Open-Weight AI Models, October 2026: 4 Open Source LLMs Tested

The open-weight AI landscape looks different as of October 2026: Meta, DeepSeek, StepFun, and NVIDIA released four major open-weight checkpoints across October 15-17 - open source LLMs in the everyday sense, although not all of them ship under an open-source license. I tested all four via the API on October 19, using the same short tasks in Polish. DeepSeek and Nemotron completed the set, while Muse Spark Open and Step 5 sometimes used up the response budget without returning an answer.

Four releases, four different propositions

Muse Spark Open arrived from Meta on October 15. StepFun released Step 5 Preview weights that day, after previously offering the model only through an API (I covered that Step 5 Preview API launch separately). NVIDIA also released Nemotron 3.5 Super VL on October 15. DeepSeek followed with V4.1 Flash 1017 on October 17, an update to the original V4.1 Flash. The distinction matters: this was not a simultaneous launch of four entirely new model families.

These specifications come from the official repositories checked on October 20, 2026. StepFun describes its model as 600B, while the repository metadata reports 604B. DeepSeek’s 763B figure is also a repository metadata value; it should not be read as an active-parameter count. Keeping those labels visible prevents a misleading comparison between total model size and the part used for each token.

All four advertise context windows of roughly one million tokens. That is a capacity specification, not a guarantee that every deployment can process that much text at a useful speed. The memory needed for a long conversation sits on top of the weights, and available capacity also depends on concurrent requests. My short test did not exercise those context windows.

Open weights vs. open source LLM: what actually changes

Open weights give developers a route to running a model on infrastructure they control. They can also make customization possible, subject to the relevant license. That changes the relationship with a hosted service: access no longer depends entirely on a remote endpoint continuing to offer the same checkpoint. It does not remove the work of deployment, maintenance, evaluation, or licensing.

The architecture numbers need similar care. A mixture-of-experts model can activate only part of its parameters for a token, reducing the computation involved. The remaining experts still belong to the model and need to be available to the inference system. Muse Spark Open’s 42B active parameters therefore do not turn it into a conventional 42B download.

DeepSeek uses MIT, including the usual obligation to preserve the copyright and permission notice. Meta uses its own Muse Model License, with a separate agreement for model-as-a-service businesses above $20 million in revenue over 12 months. StepFun also uses a custom license: products above 100 million monthly active users or $20 million in monthly revenue must show “Powered by StepFun AI”, and hardware makers and larger model-as-a-service operators need a separate commercial agreement. NVIDIA uses OpenMDW 1.1. Calling all four simply “free for any commercial use” would erase differences that can affect a product before it ships.

NVIDIA additionally publishes training resources and recipes. That makes the release useful beyond downloading a checkpoint, but it should not be interpreted as a promise that every item used in training can be independently obtained and reproduced. Open weights, published training information, and a completely reproducible training run are separate claims.

What happened in my test in Polish

On October 19, I sent each model the same 12 prompts via the API, once per task (full method and every failed item: four open models tested in Polish). The test used temperature 0 and a 6,000-token response allowance covering reasoning as well as the final answer. Tasks included arithmetic, Polish grammar, extracting invoice details into JSON, a checksum, and a historical weekday. This was a practical check of short responses, not a broad intelligence benchmark.

Every failed item was an empty answer after the completion count reached 6,000 tokens. Muse stalled on converting 2.5 hours into seconds. Step stalled on the three-word instruction, the question about Poland’s administrative regions, and the spelling correction. The recorded nonempty answers were accepted after the grading correction described below.

I corrected one rule in the automatic grader. It initially required the Polish infinitive “wziąć,” but three models returned the valid imperative “Weź parasol, bo na pewno będzie padać.” That means “Take an umbrella, because it will definitely rain.” Accepting that wording raised the scores for Muse, DeepSeek, and Nemotron. Step’s answer to that item was empty.

The timings describe these API requests, including service conditions. Requests could use different serving backends, and each task was run only once. They do not establish a universal speed ranking or predict performance on local hardware. Nor does a perfect score establish translation quality, long-form writing ability, or reliability in another language.

Running locally versus using an API

The API route makes it possible to evaluate these models without first assembling a large inference system. That was the route used for this test. Local deployment becomes relevant when control over the checkpoint, data handling, customization, or sustained workloads justifies the hardware and operational work. Those are different purchasing decisions from trying a model for an afternoon.

Meta’s card gives about 700 GB of GPU memory for NVFP4 and about 2 TB for BF16. Its Mac example uses two M5 Ultra Mac Studio systems with 512 GB of unified memory each. This is a specific high-memory configuration, not evidence that an ordinary MacBook can run the model comfortably.

StepFun’s deployment examples use four B300 GPUs or eight H200 GPUs for FP8. NVIDIA lists a single DGX Spark or H200, or two H100 GPUs, among the NVFP4 options for Super VL. These are published configurations, not measurements from my own hardware. Different precision formats and context settings cannot be treated as interchangeable.

• Choose the exact checkpoint and precision before estimating memory.
• Use its documented inference runtime and conversation formatting.
• Start with a modest context and realistic concurrency.
• Validate final answers, structured output, and tool behavior on your own tasks.
• Increase context or batch size only after measuring the full deployment.

A successful model load is only the first step. An application also needs room for conversation state and intermediate computation. For these releases, “local” often means a workstation or a cluster under your control. A single consumer GPU with 24 GB should not be the default assumption for the full checkpoints discussed here. Models that do fit a home PC are collected in my guide to running AI locally, and my earlier test of local models on an RTX 4090 shows what that class of hardware delivers.

Choosing a model by the work it must finish

For short Polish tasks resembling this test, DeepSeek and Nemotron are the clearest starting points from the observed results. Both completed every item. That is a useful reason to include them in an evaluation, but not enough to choose a production model without representative documents, longer conversations, and repeated requests.

Muse Spark Open deserves a different evaluation when the goal is tool use or extended engineering work, which Meta emphasizes. The unanswered conversion question exposes a problem under this test’s budget; it does not settle the model’s usefulness for every agent workflow. Step 5 needs the same distinction, with particular attention to whether reasoning leaves room for a final answer.

For a document or visual workflow, Nemotron’s multimodal focus is relevant, but the invoice task here contained text. No scanned invoices, screenshots, or videos were submitted. A model’s advertised inputs and an author’s tested inputs should never be allowed to blur together.

API cost also needs a dated, conditional comparison. DeepSeek’s official pricing distinguishes cache hits from uncached input and peak from off-peak hours. The cost recorded for a short historical test is not a current list price. Output consumed during reasoning can matter as much as the visible answer, especially when an empty result still uses the entire allowance.

The four guides in this series

• Muse Spark Open - Meta’s 816B model: hardware, the Mac setup, license conditions and its 11/12 result.
• DeepSeek V4.1 Flash 1017 - what changed, official API prices in USD, the Search and Fetch APIs and its 12/12 result.
• Four open models tested in Polish - method, all 12 tasks, timings, token use and the grading correction.
• Nemotron 3.5 Super VL and Step 5 - two different deployment routes, two very different licenses.

More of my independent tests are collected in the Lab of this blog.

FAQ

Which model won?

DeepSeek and Nemotron tied at 12/12 in my test in Polish. DeepSeek had the lower median request time in this run. The test does not establish a winner for coding, vision, English writing, or autonomous agents.

Can all four be used commercially?

The licenses provide commercial-use routes, with different conditions. StepFun’s current license requires a separate agreement before hardware makers ship the model as a device feature and for model-as-a-service operators above $20 million in revenue over 12 months; Meta’s license has a similar $20 million threshold for model-as-a-service businesses. Read the applicable license for the planned use.

Will a larger response budget fix the empty answers?

It might help, but this test did not include that rerun. A larger budget can also increase waiting time and cost. Treat budget changes as an experiment rather than a confirmed fix.

Does the Polish result transfer to other languages?

It gives a reason to test those languages directly. Grammar, spelling, local formats, and acceptable alternative answers need language-specific checks. Translating the prompts does not automatically preserve the same difficulty or grading rules.