Actualités DeepSeek

Actualités DeepSeek

Suivez les actualités des modèles DeepSeek, de l’API, des tarifs et des outils pour développeurs, ainsi que les nouvelles de l’IA. DeepSeek V4 Hub est une publication indépendante.

Dernières actualités

135
DeveloperArticle en anglais

DeepSeek Open-Sources DeepInfer: vLLM and SGLang Match Its Performance Within a Week

DeepSeek published DeepInfer, its own inference engine, in late October 2026. Within a week, vLLM and SGLang both shipped changes that land at or above it on the configurations each team disclosed. vLLM measured a 6% decode-throughput gap on GB200 with 131K context and 954K tokens/s per GPU on a GB300 NVL72 rack; SGLang reported up to 4% higher prefill throughput and 28.88 ms per decode step at 256 requests per GPU against DeepInfer's 31.19 ms. Every number is self-reported on different hardware, so treat them as per-config results rather than a universal ranking.

Lire l’article
AnalysisArticle en anglais

Kimi K3.1 Pro vs Gemini 4 Argon: Two Ways to Build a Flagship

Both are top flagship models, but they structure speed and cost tradeoffs differently. Moonshot's Kimi K3.1 Pro offers one underlying model with two serving tiers: Moonshot says Highspeed delivers 2 to 3 times the output speed at double the token price. Gemini 4 Argon ships as a single flagship endpoint, with latency and cost tuned via thinking levels and API service tiers. For Kimi, you pick a model ID on each request. For Argon, you set a thinking budget and service level.

Lire l’article
AnalysisArticle en anglais

Kimi K3.1 Pro vs DeepSeek V4.1 Flash: Two Different Ways to Purchase Speed

Kimi K3.1 Pro and DeepSeek V4.1 Flash both target agent workloads, and they lower the cost per step in different ways. Kimi keeps one flagship model and sells a Highspeed tier that Moonshot says runs 2 to 3 times faster at twice the token price. DeepSeek builds a separate lightweight variant with a small active path, a $0.006 cached-input rate and half-price off-peak hours. Long, multimodal runs fit Kimi; high-volume loops with repeated prefixes fit DeepSeek.

Lire l’article
AnalysisArticle en anglais

Kimi K3.1 Pro vs Claude Opus 5.5: Two Routes to Fast Agent Work

Kimi K3.1 Pro and Claude Opus 5.5 both target long-running coding and Agent workflows with a 1M-token context. The key difference is packaging. Kimi is a 3T model on an open-weight roadmap, offered via a standard ID plus a separate Highspeed ID at double the token price. Claude Opus 5.5 is a closed flagship priced at $4 input / $20 output per million tokens; its fast mode lives as a request parameter on the same model. Kimi offers a planned self-hosting path and developer-controlled search endpoints. Claude delivers a consistent cross-cloud model and model-triggered cited search.

Lire l’article
AnalysisArticle en anglais

Kimi K3.1 Pro vs GPT-6.1 Sol: Two Paths for Faster Agent LLMs

Kimi K3.1 Pro and GPT-6.1 Sol both target cheaper Agent workloads, but take opposite approaches. Moonshot maintains one 3T-class model with up to 1M-token context on premium tiers, selling speed as an extra serving tier. OpenAI releases Sol as a lighter, independently tuned model, priced at $2 input / $10 output per million tokens. Long, context-heavy sessions suit Kimi; short, high-volume loops suit Sol.

Lire l’article
AnalysisArticle en anglais

Breakout: Kimi K3.1 Pro Hits Global Top 4, Outranks GPT‑6 and Gemini 4

Kimi K3.1 Pro scores 55 on the Artificial Analysis Intelligence Index, fourth worldwide behind Claude Opus 5.5, Claude Sonnet 5.5 and Meta's Muse Wonder Full, and two points ahead of GPT-6 Astra and Gemini 4 Argon. It has the same 2.8T total and 104B active parameters as K3; the 11 point gain came from post-training. It leads the independent Terminal-Bench 4.0 run at 67.2% and costs about $2 input and $10 output per million tokens, the same tier as Claude Sonnet 5.5. Weights are promised before November 4. Human voters still rank it eighth.

Lire l’article
OfficialArticle en anglais

Kimi K3.1 Pro released: 1M context, API pricing and what changed

Moonshot AI released Kimi K3.1 Pro and Kimi K3.1 Pro Highspeed on October 23, 2026. The company describes Pro as a 3-trillion-parameter-scale model with a 1-million-token context window. Both versions are available through Kimi's apps, Kimi Code and its API. Highspeed is advertised as roughly 2 to 3 times faster at output and costs twice as much per token. Full model weights are promised by November 4; they were not part of the October 23 launch.

Lire l’article
AnalysisArticle en anglais

Z.ai Open-Sources HarnessPro — Agent Self-Evolution That Cuts Costs by Up To 53%

Z.ai and Tsinghua's Jie Tang team have open-sourced HarnessPro, a framework that lets a coding agent rewrite its own harness between runs. It keeps a Pareto set of three operating points: Lite for cost, Medium for balance, and Heavy for performance. On SWE-Bench Pro with GPT-5.5 and Codex, Medium reached 79.0% pass rate at $1.26 per task, 53.0% below RHO's $2.68. The trade-off is explicit: Medium gives up 1.0 percentage point against RHO. Heavy beat RHO on all three benchmarks while costing less.

Lire l’article
LeakArticle en anglais

Kimi K3.1 Leaks: Three.js Demos and the 1/14 Cost Claim

Early reports describe Kimi K3.1 building and revising Three.js and WebGL scenes, with promising static composition and a reported weakness in character animation. The claimed 1/14 cost versus GPT-6 Astra and 70% of Kimi K3 cost are tester reports without a published evaluation method. The official Kimi API model list checked on October 23, 2026 does not list K3.1. Treat the demos as evaluation leads, not a production recommendation.

Lire l’article
AnalysisArticle en anglais

What DeepInfer Means for the LLM Inference Ecosystem (Oct 2026)

DeepSeek open-sourced DeepInfer on October 20, 2026. The release provides offline inference for DeepSeek-V4.1-Flash on NVIDIA SM100 and Huawei Ascend 950. Its value is a concrete reference for model and engine co-design: architectural KV reuse, bounded replay, and shared model logic across two hardware paths. A complete online service still needs cache storage, disaggregated deployment, and load balancing. The chip-market implications below are editorial analysis, not a cross-platform performance test.

Lire l’article
AnalysisArticle en anglais

Personal Agents: The Vision and Reality of Autonomous AI Assistants (Oct 2026)

By October 2026, products such as Muse and Grok Bot made personal agents a concrete proposition: assistants that can keep working after you close the app. The defining feature is persistent work, backed by saved task state, tools, and schedules or event triggers. A dedicated cloud computer or separate identity can help, but neither is universal. This guide explains what these agents can do and what to check before giving one access to your accounts.

Lire l’article