The Archive

from Particle AI is now available on . It is post-trained on DeepSeek V4.1 Flash, with a focus on completing software engineering tasks using fewer tokens. Particle describes the gains as shorter reasoning, fewer rounds of trial and repair, and more focused exploration.Particle-1AI Gateway Use it for debugging, planning code changes, and breaking down multi-step engineering problems. Particle-1 has a 1M-token context window for working with code and extended task histories. Set in your environment. Use as the model name:AI_GATEWAY_API_KEYparticle/particle-1 To use it in Claude Code, Codex, Cu...

Independent third-party safety evaluation of DeepSeek V4.1 Flash-1017 for loss-of-control and manipulation risks following EU AI Code of Practice.

DeepSeek V4.1 Flash 1017, MiniMax M3.1 Flash, and Mercury 3 released; limited immediate adoption expected.

now includes DeepSeek V4.1 Flash, alongside Kimi K3 and GLM 5.3, each accessed through its preferred provider, with more models to come. The DeepSeek model uses the latest weights (1017).Token Pass Token Pass is a $30/month subscription on Pro teams that grants $80 in monthly credit and runs through . Use it with any coding agent such as , , , etc. or build your own with the AI SDK, with Zero Data Retention enabled by default.AI GatewayfxPiClineCommand Code Add DeepSeek V4.1 Flash from the in your team's AI Gateway dashboard, or read the for setup.Token Pass pageToken Pass documentation Read ...

is now available on .DeepSeek V4.1 Flash 1017AI Gateway DeepSeek V4.1 Flash 1017 is a minor update to that improves reasoning, coding, agentic tasks, and image understanding.DeepSeek V4.1 Flash The coding gains show up in terminal work and in building and editing whole repositories, and the agentic gains in multi-step automation. With images, the model reads charts and handles visual reasoning more accurately when it can call tools. DeepSeek V4.1 Flash 1017 also finishes tasks faster and with shorter outputs than the previous snapshot, so agent runs tend to take less time and use fewer output...

DeepSeek releases v4.1-Flash with 763B parameters, novel causal encoder-decoder architecture, and vision capabilities.

Qwen 3.8 27B scores 52 on the Artificial Analysis Intelligence Index That's the same score as GPT-5.6 Luna (max), and just one point behind GLM-5.2 (max) and DeepSeek V4 Pro 0813 (max) - that GLM is 753B and that DeepSeek is 1.6B parameters, and Luna is size unknown but presumably a whole lot bigger than 27B. Qwen 3.8 27B is a truly astonishing model . Via Hacker News Tags: ai , generative-ai , llms , qwen , ai-in-china , artificial-analysis

DeepSeek V4 Pro 0813 (on OpenRouter) The latest DeepSeek Pro model is now available, via API only. I had to link to OpenRouter because DeepSeek don't have any obvious announcement page for their new model. I haven't been able to confirm if they plan to release the open weights, but given the weights are available for both April's deepseek-ai/DeepSeek-V4-Pro and July's deepseek-ai/DeepSeek-V4-Flash-0731 it seems likely. Interestingly I got very different looking pelicans for the three different reasoning levels of low, medium, and high. I've not noticed this kind of difference from any other m...

DeepSeek V4-Flash 0731 released; minimal other activity.

Laguna S 2.1, a 118B MoE model from Poolside AI, achieves Deepseek v4 Pro performance at lower cost than v4 Flash.

Signal-coverage matrix stratifies LLM autoformalization errors (type vs. semantic) across ProofNet# and MiniF2F using DeepSeek V4-Pro.

DeepSeek V4 Pro pricing undercuts US models by 11–35x, challenging premium API pricing models of OpenAI and Anthropic.

Developer refactored 120-file FastAPI service using DeepSeek V4 and Hunyuan with 80x cost savings vs Opus; open-weight models matched Opus latency but introduced production bugs.

Three months ago I pressure-tested which LLMs would cave and help build the apocalypse. Claude was the only one that consistently said no. Since then I've tested 30 more models across 6 dystopia modules (Orwell, Huxley, Petrov, Basaglia, LaGuardia, Baudrillard). The gap between Anthropic and everyone else is getting *wider*, not smaller. New results: * Grok 4.3: Will happily design citizen scoring systems if you ask nicely twice * GPT-5.5: More capable, still compliant when pushed * Gemini 3.1 Pro: Talks about safety while writing the surveillance code * DeepSeek V4: "How many warheads did...

Empirical testing of DeepSeek V4's 1M context window on real codebases (45k–520k tokens) shows sustained recall under 300k but precision degradation at larger spans.

User reports optimized inference performance for DeepSeek V4 Pro using ktransformers on AMD EPYC + RTX PRO hardware with tuned NUMA settings.

User successfully quantized and ran DeepSeek V4 Pro locally on AMD EPYC + RTX PRO hardware using modified llama.cpp with Q4_K_M compression.

DeepSeek V4 full paper details FP4 quantization-aware training, achieving 27% FLOP reduction and 10% KV cache vs. V3.2 with 99.7% recall.

Developer benchmarked local Qwen 3.6 27B vs cloud models on 150 real coding tasks, finding local matched cloud 97% on 35% of workload, suggesting cost arbitrage opportunity.

DeepSeek V4 Pro matches GPT-5.2 on FoodTruck Bench agentic task; first Chinese frontier model, 17× cheaper, within 10 weeks of GPT-5.2 release.

DeepSeek V4 benchmarks below Opus 4.6 and GPT-5.5 but ranks as strongest open-source model; real-world performance estimated at GPT-5.2 level.

DeepSeek v4-Pro priced at $0.145/M input tokens (35x cheaper than Claude Opus 4.7), with promotional rates reaching 138x cheaper on cached tokens through May.

Deepseek V4 Pro offers 100M tokens for $2.65, dramatically undercutting API pricing across the industry.

Community discussion comparing Kimi K2.6 and DeepSeek V4 Pro for real-world use cases, noting K2.6 strength in coding.

GPT-5.5 ranks 2nd on Extended NYT Connections Benchmark; Kimi K2.6 leads open-weights; DeepSeek V4 Pro and Qwen 3.6 Max show significant gains.

Reddit discussion seeking user feedback on DeepSeek v4 Pro coding performance versus Kimi K2.6 and GLM 5.1.

llama.cpp adds experimental DeepSeek v4 Flash support with aggressive 2-bit quantization, achieving 17 tokens/sec on M3 Max with 128GB RAM requirement.

DeepSeek V4 achieves ~7.9x KV cache reduction vs V3.x at 1M context through improved architecture, reducing memory from 68.6GB to 8.7GB for Pro variant.

DeepSeek V4 Pro exhibits degraded token efficiency vs. V3.2 despite 2.5x scale increase, suggesting intelligence density declined.