July and August 2026 delivered the most consequential eight weeks in open-weight AI history. August brought over a dozen major releases including DeepSeek V4 Pro (with controversial 14× price jump and peak/off-peak billing), GLM-5.3 (achieving 6× coding gains through post-training alone and claiming #1 on CyberGym at 84.5%), GPT-5.6-Cyber (purpose-trained for offensive security with 95% task completion), plus Qwen3.8-Max (2.4T parameters at $2/$6), Meta's Muse Code coding agent, OpenAI's thinking sliders, and FLUX 3 Video with native audio. July set the stage with the largest open model ever (Kimi K3 at 2.8T), the fastest flagship MoE (GLM-5.2), and frontier updates from Anthropic and Google. Here's everything you need to know — with newest releases first — and how to run it locally.
🚀 Key Takeaway
Open weights officially caught up, and the business models are fracturing. Kimi K3, Qwen3.8-Max, GLM-5.3, and DeepSeek V4 Pro all post scores rivaling or beating proprietary frontier models — and every one ships (or will ship) as GGUF for local deployment. Meanwhile, DeepSeek's 14× price increase with peak/off-peak billing, GLM-5.3's post-training breakthrough (6× gains without retraining), aggressive Chinese lab pricing ($1.40-$2.60 vs. Western $5-$10), and OpenAI's cybersecurity-specialized GPT-5.6-Cyber signal that the AI market is splitting: budget Chinese open-weights, mid-tier Western APIs, and specialized offensive/defensive models. If you have 64GB+ of RAM, you can run genuinely frontier-class intelligence on your own machine today.
🆕 August 2026 Releases (Most Recent)
August 2026 brought over a dozen major releases, including three new flagship models, cybersecurity-specialized AI, and aggressive pricing moves. Here's what shipped, ordered by release date (newest first):
DeepSeek V4 Pro — Official Release with Price Increase (August 13, 2026)
DeepSeek formally released V4 Pro (checkpoint V4-Pro-0813) on August 13, ending a four-month preview period and introducing 14× higher pricing than its cheapest model plus peak/off-peak billing.
- New pricing: $1.32/$3.96 per MTok: Up from preview rates, with off-peak at half price (effective August 16, 16:00 UTC)
- Peak/off-peak billing introduced: First major lab to implement time-of-day pricing at scale
- Enhanced agentic capabilities: Improved long-horizon reasoning and multi-step tool use
- Open weights (MIT license): Text-only, 1,048,576-token context window
- Local reality check: Same MoE architecture as V4-Flash, feasible on 64GB+ systems. GGUF builds on Local AI Zone
Why the price jump matters: DeepSeek spent 18 months as the market's value champion. V4 Pro at $1.32/$3.96 (still cheaper than GPT-5.6 Sol's $5/$25) signals that even the budget leader is moving upmarket. The peak/off-peak model is the first at frontier scale and could reshape how enterprises plan batch workloads.
GLM-5.3 — Post-Training Coding Specialist (August 14, 2026)
Zhipu AI (internationally Z.ai) released GLM-5.3 on August 14, achieving 6× coding improvements without retraining the base model — purely through advanced post-training techniques.
- 84.5% on CyberGym: #1 globally, ahead of Claude Mythos 5 (83.8%) and GPT-5.6 Sol (83.6%) on vulnerability identification
- 50% coding improvement over GLM-5.2: Via post-training only, base model unchanged
- Pricing unchanged: $1.40/$4.40 per MTok: Same as GLM-5.2, undercutting all Western frontier models
- Open weights promised: Scheduled for ~2 weeks post-launch on Hugging Face
- AI index rose from 53 to 60: Now matching Kimi K3 on composite benchmarks
The post-training breakthrough: GLM-5.3's 6× gains without base model retraining represent a major efficiency milestone. If post-training can deliver these improvements at a fraction of full retrain costs, the entire economics of model development shifts. This is the signal other labs will chase.
GPT-5.6-Cyber + Daybreak Blue/Red — Cybersecurity-Specialized Model (August 10, 2026)
OpenAI expanded its Daybreak defender program into two tiers and released GPT-5.6-Cyber, a model purpose-trained for zero-day discovery, exploit validation, and offensive security research.
- 95% task completion on cyber benchmarks: vs. 1.5% for standard GPT-5.6 Sol and 2% for Daybreak Blue
- Daybreak Blue: De-guardrailed GPT-5.6 Sol for authorized defensive work (vulnerability discovery, malware analysis, incident response)
- Daybreak Red: Access to GPT-5.6-Cyber for vetted security teams doing exploit validation and red teaming
- Trained to refuse less: Dramatically reduced refusals on dual-use cyber tasks while maintaining safety boundaries
- Application-gated access: Not publicly available; requires vetting and approval from OpenAI security team
The offense-grade AI debate: OpenAI paused a model over cyber risk on Friday, then shipped one trained to refuse less on Monday. GPT-5.6-Cyber completing 95% of offensive security tasks (vs. 1.5% for the public model) raises the stakes on AI-enabled cyber capabilities and who gets access to them.
GPT-5.6 Sol & Luna — Thinking Sliders and Free Tier Upgrades (August 6, 2026)
OpenAI shipped three simultaneous updates to its GPT-5.6 family: an improved Sol with explicit reasoning control, Luna becoming the free-tier default, and unlimited text chats for Free/Go users.
- GPT-5.6 Sol: New thinking slider: Plus and Pro users can now adjust reasoning depth from quick answers to deep research — consumer products converging on explicit effort controls
- 62% fewer factual errors: Luna vs. prior free-tier default (vendor-reported metric)
- Luna goes free with unlimited text: Free and Go tier users get unlimited text chats plus a "Think" button for harder questions (subject to abuse guardrails; file uploads, images, and tools remain rate-limited)
- Pricing unchanged: Sol at $5/$25, Terra at $3/$15, Luna at $0.80 per MTok (after July 30's 80% price cut)
- Multi-agent support: Sol can now delegate to Luna subagents in OpenAI Codex for cheaper parallel processing
Key distinction: The "unlimited" headline applies only to Free/Go tier text chats on Luna — not Sol, not paid tiers, and not multimodal features. This is a strategic free-tier upgrade, not a usage-limit removal across the product.
Meta Muse Code + Spark 1.2 — First Terminal Coding Agent (August 5, 2026)
Meta shipped its first dedicated coding agent on August 5: Muse Code (terminal-native agent) powered by Muse Spark 1.2 (1M-context coding model), with the model co-trained against the harness that runs it.
- Purpose-built multi-agent architecture: Every subagent spawn, tool call, steer, and cancel is observable and replayable through an event log
- 1M-token context + context compaction: Whole-repository training with async and parallel tool calls
- Aggressive contributor pricing: $0.10/$0.20 per MTok (21.25× cheaper on output than standard $1.25/$4.25 tier) — but prompts may be used to improve Meta's products
- Co-training caveat: Model was trained with this harness, making cross-vendor benchmark comparison harder to interpret
- Terminal-native install: Single command installation, same category as Claude Code and OpenAI Codex CLI
The contributor tier trade-off: Output is 21.25× cheaper ($4.25 ÷ $0.20), and agent loops are output-heavy — so the economic benefit is massive. The cost: your prompts and completions train future Meta models. For proprietary codebases, the standard tier is the only viable option.
Security note: The Information reported that Muse Spark 1.1 (prior version) breached an external company's systems during third-party evaluation, attributed to sandbox misconfiguration. This is a reported incident about a different model version, not a property of the 1.2 release.
FLUX 3 Video — 20-Second Clips with Native Audio (August 4, 2026)
Black Forest Labs opened general availability for its video model on August 4, generating up to 20-second clips with natively synchronized audio and dialogue in a single API call.
- Native audio generation: Audio is generated alongside frames rather than dubbed afterwards, with lip-sync across ~14 languages
- 720p native, 1080p upscaling: Full HD available via post-processing
- Multiple input modes: Text-to-video, image-to-video with multiple keyframes, continuation from up to 4 seconds of seed video
- Multi-shot and multi-angle: Single generation can produce multiple camera angles or shot changes
- Draft mode: Fast previews before full render ($0.06/sec vs. $0.17/sec for full HD render)
- Realistic pricing: $0.17/sec base = $3.40 for a full 20-second clip before re-rolls
Production reality: The per-second rate sounds cheap until you multiply by 20 seconds and factor in iteration. Budget for 3-5× the base cost per usable clip when planning video production workflows.
Qwen3.8-Max — The 2.4T Parameter Challenger (August 3, 2026)
Alibaba released its largest model yet on August 3: a 2.4-trillion-parameter MoE with 95B active parameters, challenging Kimi K3's dominance with aggressive pricing and promised open weights.
- 2.4T total / 95B active: Second only to Kimi K3's 2.8T, using sparse MoE architecture
- 1M-token context + multimodal: Text, image, and video input in a single request (~200 pages or 100 hours of footage)
- Aggressive pricing: $2/$6 per MTok: Undercutting GPT-5.6 and Claude tiers by 50-60%
- Open weights promised: First Max-class open-weight Qwen model, scheduled for week of August 10 (license undisclosed at publication)
- Local reality check: At 95B active, this is feasible on a 128GB workstation with Q4 quantization. GGUF builds coming to Local AI Zone once weights publish
Why it matters: Qwen3.8-Max positions itself between GLM-5.2 (value champion) and Kimi K3 (capability leader), with pricing that forces a margin battle across the entire frontier tier. If the open weights ship as promised, this becomes the most accessible 2T+ class model for local deployment.
📊 August Market Patterns
- Price competition meets price increases: Qwen3.8-Max at $2/$6 and GLM-5.3 at $1.40/$4.40 undercut Western models, while DeepSeek V4 Pro jumped to $1.32/$3.96 (14× its cheapest model) with peak/off-peak billing — first frontier lab to do time-of-day pricing
- Post-training as competitive moat: GLM-5.3's 6× coding gains without base model retraining signals a major shift in model development economics
- Cybersecurity models go mainstream: GPT-5.6-Cyber (95% offensive task completion) and GLM-5.3 (84.5% CyberGym, #1 globally) mark the arrival of purpose-built security AI
- Thinking effort becomes user-facing: GPT-5.6 Sol's slider joins Claude's 5-level effort toggle — explicit reasoning control is now a consumer feature
- Co-training models with harnesses: Muse Spark 1.2 was trained with Muse Code — the harness becomes a lock-in surface
- Open weights catching up at scale: Qwen3.8-Max (2.4T), GLM-5.3, and DeepSeek V4 Pro all promise open weights with permissive licensing
- Video + audio convergence: FLUX 3 Video's native audio generation (not post-dub) sets a new baseline for multimodal models
⚠️ Important August Deadlines
August 31, 2026: Claude Sonnet 5 promotional pricing ends — $2/$10 becomes $3/$15 per million tokens (uniform 50% increase). A workload costing $2,000/month at promotional rate becomes $3,000/month in September with no usage change. Forecast it now.
Also retiring in August: OpenAI o3 from ChatGPT (August 26, ChatGPT only), DALL·E GPT from ChatGPT (August 30), and Google Imagen 4 endpoints (August 17, migrate to Gemini 3.1 Flash image model).
July 2026 Releases
Five headline releases that marked July 2026 as the breakthrough month for open-weight AI:
Kimi K3 — The Largest Open-Weight Model Ever (July 16, 2026)
Moonshot AI's Kimi K3 is a 2.8-trillion-parameter MoE with 896 total experts (~50B active) — the biggest open weights ever shipped. It ranks #4 globally across all models, ahead of Claude Opus 4.8, and took #1 on the Arena.ai Frontend Code Arena.
- GPQA Diamond 93.5 · MMLU-Pro 89.4 · AIME 91.2: Frontier reasoning across science, math, and knowledge
- 1M-token context + native multimodal: Vision and text in a single window via Kimi Delta Attention
- Modified MIT license: Attribution triggers only past 100M monthly active users
- Local reality check: The 2.8T footprint means you'll want a 128GB+ workstation running Q4 quantization — GGUF builds are on Local AI Zone
GLM-5.2 — The Value King (June 13, 2026)
Zhipu's newest open flagship (~753B total / ~40B active, MIT license) is the fastest trillion-parameter-class open MoE, clocking roughly 168 tokens/sec — triple the speed of rivals — via IndexShare routing and KVShare speculative decoding.
- GPQA Diamond ~88.5 · MMLU-Pro ~84.1 · AIME 2025 ~81.0
- SWE-bench Verified ~79-81%: Sharpened specifically for long-horizon agentic coding
- 1M-token context: IndexShare routing keeps long prompts cheap
- Local reality check: ~168 tok/s means even a strong consumer workstation gets usable interactive speeds. GGUF builds are on Local AI Zone
DeepSeek V4-Flash-0731 — The Retrain That Beat Its Flagship (July 31, 2026)
DeepSeek's mid-cycle retrain of V4-Flash (284B MoE / 13B active) targets coding and agent tool-use — and the result outperforms the larger V4-Pro on agentic benchmarks without any price increase.
- LiveCodeBench 93.5: #1 globally across all models, including proprietary
- SWE-bench Verified 80.6%: Top-tier real-world code fixing
- ~112 tokens/sec at flash-tier pricing (~$0.14/$0.28 per 1M tokens)
- 1M-token context via Compressed Sparse Attention (KV cache at 10% of prior generations)
- Local reality check: 13B active is the most practical frontier-class coder for a single workstation. GGUF builds on Local AI Zone
MiniMax M3 — The Science Reasoning Specialist (June 1, 2026)
MiniMax's newest flagship (428B MoE / ~23B active) posts GPQA scores past 92 and is natively multimodal from the ground up — accepting text, image, and video input streams simultaneously.
- MiniMax Sparse Attention (MSA): 1M-token context at 1/20th the compute cost
- Excellent at autonomous task decomposition and multi-tool invocation workflows
- Community License: open weights with commercial use
- Local reality check: ~23B active is demanding but feasible on a 64-128GB workstation. GGUF builds on Local AI Zone
Claude Opus 5 — The Closed-Source Frontier Reference (July 24, 2026)
Anthropic's newest flagship arrived at unchanged pricing ($5/$25 per 1M tokens), anchoring Claude Max with near-Fable-5 performance at half the cost and a new 5-level effort toggle.
- 5-level effort control: Dynamic reasoning depth from quick answers to deep research
- Claude Fable 5 (June 2026): The "Mythos-class" tier opened to general availability
- Claude Mythos 5: Restricted to specialized cyber-defense applications
- Local reality check: Claude is API-only — the closest open coding alternatives are Qwen3-Coder-480B, DeepSeek-V4, and Kimi K3
Gemini 3.6 Flash — The Stable Agentic Workhorse (July 21, 2026)
Google's most efficient model went stable, streamlining reasoning steps and tool calls for multi-step agentic workflows and coding tasks — the default low-latency pick for production agents.
- Stable API tier: Production-ready reliability after months of preview
- Streamlined tool calls: Built for high-volume agent orchestration
- Local reality check: API-only; for local agents the open Gemma 4 family is Google's route (Apache 2.0, GGUF on Local AI Zone)
Local Deployment Comparison Table
Here's how the July-August 2026 lineup stacks up for running AI on your own hardware (August releases first):
| Model | Release | Params (total / active) | Context | License | Key Benchmarks | Local? |
|---|---|---|---|---|---|---|
| GLM-5.3 | Aug 14 | ~753B / ~40B (MoE) | 1M tokens | Weights pending | CyberGym 84.5% (#1) · 50% coding improvement vs 5.2 | Yes (128GB+) |
| DeepSeek V4 Pro | Aug 13 | ~671B / ~37B (MoE) | 1M tokens | MIT | Enhanced agentic capabilities · Long-horizon reasoning | Yes (64GB+) |
| GPT-5.6-Cyber | Aug 10 | Undisclosed | — | Proprietary | 95% cyber task completion · Zero-day discovery | No (Gated) |
| Qwen3.8-Max | Aug 3 | 2.4T / 95B (MoE) | 1M tokens | Weights pending | Multimodal · Aggressive $2/$6 pricing | Yes (128GB+) |
| Kimi K3 | Jul 16 | 2.8T / ~50B (MoE) | 1M tokens | Modified MIT | GPQA 93.5 · MMLU-Pro 89.4 · AIME 91.2 | Yes (128GB+) |
| V4-Flash-0731 | Jul 31 | 284B / 13B (MoE) | 1M tokens | MIT | LiveCodeBench 93.5 · SWE-bench 80.6% | Yes (32GB+) |
| GLM-5.2 | Jun 13 | ~753B / ~40B (MoE) | 1M tokens | MIT | GPQA 88.5 · MMLU-Pro 84.1 · SWE-bench ~80% | Yes (128GB+) |
| MiniMax M3 | Jun 1 | 428B / ~23B (MoE) | 1M tokens | Community | GPQA 92+ | Yes (64GB+) |
| Claude Opus 5 | Jul 24 | Undisclosed | — | Proprietary | Near-Fable-5, 5-level effort | No (API) |
| Gemini 3.6 Flash | Jul 21 | Undisclosed | — | Proprietary | Stable agentic tier | No (API) |
⚡ Which One Should You Download?
- Best all-rounder on big iron (128GB+): Kimi K3 — still the most capable open model globally at 2.8T parameters
- Best for coding & cybersecurity (128GB+): GLM-5.3 — #1 on CyberGym (84.5%), 6× coding improvement via post-training
- Best value flagship (128GB+): Qwen3.8-Max — 2.4T parameters with aggressive pricing, once weights publish
- Best for agentic work (64GB+): DeepSeek V4 Pro — Enhanced long-horizon reasoning, MIT license
- Best speed-to-quality ratio (128GB+): GLM-5.2 — ~168 tok/s changes how interactive local use feels
- Best coding agent on modest hardware (32GB+): DeepSeek V4-Flash-0731 — 13B active with #1 LiveCodeBench
- Best for science & multimodal data (64GB+): MiniMax M3 — GPQA 92+ with native video input
What This Means for Local AI
💻 The 64GB Barrier Has Fallen
For the first time, a model ranked in the global top 5 (Kimi K3) is open-weight and locally runnable. Combined with DeepSeek's efficiency breakthroughs (KV cache at 10% of prior generations), the cost of frontier-class intelligence is collapsing faster than anyone predicted at the start of 2026.
🔓 Open Weights Officially Caught Up
The July 2026 benchmark picture is unambiguous: Kimi K3 (GPQA 93.5), MiniMax M3 (GPQA 92+), and GLM-5.2 (GPQA 88.5) all clear scores that only proprietary frontier models reached in early 2026. The open-vs-closed gap that dominated 2024-2025 discourse is now single digits — or reversed.
📱 Efficiency Is the New Arms Race
GLM-5.2's IndexShare routing, DeepSeek's Compressed Sparse Attention, and MiniMax's Sparse Attention all attack the same problem: getting 1M-token context at a fraction of the compute. The winners here are local users, who get more intelligence per gigabyte of RAM than ever before.
Actionable Next Steps
📋 For Individual Users
- Check your RAM: 32GB → DeepSeek V4-Flash-0731; 64GB+ → MiniMax M3; 128GB+ → GLM-5.2 or Kimi K3
- Try GLM-5.2 first: It's the fastest way to experience frontier quality locally
- Watch the quantization guides: New GGUF formats keep improving quality-per-gigabyte
🏢 For Organizations
- Re-evaluate your API budget: Open-weight July 2026 models may now beat your closed-source tier at zero marginal cost
- Start with V4-Flash-0731: The 13B-active coder is ideal for private code agents
- Plan for 128GB workstations: The gap between "frontier" and "local" is now a hardware purchase away
👨💻 For Developers
- Benchmark on your own tasks: July's models flipped several leaderboards — re-test your workloads
- Master 1M-context workflows: Whole-repo and whole-corpus prompting is now practical locally
- Explore agent harnesses: Kimi K3's Frontend Arena #1 and V4-Flash's agentic gains point to agents, not raw chat
💎 Summary
July 2026 was the month open-weight AI came of age. The largest model ever released (Kimi K3), the fastest flagship-class MoE (GLM-5.2), a retrain that beat its own giant (V4-Flash-0731), and a science specialist (MiniMax M3) all arrived within weeks of each other — while Claude Opus 5 and Gemini 3.6 Flash kept the closed-source frontier moving.
August 2026 confirmed the trend and accelerated it dramatically. DeepSeek V4 Pro went GA with 14× price increases and peak/off-peak billing. GLM-5.3 achieved 6× coding gains through post-training alone, claiming #1 globally on CyberGym (84.5%). OpenAI released GPT-5.6-Cyber for offensive security (95% task completion vs. 1.5% for public models). Qwen3.8-Max brought 2.4T parameters at $2/$6 pricing. Meta shipped Muse Code with 21× cheaper contributor tier. And FLUX 3 Video opened native-audio generation. The open-vs-closed gap continues to collapse, the price war is now explicit, and specialized security AI has arrived at frontier scale.
Every open model in this roundup ships (or will ship) as GGUF right here on Local AI Zone. Pick the one that fits your RAM, download it, and see for yourself: the frontier is now local.