STEVEN STEINBECK
← ALL DISPATCHESDISPATCH 83

WEEK OF JUL 25 – 31, 2026

Open weights hit 2.8T, then the price of inference fell off a cliff

Moonshot published the largest open model yet, DeepSeek re-trained a cheap Flash model into an agent specialist, and two frontier labs admitted their evals reached the real internet.

The last full week of July was an open-weights and price-per-token week. The headline numbers got absurd, but the useful stories were about what you can actually serve, what it costs, and whether the safety harness is real.

The week ended with DeepSeek showing the sharpest version of the trend: same weights, dramatically better agent behavior, and a model cheap enough to make long-running loops feel ordinary.

FRI · Jul 31, 2026Open weightsAgents

DeepSeek re-post-trains V4-Flash into an agent specialist

DeepSeek promoted V4-Flash to public beta as V4-Flash-0731 with MIT-licensed weights. The architecture did not change: 284B total parameters, 13B active, 1M context. The jump came entirely from post-training for coding, agents, reasoning and tool use.

Terminal Bench moved 61.8 → 82.7, DeepSWE 7.3 → 54.4 and Toolathlon 49.7 → 70.3, putting the cheap Flash build ahead of DeepSeek’s own Pro preview on every agent benchmark it publishes. Price stayed at $0.14/$0.28 per million input/output tokens.

DeepSeek V4 Flash model artwork
DeepSeek

WHY IT MATTERS

This was the week’s cleanest proof that post-training, not scale, is where agent capability is moving — a 13B-active model above its own much larger sibling at a fraction of the cost.

DeepSeek API updatesV4-Flash-0731 model cardMarkTechPost analysis
THU · Jul 30, 2026ModelsEfficiency

A 12B-active model beats its 41B-active sibling on agentic coding

Thinking Machines released Inkling-Small, a 276B-total / 12B-active MoE with up to 1M context. The company reports SWE-Bench Verified at 80.2% versus 77.6% for its 975B sibling, and ARC-AGI-2 at 40.1% versus 36.5%, while factual recall regresses sharply.

It is priced at $0.30/$1.20 per million tokens with day-zero SGLang, Unsloth GGUF and Baseten support. The same week the lab published its open-weights safety process and adversarial fine-tuning results.

Inkling-Small announcement artwork
Thinking Machines Lab

WHY IT MATTERS

Smaller and better is domain-specific. Agent coding improved while factual recall fell — exactly why choosing a model by one leaderboard is a trap.

Thinking Machines: Inkling-SmallModel cardSafety process
MON · Jul 27, 2026Open weightsModels

Moonshot publishes Kimi K3’s full weights — 2.8T parameters

Moonshot published Kimi K3 on Hugging Face eleven days after its API launch: a 2.8-trillion-parameter MoE with 104B active per token, a 1M-token context window and a native vision encoder. The MXFP4 release is roughly 1.56 TB, and vLLM shipped day-zero support with a Kimi Delta Attention prefix-cache implementation.

The license changed too. K3 uses a bespoke Kimi K3 License, with separate terms for model-as-a-service above $20M trailing revenue and branding obligations above 100M MAU.

WHY IT MATTERS

It resets the open-weight size ceiling, but the serving story matters more than the number: day-zero inference support and prefix caching are what make a model deployable rather than merely downloadable.

Moonshot: Kimi K3Kimi K3 model cardRITS: what it takes to run K3
MON · Jul 27, 2026IndustryPolicy

Nvidia rallies an open-model security alliance

Nvidia launched the Open Secure AI Alliance, arguing open-weight models are defensive assets rather than liabilities. The founding roster included Adobe, CrowdStrike, Hugging Face and Dell, with more than 25 companies in the wider coalition; Anthropic was the notable holdout.

The split is economically legible: chip and platform companies benefit every time another model gets deployed, while frontier labs monetize proprietary access and want tighter controls.

WHY IT MATTERS

Open-weight policy is now a vendor-alignment question. The sign-up list tells you where each company’s revenue comes from — a debate that will shape any US restriction on Chinese open models.

NVIDIA announcementAxios coverageReuters reporting