Kimi K3 topped an LMArena flagship board, priced itself onto the closed vendors' rate card — and shipped without weights. Meanwhile the largest US-origin open-weights model actually landed, and OpenAI showed what it keeps for itself.

Registry movement

GPT-Red is excluded from the tree by design — OpenAI states it will never be released, and the tree tracks deployed or deployable models. The reported Gemma 4 in-place refresh is weakly sourced (single secondary blog) and is covered under vendor signals rather than as a tree update.

Added

kimi-k3, inkling

Updated

None

OpenAI · 2026-07-15 · Specialist · Reasoning

GPT-Red

An internal-only self-play RL red-teaming model, trained at flagship post-training scale, that beat human red-teamers 84% to 13% on indirect prompt injection — and will never be released

Procurement should now treat prompt-injection robustness as a quantified, vendor-differentiating spec: GPT-Red hardened GPT-5.6 Sol to a 0.05% direct-injection failure rate (6x fewer than the best model four months prior), and its attack corpora have been folded into every OpenAI release since GPT-5.3 — numbers you can demand equivalents of from every vendor. The withholding itself is the second signal: frontier labs are starting to keep their strongest capabilities as internal advantages, so vendor selection increasingly buys the byproducts of models you will never see. All figures are vendor-reported; the technical preprint promised the week of Jul 20 is the verification checkpoint.

Sources OpenAI (vendor-reported; preprint promised week of Jul 20); live Andon Labs vending-agent compromise demo

OpenAI via AWS · 2026-07-13 · Frontier · Reasoning

GPT-5.6 family (Sol / Terra / Luna) on Amazon Bedrock

Full family GA on Bedrock's Responses API endpoint at OpenAI first-party rates that count toward AWS commitments — with 90% prompt-cache discounts via explicit cache breakpoints

Teams with AWS enterprise commitments can now route frontier OpenAI traffic through existing spend at no rate premium — re-run the make-vs-route math on any direct OpenAI contract renewal. Mind the caveats before migrating: context is capped at 272K on Bedrock, and Sol is limited to us-east-1/us-east-2 (Terra and Luna add us-west-2), so region-sensitive workloads need the smaller tiers. The 90% prompt-cache discount with explicit breakpoints makes cache design, not list price, the dominant cost lever for agent fleets on this endpoint.

Sources AWS What's New; Bedrock model card

Moonshot AI · 2026-07-16 · Frontier · Moe

Kimi K3

2.8T MoE (16 of 896 experts, ~50B active), 1M context, Kimi Delta Attention, MXFP4 quantization-aware training — hosted-only at launch, with weights promised by Jul 27

Hold the 'open' celebration until Jul 27: K3 launched API-only, the license is unpublished, the HF repo 404'd the day after launch, and AA classifies it proprietary in the interim — so treat it as a closed model with an open-weights option pending, and gate any self-hosting plan on the actual license text. The capability is real regardless: #1 on the LMArena Frontend Code Arena (first open-weight-lab model to top a flagship board), #4 on the AA Intelligence Index, and the best overall SWE Marathon score (42.0, vendor-reported on the KimiCode harness). Plan for heavy serving if weights do land — Moonshot recommends 64+ accelerator supernodes — and note it launched with only reasoning_effort=max, so token spend runs verbose (130M tokens on the AA eval vs a 63M peer average).

Model registry ID

kimi-k3

Sources Moonshot AI; Artificial Analysis; LMArena; Simon Willison

Thinking Machines Lab · 2026-07-15 · Open Frontier · Moe

Inkling

The largest US-origin open-weights model: 975B/41B-active trimodal MoE under Apache 2.0, 1M context, 45T training tokens, BF16 + NVFP4 checkpoints, day-0 transformers/vLLM/SGLang/llama.cpp

If your open-model strategy assumed the frontier-scale open lane belonged to Chinese labs, revise it: Inkling is a clean Apache 2.0, US-origin base explicitly positioned for fine-tuning and domain adaptation, with serving partners (Baseten, Modal, Databricks, TogetherAI, Fireworks) live at launch and a managed post-training platform (Tinker) attached. Budget realistically — ~2TB VRAM for BF16 or ~600GB for the calibrated NVFP4 checkpoint — and watch the previewed Inkling-Small (276B/12B active) as the practical deployment tier for most teams. The day-0 NVFP4 release on GB300-trained infrastructure also signals where open-model serving economics are heading: 4-bit datapaths as the default, not an afterthought.

Model registry ID

inkling

Sources Thinking Machines Lab via Hugging Face; NVIDIA (GB300 NVL72 training confirmation)

Extreme MoE sparsity converges on ~2-4% activation at trillion scale

Three independent trillion-scale designs now activate only 2-4% of parameters per token, which means serving cost tracks memory capacity and interconnect bandwidth, not FLOPs — the constraint that matters when sizing inference fleets for these models. It also makes routing stability a first-order engineering concern (Quantile Balancing, Per-Head Muon are the current techniques), because a mis-routed expert at 1.8% activation wastes a much larger share of the useful compute. Capacity planners should price these models on HBM footprint and east-west bandwidth, and treat vendor active-parameter counts as the real sizing number.

Examples

Kimi K3 (16 of 896 experts, ~1.8%), Inkling (41B of 975B, 4.2%), DeepSeek V4 Pro (49B of 1.6T, ~3%)

Sources Moonshot AI; Thinking Machines Lab; DeepSeek

4-bit-native training collapses the release-to-deployment gap

Both of the week's big releases trained with 4-bit precision in the loop rather than quantizing after the fact — K3 ran quantization-aware training from the SFT stage, and Inkling shipped a calibrated NVFP4 checkpoint the same day as BF16. That removes the accuracy-loss lottery that used to sit between an open-weights release and a production deployment, and it couples open models to FP4-capable datapaths. Procurement implication: hardware refresh decisions should now weight native FP4 support heavily, because the open-model ecosystem is standardizing on it at the source.

Examples

Kimi K3 (MXFP4/MXFP8 QAT from the SFT stage), Inkling (calibrated NVFP4 day-0, GB300-trained)

Sources Moonshot AI; Thinking Machines Lab via Hugging Face

1M context becomes the default tier — with flat pricing

Hybrid linear/sparse attention designs (K3's Kimi Delta Attention, DeepSeek Sparse Attention) have made million-token context cheap enough that Moonshot prices it flat — the same $3/$15 per M tokens at 1M as at 1K, with no long-context surcharge. Once long context stops carrying a price penalty, prompt-cache hit rate becomes the dominant cost lever for agent workloads: the difference between a well-designed cache ($0.30 cached-input on K3, 90% discounts on Bedrock GPT-5.6) and a naive one dwarfs the model-choice delta. Architect agent memory around cache breakpoints before optimizing anything else.

Examples

Kimi K3 (KDA, flat $3/$15 across 1M), Inkling (1M), DeepSeek Sparse Attention

Sources Moonshot AI; AWS Bedrock documentation

LMArena Frontend Code Arena

Kimi K3 debuted at #1 with 1,679 (preliminary) — the first model from an open-weight lab to top an LMArena flagship board, 48 points clear of Claude Fable 5

Kimi K3

1,679 (prelim) — new #1

Claude Fable 5

1,631

GPT-5.6 Sol

1,618

GLM-5.2

1,587

Sources LMArena

Artificial Analysis Intelligence Index

K3 entered #4 of 189 at 57.1, above Opus 4.8 — but generated 130M tokens on the eval (63M peer average) at $0.94 per Index task, and AA classifies it proprietary until weights land

Claude Fable 5

59.9 — retains the lead

GPT-5.6 Sol

58.9

Kimi K3

57.1 — new, #4 of 189

Claude Opus 4.8

56

Sources Artificial Analysis

LMArena Text Arena

K3 entered #9 at 1,486±11 — a 28-place jump over K2.6's #37, moving Moonshot from mid-pack to the top-10 conversation tier in one generation

Kimi K3

1,486±11 — #9, new this week

Kimi K2.6

#37 (prior generation)

Sources LMArena

Agentic coding suite (vendor-reported, KimiCode harness)

K3 posted the best overall SWE Marathon score and near-frontier terminal results — but every number comes from Moonshot's own harness, and W28's lesson was that harness choice can swing cost and scores 2x, so compare against Sol's Codex-harness numbers with care

Kimi K3 — SWE Marathon

42.0 (best overall, vendor-reported)

Kimi K3 — Terminal-Bench 2.1

88.3

Kimi K3 — FrontierSWE

81.2 (Fable 5: 86.6)

Kimi K3 — GPQA-Diamond / BrowseComp

93.5 / 91.2

Sources Moonshot AI (vendor-reported, KimiCode harness)

2026-07-16 · Moonshot AI

K3 pricing resets the Chinese-lab tier: $0.30 cached / $3.00 input / $15.00 output per M tokens — input up 216% and output up 275% versus K2.6, matching Sonnet 5's standard rate card (~50% above its live intro price)

The 'Chinese open models are 10x cheaper' planning assumption is dead: Moonshot is pricing on capability, not on lane. Re-run any routing or budget model built on that assumption — the K2.x tiers remain available as the value option, but the frontier Chinese tier now costs what the US mid-frontier costs.

Sources Moonshot AI pricing documentation; Simon Willison

2026-07-16 · Moonshot AI

K3 launched as 'open' with no weights: hosted-only, license unpublished until the promised Jul 27 drop, HF repo 404'd the day after launch — AA classifies it proprietary in the interim

'Open' is becoming a launch-marketing label decoupled from the license-and-weights fact. Contract and compliance teams should gate any K3 self-hosting or fine-tuning commitment on the actual license text, and treat 'weights coming' announcements as vendor roadmap, not product.

Sources Artificial Analysis; Simon Willison

2026-07-14 · DeepSeek

Hard retirement of the deepseek-chat / deepseek-reasoner aliases lands Jul 24, alongside the first time-of-day surge pricing from a major API — 2x listed rates during Beijing peak hours — with V4 GA signaled as imminent

Surge pricing is a structural first: inference is being priced like a capacity-constrained utility, so cost models that assume flat per-token rates need a time-of-day dimension, and latency-tolerant batch workloads should be scheduled into off-peak windows. The alias retirement is a hard migration deadline — audit for pinned deepseek-chat/deepseek-reasoner IDs now.

Sources DeepSeek API news

2026-07-15 · OpenAI

GPT-Red stays internal: a model trained at flagship post-training scale that OpenAI states will never be released — its output ships only as hardening folded into released models

This is the clearest capability-gating precedent yet: the strongest capabilities may increasingly arrive as vendor-internal advantages you rent the byproducts of, not products you evaluate. Expect competitors to follow, and adjust diligence accordingly — ask vendors what internal-only systems shaped the model you are buying, and demand the quantified safety deltas (0.05% direct-injection failure is now the number to beat).

Sources OpenAI (vendor-reported)

2026-07-15 · Google

Reported in-place refresh of the Gemma 4 weights, FlashAttention-4 kernels, and chat templates across the HF collection — targeting tool-call JSON consistency, vision OCR, and long-prompt prefill (weakly sourced; single secondary blog)

If confirmed, in-place weight refreshes under an unchanged model name break reproducibility assumptions: self-hosted Gemma 4 operators should re-pull stale local caches and pin content hashes, not model names, in any evaluation or compliance pipeline. Treat this as a caution flag until Google documents it first-party.

Sources ExplainX (secondary; not yet confirmed first-party)

By Jul 27

Kimi K3 weights, license, and tech report

The week's biggest open question: if clean weights and a permissive license land as promised, K3 becomes the strongest open-weights model on record and the open-frontier scorecard flips; if the date slips or the license restricts, the 'open launch without weights' pattern hardens into a playbook.

Jul 24

DeepSeek V4 GA + hard alias retirement

deepseek-chat and deepseek-reasoner go dark the same window V4 is signaled to arrive ('mid-July' per DeepSeek) — the forced migration will produce the first clean V4 adoption data and test whether surge pricing survives contact with customers.

Week of Jul 20

GPT-Red technical preprint

Every GPT-Red number is currently vendor-reported; the preprint is the first chance for independent scrutiny of the 84%-vs-13% red-teaming claim and the Fake Chain-of-Thought injection class, which together justify making injection robustness a procurement spec.

By Jul 31

Gemini 3.5 Pro slip watch

The leaked Jul 17 target passed — the third slip. Prediction-market odds (~81% by Jul 31 on Polymarket) are the only forward signal available and are speculation-grade; treat any capacity or contract planning around a 2M-context flagship as unanchored until Google commits a date.

No date announced

Inkling-Small (276B/12B active) weights

The previewed small tier is the practical deployment target for most teams eyeing the Inkling base — its release (and whether the Apache 2.0 terms carry over) determines whether Thinking Machines' open push reaches beyond 600GB-class serving budgets.

• Two tree rows added (kimi-k3, inkling). K3 establishes an openness convention for this publication: models launched hosted-only with weights promised are recorded as closed until the weights and license actually ship — the row will be updated if the Jul 27 drop lands.
• The reported Gemma 4 in-place refresh was handled as a vendor signal rather than a tree update due to weak sourcing (single secondary blog, no first-party confirmation).
• GPT-Red was covered in frontier movements but excluded from the tree: OpenAI states it will never be released, and the tree tracks deployed or deployable models.