Quick answer: As of July 2026 the open-weight tier splits three ways — Kimi K3 (Moonshot, 2.8T MoE) leads frontend coding, GLM-5.2 (Zhipu, 744B/40B MoE, MIT) tops the open-weights intelligence index and is the value pick, DeepSeek V4 leads raw SWE-bench. For on-device, Qwen3.5-4B and Llama-3.2-3B-Uncensored run on phones. This pillar links the geo/industry spokes. Verify weights shipped before downloading — dates move fast.
Educational disclaimer: Model landscape is fast-moving; figures below verified 2026-07-28. A named model can be stale in weeks — re-check before building on it.
The open-weight leaders (July 2026)
Per Morph's coding-model comparison and Artificial Analysis:
| Model | Params | Coding | Context | License | Weights |
|---|---|---|---|---|---|
| Kimi K3 | 2.8T MoE | #1 Frontend Arena 93.4% | 1M | Modified MIT (exp) | due ~Jul 27 |
| GLM-5.2 | 744B/40B MoE | #1 open Intelligence Index | 1M | MIT | ✅ downloadable |
| DeepSeek V4 | — | #1 SWE-bench Verified | — | open | ✅ |
| Qwen3.5 | 4B–max | strong, proprietary Max | — | Apache (some) | ✅ small sizes |
GLM-5.2 (Zhipu, released Jun 16 2026) is the value pick: ~$1.10/M in, $4.10/M out on some hosts — ~¼ of frontier price for leading open-weights intelligence (Morph). Kimi K3 (Moonshot, Jul 16 2026) leads frontend but was API-only until weights shipped ~Jul 27 — confirm before self-hosting.
The country race
- China — Zhipu (GLM), Moonshot (Kimi), Alibaba (Qwen), DeepSeek. Open-weights strategy dominates; MIT/Apache licenses widen adoption.
- USA — frontier closed (GPT-5.5, Opus-4.8); open via Meta Llama, Mistral.
- India — adoption + infra growing (India's AI Moment); few foundation models, strong application layer.
See spokes: The AI Gap by Country, Country-Wise AI Report.
What runs on your hardware
| Hardware | Best local model | Quant | Speed |
|---|---|---|---|
| Phone 8 GB | Qwen3.5-4B | Q4_K_M | 8–15 t/s |
| Laptop 16 GB | 7–8B (Qwen3.5/Llama) | Q4_K_M | 20–40 t/s |
| Desktop 24 GB VRAM | 70B MoE | Q4 | 30+ t/s |
| Server 80 GB VRAM | GLM-5.2 / DeepSeek V4 | Q4 | fast |
Phone picks: Llama-3.2-3B-Uncensored (abliterated, ~2.2 GB) and Qwen3.5-4B (coherent, multilingual).
Spoke: The Search Engine Gap
AI Overviews miss what matters when models are stale or context-poor. Our Search Engine Gap piece covers why GEO (generative-engine optimization) needs clear H1 identity + structured answers — exactly what these pillars do.
Spoke: Industry moves
- Nemotron 3 Ultra & Agentic Chip Design
- Nvidia Cosmos World Model
- Hugging Face vs Civitai
- Humans Upgraded, not Replaced
Why open weights matter for you
- Self-host — no per-token bill, no data leaving.
- Pin & audit — you control the version; APIs change under you.
- Fine-tune — QLoRA on your data (see Tune Local AI).
- Compliance — regulated fields need on-prem; open weights enable it.
Freshness warning (read this)
Model names go stale in weeks. By the time you read this, K3 weights may have shipped (or not), GLM-5.3 may exist. Always:
- Verify weights are actually downloadable (HuggingFace/MODELSCOPE), not "coming soon."
- Check license for commercial use (MIT = free; Modified MIT may restrict).
- Re-run a tiny benchmark on your task before committing.
How to choose — a decision tree
Need frontier reasoning (math, long agents)?
├─ yes → API (GPT-5.5 / Opus-4.8) OR desktop 70B MoE
└─ no → open weights?
├─ coding → GLM-5.2 (value) or DeepSeek V4 (SWE) or K3 (frontend)
├─ private/docs → 7B + RAG on your hardware
├─ phone → Qwen3.5-4B Q4 / Llama-3.2-3B-Uncensored
└─ agentic → K3 (1M ctx) when weights ship
Don't pick by leaderboard alone — pick by your constraint (cost, privacy, hardware, license).
Benchmark methodology (don't trust one number)
A model "wins" only on the benchmark that matches your use:
- SWE-bench Verified → code-gen capability (DeepSeek V4 leads).
- Frontend Code Arena → UI building (K3 #1 at 93.4%).
- Artificial Analysis Intelligence Index → general open-weights smarts (GLM-5.2 #1).
- MMLU / GPQA → knowledge/reasoning (frontier still ahead).
Run a 10-task mini-bench on your workload before committing. Leaderboards lie for narrow use.
The licensing trap
"Open weights" ≠ "free for commercial." Watch:
- MIT (GLM-5.2) → commercial OK.
- Modified MIT (Kimi K3 line) → check regional/restriction clauses.
- Apache 2.0 (some Qwen) → commercial OK.
- Community license (Llama) → revenue cap unless approved.
If you ship a product, read the license. "Open" often means "open for research."
Inference cost table (real numbers, July 2026)
| Model | Hosted $/M out | Self-host (24GB VRAM) |
|---|---|---|
| GLM-5.2 | ~$4.10 (Morph) | power only (~₹500/mo) |
| Frontier GPT-5.5 | $15+ | n/a (API only) |
| DeepSeek V4 | ~$2–4 | power only |
| Qwen3.5-4B | n/a (small) | phone, free |
Self-hosting wins at volume. At 50k tokens/day, API runs ₹1k–3k/mo; your GPU is sunk cost + power.
What's next (H2 2026 watchlist)
- K3 weights ship + ecosystem builds (agents on 1M ctx).
- GLM-5.3 / Qwen3.6 refreshes.
- More 3B–4B phone-grade models with reasoning.
- Agentic chip design (Nemotron 3 Ultra) → faster local inference.
- India's application layer deepening; foundation models still thin.
Re-verify this list quarterly — the field moves faster than docs.
Local vs API — final call
| Factor | Local (open weights) | API (frontier) |
|---|---|---|
| Cost at scale | cheap (own HW) | per-token |
| Privacy | total | vendor sees prompts |
| Control | full | locked persona |
| Peak reasoning | 70B MoE | best |
| Setup | dev comfort | zero |
Pick local for ownership/privacy/cost; API for max reasoning with zero setup.
The agentic shift (why 1M context matters)
2026 is the year of agents: models that plan, call tools, and loop for minutes. That needs long context (K3/GLM-5.2 both 1M) and code/tool use. A 4B phone model can't agent well; a 70B MoE can. If your use is "answer a question," small is fine. If it's "build my app," you need the big open weights or frontier API.
Model size vs quality — the real curve
- <3B: chat, simple tasks, phone. Quality drops on reasoning.
- 7–8B: daily driver, good RAG, decent code.
- 70B MoE: near-frontier on most tasks, runs on 24GB VRAM (Q4).
- Frontier (closed): still best at novel reasoning, but you rent it.
The "sweet spot" for self-hosters in 2026 is 70B MoE at Q4 — desktop-class quality, consumer hardware. Phones stay at 4B.
How we verified (sources, not memory)
- Model table: Morph coding comparison (Jul 2026) + Artificial Analysis GLM-5.2 vs Qwen3-Max.
- Weights dates: Medium Kimi K3 vs GLM-5.2 (Jul 16/17 2026).
- On-device picks: our own Termux/Ollama testing (Ollama 0.30.10, 4B Q4, 8–15 t/s).
- Freshness gate: re-checked named entities this session (2026-07-28), not from cached text.
Risk of vendor lock
APIs change: price, rate limit, model deprecation, ToS. If your product depends on GPT-5.5 and it doubles in price or adds a clause, you're stuck. Open weights (especially MIT like GLM-5.2) let you pin a version and self-host — no surprise. For a business, that resilience is worth the setup cost.
What this means for a builder in India
India's strength is the application layer (India's AI Moment): UPI-scale distribution, vernacular reach, cost sensitivity. Open weights (MIT GLM-5.2, Apache Qwen) let Indian builders self-host cheaply — no dollar-denominated API bill, rupee infrastructure. A 4B model on a ₹15k phone serving Hinglish is a real product here, not a lab demo.
The gap: few Indian foundation models. That's fine — application + fine-tune on open weights is the pragmatic path. Our own stack (NIFTY engine, BTC flow) runs on open-source + public APIs, proving the model.
FAQ
Q: Which open model is "best"?
A: Depends. Frontend → K3. Value/intelligence → GLM-5.2. SWE-bench → DeepSeek V4. On-device → Qwen3.5-4B.
Q: Is Kimi K3 downloadable?
A: Weights were due ~Jul 27 2026. Verify on HuggingFace before assuming; until then API-only.
Q: Why not just use GPT-5.5?
A: If you need frontier reasoning and don't mind cost/logging, fine. Open weights win on ownership, cost, privacy.
Q: Best for a phone?
A: Qwen3.5-4B Q4 (2.7 GB) or Llama-3.2-3B-Uncensored (2.2 GB). Both chat smoothly on 8 GB RAM.
Q: How do I verify a model is real, not rumored?
A: Check HuggingFace/MODELSCOPE for actual weights + download count. "Coming soon" = not shippable. Our skill flags rumored-as-shipped as a correction risk.
Q: What hardware for agentic work?
A: 24GB VRAM desktop running 70B MoE Q4, or frontier API. 4B phone models can't agent reliably.
Q: Why does this matter for Indian builders?
A: Open MIT/Apache weights = self-host in rupees, no dollar API bill. 4B Hinglish on a budget phone is a shippable product.
Key takeaways
- July 2026: K3 (frontend), GLM-5.2 (value/intel), DeepSeek V4 (SWE) lead open.
- GLM-5.2 is MIT, downloadable, ~¼ frontier price — the value pick.
- China drives open-weights; India strong on application layer.
- Phones run 4B Q4 today; desktop runs 70B MoE.
- Verify weights + license before building; names stale fast.
- Pick by constraint (cost/privacy/hardware), not leaderboard.
- Re-verify quarterly — the field outruns docs.
- Agentic work needs 1M ctx + 70B MoE; vendor lock is the real API risk.
- India's edge = application + open-weights fine-tune, not foundation models.
Written by **Shakti Tiwari* — Nifty Option Trader, XGBoost Expert. AI + trading coverage at optiontradingwithai.in.*
Books: Option Trading with AI (B0H9ZNTBPK) | The AI Opportunity (B0HBBFKDQF)
Top comments (0)