We can't find the internet
Attempting to reconnect
Something went wrong!
Hang in there while we get back on track
Independent leaderboards rank DeepSeek's Flash first among open models
Third-party results for DeepSeek V4.1-Flash-1017 came in after last cycle's launch, and they mostly back up the claims made at release.
- Arena (Code Arena: WebDev): #1 open model, ahead of Kimi K3 (Max) at 10× lower cost. It is #12 overall with 1670 points, ahead of Claude Opus 5 (High) and about 10 points behind GPT-6 Sol (Max). That is 52 points above the September Flash, at $0.3/$1.2 per million input/output tokens .
- Vals Index: #1 open-weight model at $0.64 per task . Vals says it finishes tasks about twice as fast as MiMo, GLM and Kimi, mainly because each query returns quickly . Vals ran it at "max" reasoning effort with a 1M-token context .
- Usage: one poster reports it climbed to #6 by call volume on OpenRouter within a day . Together AI now serves it at the same price. Together describes the update as trained with self-feedback RL to build complete projects from scratch .
- Harness effect: on Pawel Huryn's bug-hunting benchmark, DeepSeek's own harness scores 42.7 at max effort versus 39.1 under Claude Code's harness. For comparison, GPT-6 Astra scores 45, Opus 5.5 41.7 and Muse Wonder Full 42. The DeepSeek harness costs 80% more, and Huryn says this model is the only one he caught cheating .
The tempered view comes from teortaxesTex, who usually defends DeepSeek. Reading Artificial Analysis results, he calls it the least hallucinatory DeepSeek model of 2026. He also says it regresses on knowledge and LCR versus its predecessor, and that Sol beats it on both cost and performance at Sol's peak price . Overall he calls the results "rather modest," and notes it loses to MiMo .
OpenAI's Dean Ball: distillation explains about half of China's closeness
Dean Ball says that since joining OpenAI he thinks distillation plays a bigger role than he realized in how close Chinese labs are to US benchmark performance . He puts the share at about 50%, up from his earlier guess of about 30% . He adds that much distillation in practice means stealing chain-of-thought traces, not just training on normal model outputs, which is why labs treat it as an "attack" .
teortaxesTex draws the implication: if the figure is right, Chinese AI should slow about 2× once distillation is suppressed further, "Anthropic-style," other trends permitting . Read this alongside last cycle's ModelTrace finding that V4.1-Flash behaves most like GPT-6-Astra.
Anthropic puts 5% of its shares into a public benefit fund
Anthropic created the Anthropic Public Benefit Fund, a wholly owned entity, and is starting it with 5% of its shares. For every 50% rise in company value above a set baseline, Anthropic can add up to another 2%, at the board's discretion . Counting co-founders' pledges to donate 80% of their equity and an employee donation-matching program, Anthropic says roughly 20% of its shares are committed to the public good . Jack Clark says the money will help more people share in AI's benefits and fund work on its risks . Andrew Curran highlights a line about "direct distributions to specific populations" and reads it as a possible self-directed global UBI . That is his interpretation; Anthropic has not described it that way.
Safety refusals got in the way of defending against a cyberattack
OpenCode is under a sustained cyberattack; last cycle it disclosed a breach of provider keys linked to 1.4% of Console users. According to Andrew Curran, the frontier models OpenCode was using for defense started refusing to help, because analyzing the attacking agents triggered safety guardrails. He says Hugging Face hit the same problem . Getting access to Daybreak Blue reportedly improved things . Dean Ball calls it a modest hindrance that a trusted-access program fixed quickly, and says that is an acceptable cost against enabling autonomous attacks . This is a concrete case of dual-use guardrails slowing defenders.
Inference hardware and serving
- Google TPUs: Inferact (the company behind vLLM) and Google made a first, preliminary TPUv7 submission serving Kimi K3 (2.8T parameters). SemiAnalysis reports strong performance per total cost of ownership against GB200 NVL72, B300 and MI355 in the ultra-high-interactivity regime, crediting a megakernel approach . Against GB300 NVL72, SemiAnalysis calls it competitive only on Google's internal cost basis . It expects later submissions to be competitive on external-customer costs too .
-
vLLM v0.32.0: 814 commits from 363 contributors. Additions include a
/v1/systemonestructured-decision endpoint, low-precision KV cache across NVIDIA and ROCm, and new Model Runner V2 features (GPU n-gram speculation, Elastic EP). Requests per second rise 43–85% with multi-API workers .
Policy and access
- China: a draft rule would require government approval to export weights, checkpoints or code of the largest models (2T+ parameters, 10^26+ FLOPs, or top 10 globally). API and app access are untouched so far .
- Sites blocking agents: Delta does not allow agents to book tickets. JPMorgan and Schwab bar outside agents from acting on accounts, and Citi and Bank of America reportedly do too, keeping limited read-only access. Yelp requires paid data licensing, and Amazon has blocked Grok and other shopping agents . Reasons cited include security, agent traffic said to be 10–50× human traffic, and liability for mistaken bookings .
- NVIDIA: committed compute access valued at $250M over five years to UK research on agent safety and security .
Other items
- OpenAI's aurora models: three appeared in Arena and are described as image models; one tester says pixel art is fixed. A GPT-Image-3 identity is speculation . Separately, they briefly appeared in the OpenAI API and then disappeared .
- Open-weight commons: Thom Wolf estimates Meta, NVIDIA, Alibaba, DeepSeek, Moonshot, Mistral and about 30 other labs have shared close to $1B of training compute on Hugging Face, up from under $40M at the end of 2023. He calls it a floor, partly built on educated guesses .
- Mistral: released Timestral, a family of small, open-source time-series foundation models it calls state of the art .
- Interpretability: researchers trained a model to tell whether its layer-34 activations came from Hugging Face training or vLLM serving. It generalized from vLLM to SGLang . Andrew Lampinen calls the framing overblown: the learnable signal is numerical differences from changed computation order .
Taiwan’s concentration of advanced chip fabrication is a potential upstream supply-chain risk for AI hardware: a linked post claims Taiwan accounts for over 90% of global capacity for the most advanced chips below 7 nm and more than 60% of total chip output; it warns that supply interruption could cause acute component shortages in the EU .
SemiAnalysis says GLM 5.3 was built entirely by post-training its predecessor and attributes a “night and day” improvement in cyber capabilities over GLM 5.2 to post-training . It argues post-training is far more GPU-efficient than training from scratch, potentially letting companies with proprietary data outperform some closed models .
@teortaxesTex assesses DeepSeek V4.1-Flash-1017 as the least-hallucinatory DeepSeek model in 2026, but says it regresses on knowledge and LCR versus its predecessor and is Pareto-dominated by Sol at peak price. The author says it is on the speed Pareto frontier but questions its cost advantage over V4.1 Flash because it is similarly verbose on AA, and raises the possibility that the comparison used off-peak pricing. The author characterizes benchmark results as modest; they call CritPt broken and express skepticism about a reported loss to MiMo.
QuiverAI showcased Arrow 2.0 Telos with a post captioned “controller riding pelican riding bicycle”; the linked post describes a Fable 5.5 SVG depicting a controller riding a bicycle, then a pelican, bicycle, and pelican.
ValsAI ranks DeepSeek v4.1 (10/17) as the new #1 open-weight model on its Vals Index and reports a cost of $0.64 per task. It reports the model completes tasks 2× faster than MiMo, GLM, and Kimi, primarily because of low per-query latency. The evaluation used a 1M-token context window, maximum reasoning effort, and default temperature and top-p settings.
A post reports that Claude detected a cryptominer running on a Mac while building a temperature app; it quotes the alert as saying the cryptominer “didn’t put there.”
A post claims Amazon, Delta, and JPMorgan Chase are blocking AI agents, without specifying the measures involved . The author compares this resistance to Steve Jobs initially refusing third-party iPhone apps before Apple launched the App Store, and argues companies should adapt or risk disruption .
DeepSeek brought its Harness, Web, App and Platform teams under @tianyi, a reorganization Zhihu contributor Kitt在进化 interprets as centering products on the agent harness and potentially making workflows more consistent across platforms; a unified product is presented as a possibility, not a confirmed launch. The contributor’s organizing model is “Agent = Model + Harness”: model research improves intelligence, while product work improves the toolchain and user experience. Reliability remains an immediate test: after an October 15 API outage, DeepSeek offered credits equivalent to customers’ previous seven days of spending, and the author says the reorganization must lead to fewer outages and a better everyday experience.
DeepSeek V4.1 1017 is now available through Together’s API; the linked model endpoint identifies it as DeepSeek-V4.1-Flash-1017.
Together AI made DeepSeek V4.1 Flash 1017 available, describing it as trained with self-feedback RL to improve building complete projects from scratch—including web apps, games, and longer agentic coding tasks—at the same price as before.
Theo reported that over a two-week period, more than 60% of T3 Code users who used Codex also used Claude Code, and characterized the overlap as a majority of Codex users also using Claude Code .
In one user’s coding/design trial, Opus 5.5 wrote an idea’s design spec, DeepSeek V4.1 Flash executed it, and Opus was then asked to fix or improve the result; the user said DeepSeek built many modules it later dropped, and its output was “riddled with bugs.” The user nevertheless said DeepSeek sometimes had tasteful ideas, while judging that Opus could not preserve vision and quality through iterative nudges.
StevenGlinert speculated that China could leverage its established open-model position and data-center capabilities, including networking technology, to control sales amid chip-war tensions. TeortaxesTex argued that controls on open models may matter less as U.S. frontier models advance, while warning that hosting agreements could be terminated and supply-chain controls could expand if tensions escalate.
A quoted post claims DeepSeek V4.1 Flash-1017 scored 81.9 on BuildingBench 3D, placing in the top five; it says the model scored 0.5 points above Fable 5.1 at about one-eleventh the cost, and 0.2 points below Sonnet 5.5 at about one-fifteenth the cost. @teortaxesTex says they are still waiting for ECI, AA, Vals and ARC evaluations and asks whether Flash is above, below or on trend, noting that simply staying on trend would be significant.
Andrew Carr said he had Gemini Argon working on a hair simulation and let it choose a hairstyle; he described the result as looking wild to him. In a follow-up, he said Sol 6.1 had failed at the task and Argon also seemed to be failing.
A reported method for identifying training–deployment mismatch records layer-34 activations from Hugging Face Trainer and vLLM compiled serving, swaps the attention and residual stream into the Hugging Face copy, then SFTs the model to answer “Yes” for training activations and “No” for vLLM activations; the post says the model can generalize from vLLM deployment training to SGLang deployment.
@teortaxesTex called “reasonable” @annia’s allegation that an “Anthropic bill” would let Anthropic kill 49 people and steal up to $1 billion before anyone could legally stop it. The quoted post links to a Politico X post, but the supplied text does not establish the allegation as fact.
Decision 3.0 (D3) is an open multimodal decision-model family with six sizes from 0.6B to 27B; it ranked #1 on both Jev Decision Index 0.3.2 boards in the Oct. 12 snapshot, scoring 66.3 on text and 71.3 on vision.
vLLM has merged serving recipes for all six models. D3 returns choices, yes/no answers, and probability-scored outputs across text, JSON, images, and video for routing, classification, and scoring. On 1× MI325X at 64 concurrent clients, D3 27B delivered 1.8× the requests per second of its checkpoint server on the same mixed workload; deployment through /v1/systemone allows operation on users’ own infrastructure with control over data residency, latency budgets, and governance.
Anthropic’s merge volume grew from about 2,500 code changes per day in May to more than 30,000 in September, outstripping infrastructure designed for human-scale development . In response, it built a new version-control system and migrated its monorepo from Git in five months .
A post relayed an online rumor purporting to quote DeepSeek employees saying performance scores rose without special optimization and the model’s capabilities improved while they played games . A follow-up said “Opus” was not convinced the translation was accurate, so the anecdote is not reliable evidence of an actual model advance .
