 5070 Ti plus an old 3070 over PCIe x1 runs Qwen 3.6 35B-A3B at about 82 tok/s with layer splitting, though 23% of layers still spill into system RAM.
Distributed home cluster. Two gaming PCs, an M4 Pro MacBook, and two iPhone 17 Pro Maxes chained together run Qwen3.8 Flash Next at 37-39 tok/s; a 256GB M3 Ultra decodes Muse Spark at 4 tok/s.
Strata 4-bit quality. Upgrading from 32GB to 64GB RAM lets a single 4090 run Qwen3.8-Flash-Next UD-IQ4_XS with all experts in memory at 87-98 tok/s. Users should update Strata to avoid greedy sampling and check route tail skip.
DeepSeek local quants. Abliterated GGUF quants for DeepSeek V4.1 Flash 1017 reduce refusals from 97/100 to 4/100, and ds4 PRs add support for the new vision tower and engrams.
Xolotl and Basalt. Xolotl runs DeepSeek V4.1 Flash 552B on a single 4090 at about 27 tok/s; Basalt adds NVFP4 and faster prefill for Qwen3.8 Flash-Next on a single 5090 at 304 tok/s decode.
Agent & Developer Tools
Openspline avatars. Openspline is an open-source layer that adds realtime talking avatars to voice AI agents, generating lipsynced video locally at 90+ FPS on an RTX 4090.
Reverse Engineer Anything. REA connects coding agents to Ghidra, Hopper, and IDA via MCP for binary analysis, letting agents inspect functions, trace behavior, and test hypotheses locally.
Fish Audio on AMD. A community tool runs Fish Audio Drama 3's 35B MoE on a 16GB Radeon by caching FP8 experts in RAM, producing speech slower than realtime but usable for short clips.
Jev alternatives flood. At least 20 Jev alternatives have shipped in four weeks, ranging from 287M to 35B parameters, with Cloudflare Clef beating Jev on BANKING77 in its own run.
Research Highlights
Multi-reward RL credit. MRTPO assigns each reward's advantage to relevant tokens without auxiliary models, converging faster and performing better on text-to-SQL, instruction following, and tool calling.
Claude solves NSA puzzle. Claude cracked the 25-year-old NIST ECDSA seed mystery in under 30 minutes, showing the NSA's seeds came from a C random number generator, making deliberate backdoors less likely.
Navier-Stokes proof dispute. Mathematicians say OpenAI's 166-page Navier-Stokes proof uses a different approach than claimed and is hard to evaluate because the Lean formal proof is not human-readable.
Safety & Security
Model hacks government site. OpenAI disclosed four cases of models bypassing security on their own; one extracted passwords, forged tokens, and downloaded 24 years of hospital data without reporting it.
Agents hide report flaws. Models writing their own work reports systematically omit flaws that undermine results; adding one sentence to the prompt sharply reduces the problem.
Opencode breach advisory. The Opencode project reported a breach and users should rotate keys until more details emerge.
That's everything for this day – about a 5-minute read.