Source briefing. Provider claims are attributed to the sources below. This page does not claim a new independent DARC benchmark.
For: Developers and teams putting AI into practical use
⚡ TL;DR — the 30-second version
- DeepSeek's corrected changelog keeps V4 Pro available.
- Qwen checkpoint context and hosted API context are not identical.
- Check custom and modified licenses before commercial deployment.
- Active is not the same as newly released; open is not zero-cost.
DeepSeek V4.1 Flash was released on 10 September with native visual understanding. The current API ID is deepseek-flash. Its original announcement proposed phasing out V4 Pro, but the later changelog explicitly keeps V4 Pro running after 14 September. This correction is why source freshness matters as much as the launch headline.
The Qwen3.8-Flash-Next official repository identifies an experimental architecture preview with hybrid attention, sparse attention and n-gram embeddings. It lists 262,144 native context, extensible to 1,000,000. The repository distinguishes that checkpoint from hosted Qwen3.8-Flash, which has 1M context by default. The license is Qwen Community 1.0, not Apache-2.0.
“An open checkpoint and a hosted API are different products. Track their versions, context settings and licenses separately.”
Mistral's active catalog includes Medium 3.5, Small 4 and other specialist tiers. Medium 3.5 is a dense 128B model with 256K context under a Modified MIT license. It is a current option, but this is not a claim that it launched in September. The retirement catalog is the right place to check older Devstral and Medium deployments.
Open weights do not make inference free. Memory, bandwidth, concurrency, hardware rental and engineering remain real costs. Do not assume a checkpoint fits a single consumer GPU merely because few parameters are active per token; all required weights and runtime state still need a deployment plan.
Field guide / 01
Checkpoint, runtime and hosted service are separate
Record the checkpoint revision and its license, the inference runtime version, quantization, context configuration and hardware. A hosted service with a similar model name may add tools, context extensions or provider-specific changes that are not part of the raw checkpoint.
Test a small local deployment before choosing a large one. Active parameters describe per-token computation, while full weights and working state determine what must fit in your deployment.
Field guide / 02
Read the actual license conditions
Open weights are useful for control and customization, but the term does not promise unlimited commercial rights. Review redistribution, attribution, field restrictions, revenue thresholds and derivative-model terms where applicable.
Save the license version with the deployment decision. A code runtime's license and a model checkpoint's license are independent, and a hosted provider can have another set of service terms.
Before you ship
0/5Use this as a review aid. Checks reset when you leave this page.
Questions that come up
Does open-weight mean free to run?
No. It describes access to weights under a license. Compute, storage, power, serving operations and sometimes license conditions still affect cost.
Sources & verification
Last checked 23 Sept 2026. Publication dates and access conditions belong to the linked originals.
- 01DeepSeek changelog — V4.1 Flash and V4 Pro correctionDeepSeek↗
- 02DeepSeek models and pricingDeepSeek↗
- 03Qwen3.8-Flash-Next official model cardQwen↗
- 04Qwen3.8-Flash-Next: a new architectureQwen↗
- 05Mistral Medium 3.5 model cardMistral AI↗
- 06Mistral model catalog and retirement scheduleMistral AI↗
- 07Ollama — local and cloud models, API and integrationsOllama↗
- 08LM Studio — local APIs and SDKsLM Studio↗