Open-Weight LLMs in 2026: Local Inference or API?

Treating the summer 2026 open-weight flagships as laptop “local LLMs” will break a technology decision. Kimi K3 (2.8T, 1M context, native vision), GLM-5.2 (1M, MIT), DeepSeek-V4-Flash-0731 (MIT, official vLLM recipe), and Qwen3.8-Max (2.4T / 95B active) are all long-horizon coding and agent candidates. Weight availability and hardware floors are not the same. The first LLM choice is not a single benchmark. Decide whether you will own full weights on a cluster, run a quantized lab copy, or stay on a hosted API.

📑Table of Contents

1. “Local” still means a cluster for full K3
2. Official specs and licenses for four 2026 open-weight LLMs
3. Which self-host path is realistic
4. API prices and how to choose by job
5. Decision checklist and next actions
6. FAQ
7. Summary

---

“Local” still means a cluster for full K3

A single consumer GPU or one Mac Studio cannot load full Kimi K3. In this article, “local” means a self-owned cluster unless you deliberately drop to a quantized lab build.

Total parameters, active parameters, quantized disk size, and KV cache are different quantities. explainx.ai, dated 2026-07-28, puts the MXFP4 safetensors near 594GB and the full-load floor at 8×H100 80GB; it rules out a single consumer GPU or Mac Studio (explainx.ai). Unsloth’s public docs list 1.56TB for full-precision storage, about 1.56TB for lossless Q8, 594GB for Dynamic 1-bit GGUF with roughly 610GB RAM, and about 20 tok/s when the model fits a B200 (Unsloth Kimi K3). Mehul Gupta (2026-07-17) still sees about 1.4TB before KV even in MXFP4/MXFP8 and cites Moonshot’s 64+ accelerator guidance (Medium). Yotta Labs (updated 2026-08-06) repeats 1.56TB and 64+ accelerators for production K3 (Yotta Labs).

Baseten’s 2026-08-13 update points consumer hardware at Qwen 3.6 and Gemma 4 12B, not full K3 (Baseten). If the job must finish on a laptop, look at smaller open models or a split local/cloud agent path such as LM Studio Bionic 初期プレビュー｜ローカルとクラウドで文書・コードを進めるAIエージェント before forcing a Mythos-scale LLM onto a desktop.

Pick one of three layers first:

1. Own the full weights (terabyte-class disk and multi-GPU).
2. Run a quantized lab copy (about 610GB RAM for 1-bit K3).
3. Use APIs only and keep self-hosting as a later experiment.

---

Official specs and licenses for four 2026 open-weight LLMs

Compare scale, context, license, and whether weights are actually public. A vendor Terminal-Bench 2.1 score does not tell you if you can serve the model or redistribute it.

Model	Official scale and context	License / weights	Official positioning	Keep this caveat
Kimi K3	2.8T, 16/896 experts, 1M, native vision	Weights landed 2026-07-26/27. Independent sources call the license Kimi K3 / Modified MIT. Official API: $0.30 hit / $3 miss / $15 out	Open 3T-class. Moonshot says overall it still trails Fable 5 and GPT-5.6 Sol	Thinking history is required. Default effort is max
GLM-5.2	1M with IndexShare. HF lists 753B	MIT on HF / ModelScope	Long-horizon coding. TB 2.1 81.0, SWE-Pro 62.1	Do not collapse HF 753B and Morph 744B/40B
DeepSeek V4 Flash 0731	HF card says 304B. Independents say 284B/13B. 1M lives on the paper side	MIT. DSpark ships in-checkpoint	API public beta on 2026-07-31. TB 2.1 82.7	Official pricing moved to peak/off-peak at 2026-08-16 16:00 UTC
Qwen3.8-Max	2.4T / 95B active. Sample config uses 1M context	Official 2026-08-02 post promised weights “next week.” $2/$6 is third-party	First promised Max-class open-weight drop	Recheck license after the files actually appear

Sources: Kimi K3 blog, Kimi K3 pricing, Z.ai GLM-5.2, Hugging Face GLM-5.2, DeepSeek-V4-Flash-0731, Qwen3.8 blog, MindsHub comparison (as of August 2026).

Moonshot’s post documents KDA, Attention Residuals, Stable LatentMoE (16/896), availability on Kimi.com / Work / Code / API, and full weights by 2026-07-27. It also says quality wobbles if the harness drops thinking history. Z.ai positions GLM-5.2 for long-horizon work, claims a 2.9× per-token FLOP cut at 1M via IndexShare, and publishes MIT weights with no regional restriction. The blog does not print a total parameter count; keep HF 753B and Morph 744B/40B side by side. The DeepSeek card is MIT, includes DSpark, reports TB 2.1 82.7, recommends temperature 1.0, and allows 384K max output at high/max. The card’s 304B label and the independent 284B/13B MoE split should not be merged. Qwen’s 2026-08-02 post is a promise, not proof that files and license terms are already frozen.

A smaller open-weight coding agent is represented by Poolside Laguna S 2.1 — 118B total parameters with about 8B active. Do not put Laguna on the same serving shelf as K3 or Qwen-Max.

Do not treat MIT GLM/Flash and giant K3/Qwen-Max as interchangeable AI SKUs.

---

Which self-host path is realistic

Full K3 wants an 8×H100-class box. Quantized K3 still wants about 610GB. Cheaper failure is to stand up Flash or GLM first.

Full K3 lab

1. Confirm 594GB+ disk.
2. After pip install -U vllm, follow the explainx example: vllm serve moonshotai/Kimi-K3 --tensor-parallel-size 8 --max-model-len 131072 --trust-remote-code.
3. Start at 128k–256k. Measure KV and kernel support before 1M.

Quantized K3 (Unsloth)

1. curl -fsSL https://unsloth.ai/install.sh | sh
2. Open unsloth studio and pick Kimi K3 / UD-IQ1_S (594GB, about 610GB RAM).
3. Use temperature 1.0 and low/high/max thinking. There is no Instant mode.

DeepSeek V4 Flash

1. Use deepseek-ai/DeepSeek-V4-Flash-0731 on Hugging Face.
2. For vLLM, pass --speculative-config '{"method":"dspark","num_speculative_tokens":7,"draft_sample_method":"greedy"}'.
3. The official example is 4×GB300. high/max allow 384K max output.

GLM-5.2

1. Use MIT weights at zai-org/GLM-5.2.
2. Official serving mentions transformers, vLLM, and SGLang. Independent notes put the floor at vLLM 0.23+ / SGLang 0.5.13+.
3. Yotta writes about 8×H200 at FP8.

Date the caveats. explainx’s 2026-07-28 “no faithful GGUF yet” note is already superseded by later Unsloth docs. Qwen3.8-Max weight status and license can drift from the 2026-08-02 post, so refresh Hugging Face before you budget a cluster.

Sources: explainx.ai, Unsloth, DeepSeek HF, Z.ai blog, Yotta Labs.

---

API prices and how to choose by job

If you do not already own a cluster, make the official or third-party API the default path and keep self-hosting as a verification lane. Host prices diverge. Do not lock a budget from a copied table.

Path	Per 1M tokens	Source and date	Use
Official Kimi K3	$0.30 cache hit / $3.00 miss / $15.00 output. Context 1,048,576	Official pricing (checked 2026-08-16)	Long-horizon coding with vision in the loop
Third-party Flash	Fireworks $0.14 / $0.28	MindsHub 2026-08-13, humble 2026-08-10	Cost-first agentic coding. Not the official DeepSeek tariff
Third-party Qwen3.8-Max	$2 / $6	MindsHub / humble	Cloud-first Max. Recheck license after weights land
Third-party GLM-5.2	$1.40 / $4.40	MindsHub 2026-08-13	Contrast price if you also need MIT redistribution
Official DeepSeek	Peak/off-peak from 2026-08-16 16:00 UTC; off-peak is half	DeepSeek API updates	Do not treat $0.14/$0.28 as official

Sources: Kimi pricing, MindsHub, humble weekly, DeepSeek API updates (August 2026).

A practical split:

• Long-horizon coding with vision: run one task on official K3 (default max).
• MIT redistribution or a long-lived self-hosted agent: put GLM-5.2 on the shortlist.
• Cost-first agentic coding: V4 Flash. Independents put it in the cheapest band; refresh official DeepSeek peak/off-peak the same day.
• Cloud-first 2.4T Max: Qwen3.8-Max API. Do not freeze a self-host plan until weights and license are rechecked.

humble’s 2026-08-10 roundup puts Kimi K3 at 57 on the Artificial Analysis Intelligence Index as the open-weight lead, with GLM-5.2 close. MindsHub places K3 about 5% and Qwen3.8-Max about 8% off the then lead. Vendor benches and third-party indexes are weak adoption evidence unless you can replay one task from your own repo.

Some summer 2026 roundups still market this cohort as Mythos-class. Treat that label as marketing, not a hardware spec.

---

Decision checklist and next actions

Do not copy a popularity recap. Decide with one page of execution layer, license, price, and verification.

• ✓ Is the job chat, or a long-horizon coding agent?
• ✓ Full weights (TB-class), quantized (about 610GB for K3), or API only?
• ✓ Can you actually reserve 8×H100, 4×GB300, or 8×H200-class hardware?
• ✓ If you must redistribute, is the license MIT (GLM / Flash) or a K3 / Qwen-specific term?
• ✓ Do you need 1M context locally, or does 128k cover the first task?
• ✓ Did you refresh official quotes (K3 $0.30/$3/$15, DeepSeek peak change) today?
• ✓ Do you have a slot to replay one vendor bench (TB 2.1 or similar) on your own repo?

Next actions:

1. No cluster: run the same task once on K3 or Flash API.
2. Cluster available: load Flash or GLM first; treat full K3 as the second stage.
3. If you try K3, choose Unsloth 1-bit or vLLM TP=8 before downloading.
4. Keep Qwen3.8-Max out of the self-host plan until Hugging Face shows weights and license.

---

FAQ

Q. Can a laptop or Mac Studio run full Kimi K3?

No. explainx rules out a single consumer GPU or Mac Studio and sets the full-load floor at 8×H100 80GB. Even Unsloth 1-bit still wants about 610GB of memory.

Q. Which of the four is easier to self-host?

Relatively, DeepSeek V4 Flash (MIT, official vLLM/SGLang recipe, 4×GB300 example) and GLM-5.2 (MIT, 753B on HF). K3 and Qwen-Max are a different scale.

Q. What is official K3 API pricing?

The official table checked on 2026-08-16 is $0.30 / $3.00 / $15.00 per 1M tokens for cache hit, miss, and output. Context is 1,048,576. Default reasoning_effort is max.

Q. Is GLM-5.2 744B or 753B?

The Hugging Face card says 753B. Morph says 744B/40B. Keep both figures and cite the source instead of rounding to one number.

Q. Are Qwen3.8-Max weights available yet?

The official 2026-08-02 blog said “next week.” Later third-party notes and license details can diverge, so recheck Hugging Face when you plan.

---

Related articles:

• 13558
• 13549
• 13550

Summary

The summer 2026 open-weight flagships are not interchangeable local LLMs. K3 is a capability candidate on a cluster-class serving layer. GLM and Flash are easier MIT self-host options, but official and third-party prices are different objects. Treat Qwen3.8-Max’s 2.4T/95B open-weight promise as a promise until the files and license are rechecked. Use the checklist, then run one identical task on an API or a smaller self-hosted model.