Which Open Model Should You Use?

Snapshot: 7 November 2026. Model releases now arrive weekly, so every figure below carries a date and a source. Where a number is the developer's own claim rather than an independent measurement, we say so. Where we could not verify something, the table says "not verified" rather than guessing.

There are now more serious open-weight language models than any team can evaluate properly. Most comparisons respond to that by ranking them on a single leaderboard score, which is a bit like choosing a car by its top speed. This guide does it differently. It starts with a plain answer for people who just want to know what to use, then adds one layer of detail per section, ending with the arithmetic that decides what will actually run on your hardware. Read until you have your answer, then stop.

Part 1: The sixty-second answer

Find the line that describes you.

• "I want a private assistant on my laptop." Use a model of 4 to 9 billion parameters. With 8 GB of memory, try Qwen3.5 4B or Qwen3 4B Instruct 2507. With 16 GB, try Qwen3.5 9B, Ministral 3 8B or Apertus 8B.
• "It must be Swiss or European, and I need to explain to an auditor what it was trained on." Use Apertus (8B on a laptop, 70B on a server). It is the only model in this guide whose training data pipeline is fully published. For European origin with more raw capability, use Mistral.
• "I need it to read images, scans or screenshots." Use Ministral 3, Qwen3.5, Gemma 4 or Apertus 1.5. All accept images.
• "I have one good GPU server and want a capable all-rounder." Use gpt-oss-120b (fits one 80 GB GPU) or Mistral Small 4.
• "I want the strongest open model, whatever it costs to run." That is Kimi K3.1 Pro. You will need a GPU cluster, and you must read its licence.
• "My organisation cannot use models from Chinese labs." Your shortlist is Mistral, Apertus, Beam, Muse Spark Open, gpt-oss, Gemma 4 and Nemotron.
• "I am building a product or a hosted service on top of it." Prefer models under plain Apache 2.0 or MIT: Beam, DeepSeek V4.1 Flash, MiMo-V2.6, Qwen3.8 27B, Gemma 4, gpt-oss, Mistral Small 4, Ministral 3, the small Qwen models and Apertus.
• "I need the absolute best answer and the data is not sensitive." Use a closed frontier model. That is not what this guide is for, and pretending otherwise would not help you.

If one of those lines settled it, you can stop here.

Part 2: The five questions that actually decide

Every model choice comes down to five questions. Their order matters, because each one removes options before the next is asked.

1. Where will it run? A phone, a laptop, a single GPU, a server or a cluster. This is the hardest constraint, because a model that does not fit in memory does not run at all. Part 6 shows how to work it out.

2. What data will it see? Public text, internal documents or regulated client data. This decides whether hosted APIs are acceptable at all, and whether you care about the model's origin and the jurisdiction of whoever hosts it.

3. What will you build on it? Private use, an internal tool, a commercial product or a hosted service for others. This decides which licences you can accept.

4. What is the task? Answering from your own documents, writing, translation, coding, long autonomous agent work, or reading images. Different models are strong at different things, and the hardest agentic tasks are where capability differences really show.

5. Which languages? Most models handle English and major languages well. Coverage of smaller languages, including Swiss German and Romansh, varies a lot.

Only after these five questions does a capability score help you choose between the models that are left.

Part 3: The master table

Every major open-weight family as of 7 November 2026. "Total / active" gives the parameter count; for mixture-of-experts (MoE) models, only the active share is computed per token (see Part 6). The AA v5 column is the Artificial Analysis Intelligence Index v5.0.1, an independent composite in which 60 percent of the tests are private. For reference, the best closed model, Claude Opus 5.5, scores 58. A dash means no v5 score was published.

Show as a plain table

Model	Developer (origin)	Weights released	Total / active params	Context	Inputs	Licence	AA v5	Runs on	Best for
Kimi K3.1 Pro	Moonshot AI (CN)	Nov 2026	2.8T / 104B (MoE)	1M	Text, image	Kimi K3.1-Pro Licence (MIT-style, custom)	46	GPU cluster	Highest open capability, agentic coding
DeepSeek V4.1 Flash	DeepSeek (CN)	Sep–Oct 2026	552B / 16B (MoE)	1M	Text, image	MIT	33	Multi-GPU server	Capability per dollar, permissive licence
GLM-5.3	Z.ai (CN)	29 Aug 2026	~744–753B / 40B (MoE)¹	1M	Not verified	Custom GLM-5.3 Licence²	32	Multi-GPU server	Agentic coding
MiMo-V2.6-Pro	Xiaomi (CN)	21 Sep 2026	1.02T / 42B (MoE)	1M	Text, image, audio, video	MIT	30	Multi-GPU server	Multimodal, permissive licence
MiniMax M3.1 Flash	MiniMax (CN)	23 Oct 2026	428B / 23B (MoE)	1M	Text, image, video	MiniMax Community Licence (restrictive)	29	Multi-GPU server	Fast, cheap agent workloads
Mistral Large 4	Mistral AI (FR)	29 Oct 2026	~1T / 49B (MoE)	1M	Text, image	Mistral Ecosystem Open-Weight Licence (modified MIT)	26	Multi-GPU server	Strongest European model
Muse Spark Open	Meta (US)	15 Oct 2026	816B / 42B (MoE)	1M	Text, image, video, audio	Muse Model Licence (custom)	25	Multi-GPU server	Broadest multimodality, US origin
Qwen3.8 27B	Alibaba (CN)	Aug 2026	27B (dense)	256K	Text, image	Apache 2.0	20	Workstation / 24–32 GB GPU	Best capability on one workstation
Beam 1.0	Reflection (US)	16 Oct 2026	501B / 23B (MoE)	1M	Text	Apache 2.0	18	Multi-GPU server	Speed, cleanest licence at this size
Nemotron 3.5 Super VL	NVIDIA (US)	2026	120B / 12B (MoE)	1M	Text, image	Custom NVIDIA licence (terms not verified)	14	Single 80 GB GPU (quantised)	High-throughput serving
Qwen3.8 2.4T	Alibaba (CN)	Aug 2026	2.4T / 95B (MoE)	262K (ext. to ~1M)	Text	Qwen3.8-Max Licence (custom)	–	GPU cluster	Frontier reasoning; always in thinking mode
Mistral Large 3	Mistral AI (FR)	2 Dec 2025	675B / 41B (MoE)	Not verified	Text, image	Apache 2.0	–	Multi-GPU server	Large European model under Apache 2.0
Mistral Medium 3.5	Mistral AI (FR)	30 Apr 2026	128B (dense)	256K	Text, image	Modified MIT	–	Multi-GPU (4 GPUs)	Chat, reasoning and code in one model
Mistral Small 4	Mistral AI (FR)	Mar 2026³	119B / 6.5B (MoE)	256K	Text, image	Apache 2.0	–	Single 80 GB GPU (quantised)	Fast European all-rounder
gpt-oss-120b	OpenAI (US)	5 Aug 2025	117B / 5.1B (MoE)	128K	Text	Apache 2.0	–	Single 80 GB GPU	Reasoning on one GPU
gpt-oss-20b	OpenAI (US)	5 Aug 2025	21B / 3.6B (MoE)	128K	Text	Apache 2.0	–	16 GB	Reasoning on a laptop
Nemotron 3 Nano	NVIDIA (US)	Dec 2025	31.6B / 3.6B (hybrid Mamba-MoE)	262K	Text	NVIDIA Open Model Licence (custom)	–	24 GB	Long context on modest hardware
Gemma 4 31B	Google (US)	2 Apr 2026	31B (dense)	256K	Text, image, video	Apache 2.0	4⁴	24–32 GB	Multimodal on one GPU
Gemma 4 26B A4B	Google (US)	2 Apr 2026	26B / ~4B (MoE)	256K	Text, image, video	Apache 2.0	–	16–24 GB	Fast multimodal
Gemma 4 E2B / E4B	Google (US)	2 Apr 2026	"Effective" 2B / 4B	128K	Text, image, video, audio	Apache 2.0	–	Phone / laptop	On-device, speech input
Apertus 1.5 70B	Swiss AI Initiative (CH)	24 Jul 2026	70B (dense)	262K	Text, image (audio experimental)	Apache 2.0 (gated download)	–	Multi-GPU (4-way)	Sovereign, auditable, multilingual
Apertus 1.5 8B	Swiss AI Initiative (CH)	24 Jul 2026	8B (dense)	262K	Text, image (audio experimental)	Apache 2.0 (gated download)	–	Laptop 16 GB	Sovereign, auditable, on a laptop
Apertus 8B Instruct (1.0)	Swiss AI Initiative (CH)	2 Sep 2025	8B (dense)	65K	Text	Apache 2.0	–	Laptop 16 GB	Multilingual text, fully documented
Qwen3.5 9B	Alibaba (CN)	Mar 2026	9B (dense)	262K	Text, image	Apache 2.0	–	Laptop 16 GB	Top sub-10B score in AA's March 2026 test
Qwen3.5 4B	Alibaba (CN)	Mar 2026	4B (dense)	262K	Text, image	Apache 2.0	–	Laptop 8 GB	Best 4B-class model
Ministral 3 (3B / 8B / 14B)	Mistral AI (FR)	2 Dec 2025	3B / 8B / 14B (dense)	256K	Text, image	Apache 2.0	–	8–24 GB	European, vision, base/instruct/reasoning variants
Qwen3 4B Instruct 2507	Alibaba (CN)	6 Aug 2025	4B (dense)	262K	Text	Apache 2.0	–	Laptop 8 GB	Fast, non-thinking answers
Qwen3 4B (original)	Alibaba (CN)	29 Apr 2025	4B (dense)	32K (131K with YaRN)	Text	Apache 2.0	–	Laptop 8 GB	Hybrid thinking / non-thinking

¹ Sources differ: Z.ai's announcement as reported gives 744B; Artificial Analysis lists 753B. Active parameters agree at 40B. ² Reported term: organisations with more than US$10bn in annual revenue must pass a security review by Z.ai. One secondary source describes the licence as MIT; we follow the more detailed report and recommend reading the repository licence. ³ Inferred from the version suffix "2603"; no release date was stated on the model card we checked. ⁴ Reasoning variant. The v5 index is weighted heavily towards private and agentic tasks, where models of this size score near the floor (Gemma 4 31B scores 0% on Terminal-Bench 4.0). Treat it as "not suited to long agent tasks" rather than as a measure of general usefulness.

Not in the table: models announced as open whose weights could not be downloaded at the time of writing, including Thinking Machines' Inkling 2 Small (AA v5 score 31, weights pending) and Alibaba's Qwen4 family. Until the files and the licence are on a public repository, treat them as closed.

Part 4: Six rankings, because one number is not enough

By independent capability (AA v5, 7 November 2026)

1. Kimi K3.1 Pro: 46
2. DeepSeek V4.1 Flash (max effort): 33
3. GLM-5.3 (max effort): 32
4. MiMo-V2.6-Pro: 30
5. MiniMax M3.1 Flash: 29
6. Mistral Large 4 (high effort): 26
7. Muse Spark Open (extra-high effort): 25
8. Qwen3.8 27B (extra-high effort): 20
9. Beam 1.0: 18
10. Nemotron 3.5 Super VL: 14

For calibration, the closed leaders are Claude Opus 5.5 at 58, GPT-6 Astra at 57, GPT-6.1 Sol at 55 and Gemini 4 Argon at 50. When v5 replaced v4.3 on 27 October, the gap between the best open and best closed models widened from 3 points to 12. That is not because open models got worse. The new private tests cannot be trained against, and the widening is consistent with Epoch AI's earlier warning that public benchmarks tend to flatter open models.

By licence freedom

• Tier 1, no strings (Apache 2.0 or MIT): Beam, DeepSeek V4.1 Flash, MiMo-V2.6, Qwen3.8 27B, Mistral Large 3, Mistral Small 4, Ministral 3, Gemma 4, gpt-oss (plus a short usage policy), Qwen3 and Qwen3.5 small models, Apertus (Apache 2.0, but downloading requires accepting an acceptable-use policy).
• Tier 2, permissive with conditions at scale or for hosted services: Mistral Large 4 and Medium 3.5 (modified MIT; model-as-a-service businesses need a separate licence), Kimi K3.1 Pro, Qwen3.8 2.4T, Muse Spark Open, GLM-5.3, NVIDIA's Nemotron licence.
• Tier 3, conditions from the first commercial use: MiniMax M3.1 Flash. It requires a "Built with MiniMax M3.1-Flash" notice on commercial products, a one-time notice to MiniMax at or below US$20m annual revenue, prior written authorisation above it, and it prohibits military use.

Notice that the top of the capability ranking and the top of the licence ranking barely overlap. Of the top five by capability, only DeepSeek and MiMo are in Tier 1.

By ease of running locally (smallest first)

Gemma 4 E2B / E4B → Qwen3 4B / Qwen3.5 4B / Ministral 3 3B → Apertus 8B / Ministral 3 8B / Qwen3.5 9B / gpt-oss-20b → Ministral 3 14B / Gemma 4 26B A4B → Nemotron 3 Nano / Qwen3.8 27B / Gemma 4 31B → gpt-oss-120b / Mistral Small 4 / Nemotron 3.5 Super → everything else.

By transparency and auditability

1. Apertus: weights, training-data reconstruction scripts, training code and recipes are all published. Website opt-outs were applied retroactively, personal data was filtered, and training was designed to limit memorisation.
2. Nemotron: NVIDIA has published training data and methods for parts of the family.
3. Everyone else: weights only, with varying levels of documentation. Open weights give you custody of the model, not insight into how it was made.

By origin, for buyers with provenance requirements

• Switzerland: Apertus.
• EU: Mistral (Large 4, Large 3, Medium 3.5, Small 4, Ministral 3).
• United States: Beam, Muse Spark Open, gpt-oss, Gemma 4, Nemotron.
• China: Kimi, DeepSeek, GLM, MiMo, MiniMax, Qwen.

Origin says nothing about where a model runs. A Chinese-developed model running on your own server in Zürich sends no data to China. Some buyers still exclude these models for supply-chain, policy or reputational reasons, and that decision is legitimate. It just should not be confused with a data-flow risk.

By context window

One million tokens: Kimi K3.1 Pro, DeepSeek V4.1 Flash, GLM-5.3, MiMo-V2.6, MiniMax M3.1 Flash, Mistral Large 4, Muse Spark Open, Beam, Nemotron 3.5 Super. Around 256K: most of the rest, including all the current small models. 128K or less: gpt-oss, Gemma 4 edge models, and Apertus 1.0 (65K).

A large window is not the same as good use of it. On a laptop, the memory cost of a long context (Part 6) usually becomes the limit well before the advertised maximum.

Part 5: The four models in our local RAG system

The local retrieval system we build and ship runs fully offline through Ollama and currently offers four answer models. Here is exactly what each one is, and when to use it.

Qwen3 4B Instruct 2507 (the default)

• What it is: a 4-billion-parameter dense model released by Alibaba on 6 August 2025 under Apache 2.0, with a native context of 262,144 tokens.
• Why it is the default: it is a non-thinking model. It answers directly, without first generating a long internal reasoning trace. In a retrieval system, where the passages are already in front of the model, this gives the best speed for a given quality.
• Model-card numbers (Alibaba's own): a large jump over the original Qwen3 4B in non-thinking mode, for example MMLU-Pro 69.6 vs 58.0 and GPQA 62.0 vs 41.7.
• What to watch: on Artificial Analysis's knowledge-and-hallucination test, this model had an 84 percent hallucination rate on hard factual questions it could not answer from memory. That is exactly why it should answer only from retrieved passages, with citations, and say so when the documents do not cover a question. Artificial Analysis now lists it as deprecated in favour of Qwen3.5 4B.

Qwen3 4B (original)

• What it is: the April 2025 release (29 April 2025, Apache 2.0). It is a hybrid model that can switch between a "thinking" mode, which reasons step by step before answering, and a direct mode. Native context is 32K tokens, extendable to 131K with the YaRN technique.
• When to use it: for questions that need multi-step reasoning across several retrieved passages, such as comparing two contracts or working through figures, where the slower thinking mode is worth the wait.
• What to watch: its shorter native context limits how many passages it can read at once. Ollama tags can change between versions, so ollama show qwen3:4b will confirm which model a given install actually contains.

Ministral 3

• What it is: Mistral's family of small dense models, released on 2 December 2025 in 3B, 8B and 14B sizes, each with base, instruct and reasoning variants. All accept images, all have 256K context, and all are Apache 2.0.
• When to use it: when a European-developed model is preferred, for documents in European languages (Mistral cites more than 40 native languages), and for scanned or image-heavy documents. The 8B is the sweet spot on a 16 GB machine; the 3B fits 8 GB.
• Independent context: on Artificial Analysis's multimodal test, Ministral 3 8B scored 46.0 percent, against 65.4 percent for Qwen3.5 4B. Ministral is a solid all-rounder, not the strongest image reader in its class.

Apertus 8B Instruct

• What it is: the Swiss AI Initiative's model (EPFL, ETH Zurich, CSCS). Version 1.0, released on 2 September 2025, is text-only with a 65K context. Version 1.5, released on 24 July 2026, extends it to 262K tokens, adds image input (with experimental audio) and an optional thinking mode. Both are Apache 2.0.
• When to use it: when the user or client needs to know exactly what the model was trained on, when Swiss provenance matters, and for multilingual material, including Swiss German.
• What to watch: it is not the strongest small model on reasoning benchmarks. The 1.5 model card reports MMLU-Pro 48.75 and GPQA 27.23. Those are Apertus's own figures, measured on a different setup from Qwen's, so they are not directly comparable, but the direction is clear. In retrieval use, where the model reads rather than recalls, that gap matters much less. Check which version a given install runs, because 1.0 and 1.5 differ substantially in context and modality.

How to choose between them

The upgrade worth testing next: Qwen3.5 4B. On the same Artificial Analysis index version, it scored 27 against 18 for Qwen3 4B 2507. It adds image input and a 262K context, needs about 3 GB at 4-bit, and keeps the Apache 2.0 licence. Its hallucination rate on hard unanswerable questions was still 80 percent, so the rule does not change: small models are readers, not oracles.

Part 6: The technical layer

Memory: the one calculation you need

A model's weights take roughly parameters × bits per weight ÷ 8 bytes. At 4-bit quantisation, that is half a byte per parameter:

These are estimates for the weights alone. Published figures line up with them: Beam's BF16 checkpoint is about 933 GiB, MiniMax M3.1 Flash downloads at about 252 GB, and Qwen3 4B at 4-bit with an 8K context needs about 5.6 GB in total, of which 1.86 GB is weights. On top of the weights you need memory for the KV cache, which stores the context being processed and grows with every token, plus runtime overhead. A practical rule: budget the weight size plus 30 to 50 percent at moderate context lengths, and more if you push towards the advertised maximum.

Total versus active parameters

Mixture-of-experts models contain many "expert" sub-networks and route each token through only a few of them. This creates the most misunderstood number in AI:

• Total parameters decide what hardware you need, because every expert must be in memory and ready.
• Active parameters decide how fast it runs, because only those are computed for each token.

So gpt-oss-120b (117B total, 5.1B active) needs an 80 GB GPU but generates text at the speed of a much smaller model. Kimi K3.1 Pro has more active parameters (104B) than most models have in total, and 2.8T in total, which is why it needs a cluster.

Quantisation

Quantisation stores weights with fewer bits. At 8-bit, quality losses are usually negligible. At 4-bit, they are generally small for chat and retrieval but can show up in maths, code and long reasoning. Some developers now train with quantisation in mind: OpenAI shipped gpt-oss with its MoE weights in 4-bit MXFP4, and Moonshot published Kimi K3.1 Pro in MXFP4 after quantisation-aware training. For those models, 4-bit is the intended format, not a compromise.

Thinking modes and the hidden cost of reasoning

Many models now offer a reasoning ("thinking") mode or an effort setting, and benchmark results are usually reported at maximum effort. That is why the AA v5 column specifies "max" or "xhigh". Reasoning improves hard tasks but multiplies the tokens generated. Artificial Analysis noted that the small Qwen3.5 models used 230 to 390 million output tokens to complete its index, high even compared with frontier models. On a laptop, tokens are time. For retrieval, direct mode is usually the right default, with thinking reserved for genuinely multi-step questions. Some models remove the choice: the open Qwen3.8 2.4T always thinks.

Reading benchmarks without being fooled

• Independent beats self-reported. Artificial Analysis and Epoch AI run their own evaluations. Model cards report the developer's own numbers, often on favourable settings.
• Private beats public. Public test sets leak into training data. The 3-to-12-point widening when Artificial Analysis moved to mostly private tests is the clearest recent evidence of how much that matters.
• Index versions are not comparable. A score on v4 cannot be set against a score on v5. This guide uses v5 for every cross-model comparison, except for the small Qwen models, where only v4-era scores are available and both models being compared were scored on the same version.
• Settings matter. The same model at "low" and "max" effort can differ by many points.
• Your task beats any leaderboard. Twenty real questions with known answers, run against two or three candidates, will tell you more than any index.

The anatomy of an "open" licence

Read every licence for five clauses:

1. Hosted-service clause: whether you may offer the model as an API or service to others, and above what revenue. Examples: Kimi's US$20m threshold over 12 months for model-as-a-service operators, Muse Spark Open's US$20m threshold for API providers, Mistral Large 4's separate licence for model-as-a-service businesses, and Qwen3.8 2.4T's US$50m threshold for model-as-a-service and AI work-assistant businesses.
2. Attribution clause: whether your product must display the model's name. Muse Spark Open requires "Built with Muse Spark Open" on products. Kimi and Qwen3.8 2.4T require the model name in the interface above 100 million monthly users or US$20m in monthly revenue. MiniMax requires it for any commercial use.
3. Size or revenue gates: for example, GLM-5.3's reported security review for organisations above US$10bn in annual revenue.
4. Use restrictions: for example, MiniMax's prohibition on military use, or acceptable-use policies such as those attached to gpt-oss and Apertus.
5. Exemptions: most custom licences exempt internal use. the Kimi K3 licence, as reported, exempts internal deployments from both its revenue and attribution clauses (K3.1 Pro's licence carries the same service and display clauses), and Muse Spark Open's API threshold does not apply to internal use, research, or powering features of your own product.

For an organisation using a model internally, most Tier 2 licences impose little. For a company building a hosted product, they can decide the choice outright.

Part 7: How this guide was built

Capability scores come from Artificial Analysis's open-source comparison (Intelligence Index v5.0.1), accessed on 7 November 2026, and from its v5 announcement of 27 October 2026. Specifications and licence terms come from model cards and from technical reporting on each release, listed in the sources. Where sources disagreed (GLM-5.3's parameter count and licence, Qwen3.8's exact release day), we say so. Where a figure was not stated by any source we could check, we have left it blank rather than estimated it, except for memory sizes, which are labelled as arithmetic estimates.

Two caveats apply to everything above. First, this field changes monthly: pending weight releases, the next index revision or a new Qwen generation could reorder the table within weeks. Second, no ranking substitutes for testing on your own documents and your own questions. Use the table to build a shortlist, not to make the decision.