Five Open-Weight Models, Three Run Paths: A Buyer's Ledger for November 2026

Mistral Large 4 scores 48% on the Artificial Analysis Coding Agent Index, effectively tied with GLM-5.3, while using 47% fewer tokens per task — by Mistral's own account. That pairing of numbers, quality parity plus token efficiency, anchors the November 13, 2026 comparison of five open-weight model families: Kimi K3.1 Pro, DeepSeek V4.1 Flash, Mistral Large 4, GLM-5.3 and Gemma 4. They share a label and almost nothing else.

What does 'open source AI model' actually mean here?

Strictly, almost none of these releases are open source. The software-industry bar covers code, data and the freedom to modify and redistribute. What labs ship is closer to open weights (downloadable parameters under a license that grants some rights and withholds others). Moonshot AI officially calls Kimi K3.1 Pro open-weight; Mistral calls Large 4 open weight. Neither claims a fully open training pipeline.

License evidence varies sharply. Mistral's announcement names the Mistral Ecosystem Open-Weight License v26.10, described by Mistral as a modified MIT license; a third-party comparison adds a model-as-a-service clause and resale restriction. Moonshot's license text isn't reproduced in confirmable sources. DeepSeek V4.1 Flash is the awkward case: official pages confirm the model, API identifier and pricing, but not a downloadable weight release. A third-party label of MIT rests on a single source and may tie to DeepSeek's Harness tooling, not the weights. Google's Gemma 4 blog post positions the family as open, but no license text surfaced. Z.ai's documentation says only 'open-source SOTA coding capabilities.' The lesson: confirm the license file in the exact repository, and keep a copy in your compliance folder.

What does each model cost?

Published per-direction rates, per million tokens:

• Mistral Large 4: $0.70 input, $0.07 cached, $2.10 output (vendor page)
• GLM-5.3: $1.40 / $0.26 / $4.40; GLM-5.3-FlashX $0.37 / $0.075 / $1.25; GLM-5.3-Flash $0.15 / $0.03 / $0.50 (Z.ai pricing page)
• DeepSeek V4.1 Flash: $0.525 blended (third-party tracker; off-peak runs 50% of peak)
• Kimi K3.1 Pro: $4.00 blended (tracker, unconfirmed on any official Moonshot page)
• Gemma 4: not confirmed

GLM-5.3-Flash at $0.15 input and $0.50 output is the cheapest verified route. Mistral's regional inference (EU or US endpoints) carries a 10% upcharge — a concrete data-residency lever the other four don't document.

What are the specs that matter?

• Kimi K3.1 Pro (Moonshot, launched Oct. 23, 2026): 3T-class parameters, 1M-token context, multimodal. The Highspeed variant is advertised at 2-3x faster; treat that as a vendor claim. API access unlocks after a $1 minimum top-up, and temperature, top_p and penalty parameters are fixed — you can't tune sampling.
• Mistral Large 4 ('Le Chonk', preview Oct. 6, open weights Oct. 29): about 1T total, ~49B active parameters, natively multimodal, positioned for agentic coding and cybersecurity. Weights on Hugging Face, serving via vLLM and SGLang. The preview is retired; anyone on preview weights should move.
• DeepSeek V4.1 Flash (released Sept. 10, 2026): identifier deepseek-flash, always points at the latest version. Web Search and Fetch cost $0.75 per 1,000 successful requests each. V4 Flash and V4 Flash Vision Exp were retired at launch.
• GLM-5.3 (Z.ai, documented current by Nov. 2): 1M-token context, built on the same base model as GLM-5.2 with gains from post-training — so no new hardware needs for existing GLM hosts. FlashX claims 200 tokens per second.
• Gemma 4 (Google, blog dated Apr. 2, 2026): trained on 140+ languages, the only family here a developer can plausibly run on a workstation.

What changed under the hood without anyone asking?

Silent rerouting already happened this autumn. Z.ai now sends GLM-5.2 and GLM-5.1 requests to GLM-5.3, and GLM-4.7 requests to GLM-5.3-Flash. If your config names an older GLM, you already run the new model. DeepSeek did the same with legacy identifiers. Pin behavior tests to the served model, not the name in your config.

Kimi K3.1 Pro also moved fast into catalogs: Microsoft published a Foundry route via Fireworks (November 10 listing, date unconfirmed), Amazon Bedrock reportedly carries the Highspeed variant, and NVIDIA's API docs carry a reference page. For procurement, that means skipping a vendor security review; the trade is price and control.

Which model for which job?

• Regulated company needing a license trail: Mistral Large 4 — the only named, versioned license in the group.
• Tight token budget, flexible timing: GLM-5.3-Flash or DeepSeek V4.1 Flash (off-peak halves rates; Z.ai subscriptions consume 50% of standard points off-peak and on weekends; Lite/Pro/Max plans reportedly start at $18/month, single source).
• Token-hungry agent loops: Mistral Large 4 or GLM-5.3.
• Long documents: Kimi K3.1 Pro or GLM-5.3, both at 1M tokens.
• Locked into one cloud: whatever your hyperscaler lists.
• Hardware-limited: Gemma 4, after checking the model card.

One recommendation the evidence doesn't support: self-hosting DeepSeek V4.1 Flash. Its weights are unconfirmed. Use it as a hosted API.

What's the verdict?

Mistral Large 4 is the safest pick for teams that must document what they run — named license, published prices, downloadable weights, the best-documented benchmarks. Its Dense 200 visual grounding win over GPT-6-Astra (42% to 41%) is inside plausible noise. GLM-5.3 takes second on value; Kimi leads on scale and catalog reach; Gemma 4 wins only where hardware is the limit.

Practical advice for the next 30 days: shortlist two models, replay 30-50 of your own prompts, buy through the route compliance already approves, and re-read the license on every point release. Watch three open items: a DeepSeek V4.1 Pro launch, published model-card licenses for DeepSeek, Gemma and GLM, and Moonshot's still-unpublished per-direction prices.