For architects tracking model capability shifts.
Open weights finally became deployable artifacts — and the closed stack answered with post-training and rate cards, not new flagships
Week 31 of 2026 · August 1, 2026
Big read
Moonshot's Jul 27 Kimi K3 weight drop converts last week's promissory open-frontier watch item into a procurement fact: a 2.8T MoE with 104B active parameters, 896 experts selecting 16, hybrid 69 KDA + 24 Gated MLA layers, 1,048,576 context, and MoonViT-V2 — under a revenue-tiered Kimi K3 License that triggers a separate commercial deal for MaaS above $20M trailing-twelve-month revenue. Self-host evaluation can now start on artifacts; license diligence is the gate, not model quality.
No closed frontier model shipped. The closed stack competed on price and post-training instead. OpenAI cut GPT-5.6 Luna 80% to $0.20/$1.20 and Terra 20% to $2/$12 on Jul 30 while holding Sol at $5/$30 and replacing Priority Processing with Fast mode at 2x token price for up to ~2.5x speed. DeepSeek put V4-Flash-0731 into public beta on Jul 31 on the same 284B/13B architecture and unchanged $0.14/$0.28 pricing; Artificial Analysis independently scored it +10 Intelligence Index points to 50. Thinking Machines released Inkling-Small the same day as the Luna cut — 276B/12B active, Apache 2.0, independent AA Index 40 versus Inkling's 41 — with a context-window conflict to resolve (vendor card up to 1M; AA article lists 256K).
The week's largest vendor-claimed capability jump did not need new weights either: OpenAI reported ARC-AGI-3 public-set scores moving from 13.3% to 38.3% on GPT-5.6 Sol by retaining reasoning between tool calls and compacting context instead of truncating, with roughly 6x fewer output tokens. That is a harness result, not an official leaderboard reprint — and no qualifying LMArena, SWE-bench, or ARC official board moves landed in-window. Gemini 3.5 Pro missed its Jul 31 GA prediction. MiniMax H3 promised open weights and had not shipped them as of Aug 1. Net/net: route on measured Index and task cost, treat open as license-plus-weights, and re-test agents against memory policy before buying a bigger model.
Tree delta
Inkling-Small joins the tree as the strongest US open-weights efficiency result to date, reaching an independent Intelligence Index of 40 against Inkling's 41 with under a third of the parameters. Kimi K3 is reclassified from closed to open weights on the July 27 artifacts and license text, with its active-parameter count corrected from a community estimate to a confirmed 104B. DeepSeek V4 Flash updates for the 0731 re-post-train checkpoint that gained 10 independent Index points on unchanged architecture.
Registry movement
Kimi K3 is an update rather than an addition: the row has existed since W29, and this week's material change is openness, license terms, and a corrected active-parameter count of 104B against the earlier community-estimated ~50B. Qwen3.7 Flash is held out of the tree pending a first-party model card, and MiniMax H3 is deferred until its promised weights actually ship. Both are covered in the cards.
- Added
- inkling-small
- Updated
- kimi-k3, deepseek-v4-flash
Frontier movements
OpenAI · 2026-07-30 · Frontier · Reasoning
GPT-5.6 Luna / Terra price cut and Sol Fast mode
Re-baseline high-volume routing immediately: Luna undercuts most Western cheap tiers on sticker price while Sol remains the premium band against Opus 5. Treat Fast mode as a purchasable latency tier rather than a separate model, and instrument per-tier completion rates before moving fleet volume on list price alone. The AI Stack Weekly owns the fixed-workload spread calculation; this Pulse stays on the architecture and routing implication.
DeepSeek · 2026-07-31 · Frontier · Moe
DeepSeek V4-Flash-0731
Drop this into existing Flash routes as a cheap-tier agent upgrade without a size or price change — but do not treat it as V4-Pro GA, and pin the harness when comparing Terminal-Bench (vendor-claimed 82.7 max effort versus Artificial Analysis 79%). Architecture and model ID stay deepseek-v4-flash; Pro, app, and web remain unchanged pending the promised official V4-Pro.
Sources DeepSeek API changelog; Artificial Analysis (independent)
OpenAI · 2026-07-29 · Reasoning · Agentic
GPT-5.6 Sol harness memory on ARC-AGI-3
Treat the ARC-AGI-3 jump as vendor-claimed on a single harness, not an official leaderboard reprint: OpenAI reports roughly 6x fewer output tokens when compaction replaces rolling truncation on the Responses API, and no qualifying LMArena or SWE-bench official moves landed in-window. Agent Techniques Weekly owns the method and the reproduction caveat.
Alibaba / Qwen · 2026-07-27 · Specialist · Multimodal
Qwen3.7 Flash
Useful as a fleet-cost floor candidate for multimodal agents, but treat specs and pricing as secondary until a first-party model card is pinned. Sources are OpenRouter listing dates and technical digests citing a QwenCloud changelog; no qwenlm.github.io launch post was found in-window.
Sources eesel.ai; BenchLM (secondary; no primary lab announcement)
Open weights
Moonshot AI · 2026-07-27 · Open Frontier · Moe
Kimi K3
Start the self-host evaluation now — GitHub and Hugging Face artifacts landed Jul 27 — but send the Kimi K3 License to counsel before any commercial hosting plan: MaaS above $20M aggregate trailing-12-month revenue needs a separate deal, with attribution triggers above 100M MAU or $20M monthly revenue. Independent AA Index reference remains 57; vendor-claimed table includes GPQA Diamond 93.5 and DeepSWE 67.5. Hosted API stays $3/$15 (cached input $0.30).
- Model registry ID
- kimi-k3
Sources MoonshotAI/Kimi-K3 GitHub; Hugging Face; license text; Unite.AI
Thinking Machines Lab · 2026-07-30 · Open Frontier · Moe
Inkling-Small
Strongest US open-weights efficiency play this week for teams that cannot absorb K3-class cluster scale: BF16 needs ~600 GB aggregate VRAM, NVFP4 ~180 GB, with vLLM v0.26.0 shipping same-week Inkling family support. Flag the context conflict honestly — vendor model card says up to 1M; Artificial Analysis lists 256K in its article blurb — and use Index 40 as the independent baseline rather than vendor-claimed SWE-bench Verified 80.2% alone.
Sources Thinking Machines Lab; Artificial Analysis (independent Index 40); VentureBeat
MiniMax · 2026-07-31 · Specialist · Multimodal
MiniMax H3
Treat current state as API-only until the repo appears: unified text/image/video/audio understanding to video with native stereo, up to 15s at 2K, with vendor claims that 2K per-second price is under one-third of mainstream models. Do not build a self-host plan around the promise; the W30 K3 lesson applies again.
Sources MiniMax blog; SCMP; secondary weights-not-yet status
Architecture watch
Post-training and harness policy move scores without new weights
Two independent vendors moved measured capability this week without shipping new parameter counts: DeepSeek re-post-trained Flash in place, and OpenAI changed Responses API memory settings on the same Sol weights. Architects should treat reasoning retention, context compaction, and post-training checkpoints as first-class evaluation dimensions alongside the model ID. Buyers who only diff weight releases will miss the largest week-over-week gains.
- Examples
- DeepSeek V4-Flash-0731: +10 AA Index on unchanged 284B/13B, OpenAI ARC-AGI-3: 13.3% to 38.3% via retained reasoning + compaction, OpenAI Jul 29 serving/harness efficiency post (~20% serving cost, >15% token-gen efficiency)
Revenue-tiered open licenses turn legal review into the deployment gate
Downloadable weights no longer imply a single open-source diligence checklist. K3 is MIT-like until commercial hosting crosses the revenue trigger, at which point counsel must negotiate; Inkling-Small remains Apache 2.0 with no such trigger. Procurement should classify each open release by license economics before sizing clusters, and keep Apache-clean alternatives on the shortlist when MaaS revenue is material.
- Examples
- Kimi K3 License: >$20M trailing-12-month MaaS commercial trigger, Kimi K3 attribution triggers above 100M MAU or $20M monthly revenue, Inkling-Small: clean Apache 2.0 contrast case
Serving stacks ship in the same week as the open MoE class
Production self-hosters got a same-week inference path for the new open MoE class rather than a multi-month serving lag. vLLM v0.26.0 (Jul 27) adds Inkling CUDA-graph and Hopper FA4 relative attention support alongside DeepSeek-V4 serving pushes and matured KV offloading. Capacity planners should price FP4 datapaths and KV-tiered storage into the model-release decision, not as a later optimization.
- Examples
- vLLM v0.26.0: Inkling modeling stack, NVFP4, MTP speculative decoding, Inkling-Small BF16 ~600 GB / NVFP4 ~180 GB VRAM floors, DeepSeek-V4 routing and speculative optimizations across NVIDIA/AMD/XPU
Sources vLLM project GitHub release v0.26.0; Thinking Machines model card
Near-iso Intelligence at a fraction of active parameters
Artificial Analysis places Inkling-Small within one Index point of full Inkling while activating 12B rather than 41B parameters, and notes no smaller open-weights model scores higher in that size class. Efficiency MoEs are now a procurement tier of their own: teams should ask whether the last Index point is worth 3x the serving footprint before defaulting to the flagship open checkpoint.
- Examples
- Inkling-Small: AA Index 40 vs Inkling 41 at <1/3 parameters, Inkling-Small: 12B active vs Inkling 41B active, DeepSeek V4 Flash prior: also sat at AA Index 40 in the same size class
Benchmark moves
Artificial Analysis Intelligence Index — DeepSeek V4 Flash
Independent Index jumps from 40 to 50 (+10) for Flash-0731 on unchanged architecture; now ~1 point behind GPT-5.6 Luna (51) and 6 points above V4 Pro on AA's comparison
- DeepSeek V4 Flash 0731 (max)
- 50 Index; GDPval-AA v2 1559 Elo; Terminal-Bench 2.1 79% (AA)
- Prior DeepSeek V4 Flash
- 40 Index; GDPval-AA v2 1189 Elo
- Kimi K3 (max)
- 57 — open-frontier reference
- GPT-5.6 Luna (max)
- 51
- GLM-5.2 (max)
- 51
Artificial Analysis Intelligence Index — Inkling-Small
New open US model lands at Index 40, within 1 point of Inkling (41) at less than a third the parameters
- Inkling-Small
- 40 (independent)
- Inkling
- 41 (independent)
- DeepSeek V4 Flash (prior max, same size class)
- 40
- MiniMax-M2.7 (per AA)
- ~38 — about 2 points below Inkling-Small
ARC-AGI-3 public set (vendor harness, not official board)
OpenAI reports GPT-5.6 Sol rising from 13.3% to 38.3% by retaining reasoning and compacting context — same model, roughly 6x fewer output tokens
- GPT-5.6 Sol + retained reasoning + compaction
- 38.3% (vendor-reported)
- GPT-5.6 Sol + rolling truncation baseline
- 13.3% (vendor-reported)
Sources OpenAI (vendor-reported; not an official ARC leaderboard reprint)
DeepSeek Terminal-Bench 2.1 (vendor harness vs independent)
Vendor-claimed max-effort 82.7 versus Artificial Analysis 79% on the same Flash-0731 checkpoint — pin the harness before ranking
- DeepSeek V4 Flash 0731 (DeepSeek harness, max effort)
- Terminal-Bench 2.1 82.7 (vendor-claimed)
- DeepSeek V4 Flash 0731 (Artificial Analysis)
- Terminal-Bench 2.1 79% (independent)
- DeepSeek V4 Flash 0731 (DeepSeek harness)
- Cybergym 76.7; Toolathlon verified 70.3; DeepSWE 54.4 (vendor-claimed)
Sources DeepSeek API changelog (vendor); Artificial Analysis (independent)
Tier scorecard
As of 2026-08-01
| Tier | Leader | Challenger | Read |
|---|---|---|---|
| Closed frontier | Claude Fable 5 | Claude Opus 5 | No new closed flagship shipped; Fable retains the absolute reference while Opus 5 remains the economic default at $5/$25. |
| Open frontier | Kimi K3 | Inkling-Small | K3 takes the lead now that weights and license text are public (AA Index 57); Inkling-Small is the efficiency challenger at Index 40 under Apache 2.0. |
| Reasoning | Claude Fable 5 | GPT-5.6 Sol | Sol's vendor-reported ARC-AGI-3 harness jump shows memory policy can move scores without new weights; independent cost-per-task still pending. |
| Coding | Claude Opus 5 | DeepSeek V4 Flash 0731 | Opus 5 holds the practical coding lead on W30 launch evidence; Flash-0731 is the cheap-tier agent upgrade at Index 50 and unchanged $0.14/$0.28. |
| Multimodal | Gemini 3.6 Flash | Kimi K3 | Gemini remains the public fleet multimodal default; K3 now adds self-host control with MoonViT-V2. Gemini 3.5 Pro still has not GA'd. |
| Edge / small | Inkling-Small | Gemini 3.5 Flash-Lite | Inkling-Small brings open multimodal agents to ~180 GB NVFP4; Flash-Lite remains the hosted speed-cost floor at roughly 350 tok/s. |
Vendor signals
2026-07-30 · OpenAI
GPT-5.6 Luna cut 80% to $0.20/$1.20 and Terra 20% to $2/$12; Sol held at $5/$30; Fast mode replaces Priority Processing
Closed labs are competing on completed-task economics via rate cards and latency tiers rather than new flagship weights this week. Route only latency-sensitive work into Fast mode and keep ordinary asynchronous agents on standard Luna or Terra after measuring completion quality.
Sources OpenAI
2026-07-27 · Moonshot AI
Kimi K3 License ships with weights: MIT-like until MaaS revenue exceeds $20M trailing twelve months
Open-weights diligence now splits into quality evaluation and license classification. Counsel should map expected MaaS and product revenue against the trigger before any production self-host commitment; Apache-clean alternatives remain available for teams that cannot accept the commercial clause.
2026-07-31 · Google DeepMind
Gemini 3.5 Pro still not generally available as the standing Jul 31 prediction deadline expires
Do not plan Q3 procurement around 3.5 Pro. Buy the callable Flash and Flash-Lite fleet now; another slip further weakens roadmap-based purchasing.
Sources Google (Jul 21 partner-testing status; no in-window GA found)
2026-07-31 · MiniMax
H3 launches with an explicit promise to open weights in the coming days; no public HF or GitHub artifacts as of Aug 1
Repeat the K3 lesson: an open-weights commitment is not a downloadable artifact. Keep H3 on the API watchlist until the repo and license appear.
2026-07-27 · vLLM project
vLLM v0.26.0 ships Inkling family support, NVFP4, MTP speculative decoding, and DeepSeek-V4 serving optimizations
Serving readiness is now part of the model-release race. Self-host teams evaluating Inkling-Small or DeepSeek V4 this week should pin v0.26.0 as the baseline inference path rather than waiting for a later stack catch-up.
Sources vLLM GitHub releases
Watchlist
Next 7-14 days
MiniMax H3 downloadable weights and license
Decides whether H3 becomes a real open video-generation option or remains an API product wrapped in open-language marketing.
Next 30 days
DeepSeek V4-Pro official GA and pricing
Flash-0731 is not Pro GA; the standing surge-pricing and official Pro triggers remain incomplete until a callable Pro ID and rate card land.
Next 14 days
Inkling-Small context window: vendor 1M vs AA 256K
Procurement specs need a resolved context number; watch for an AA methodology note or Thinking Machines clarification.
By Aug 15
Independent Opus 5 cost-per-task results
Vendor benchmark leadership from W30 is still not enough; the market needs task cost, token use, and latency against Fable 5 and GPT-5.6 Sol.
Next 2-3 weeks
LMArena Opus 5 Elo stabilization
No qualifying in-window official board move; secondary pages remain unstable — wait for a date-stamped stable print before re-ranking coding.
Next 30 days
Gemini 3.5 Pro callable model ID and public pricing
The Jul 31 miss is now a track-record item; only a public GA closes the W30 prediction cleanly.
Changelog
- W31 opens on the Kimi K3 weights-and-license conversion, Inkling-Small Apache 2.0 efficiency play, DeepSeek Flash-0731 independent +10 Index jump, and OpenAI Luna/Terra cuts plus harness ARC-AGI-3 result.
- Negative findings recorded as content: Gemini 3.5 Pro missed Jul 31 GA; no qualifying LMArena/SWE-bench/ARC official leaderboard moves; MiniMax H3 weights promised but not shipped.
- Tree delta lists kimi-k3 under added for this issue's openness conversion while models.yaml still awaits an Inkling-Small row; editor should restructure to added:[inkling-small] / updated:[kimi-k3] once the YAML lands.