Key takeaways
- Two major open-weight releases landed within 48 hours. Meta published Muse Spark Open on 15 October (816B total parameters, ~42B active, 1M-token context) (Meta for Developers). DeepSeek followed on 17 October with V4.1 Flash 1017, also with a 1M-token context (MarkTechPost).
- On the same independent index, DeepSeek leads. Artificial Analysis scores V4.1 Flash 1017 at 45 (Artificial Analysis), against 39 for Muse Spark Open (OrcaRouter summary of Artificial Analysis).
- DeepSeek is roughly 4× cheaper through an API. At ~₹94 per US$, DeepSeek's list price works out to about ₹28 per million input tokens and ₹113 per million output tokens. Hosted Muse Spark Open costs about ₹118 and ₹400.
- The licences are very different. DeepSeek's weights are listed as MIT. Meta's Muse Model License requires "Built with Muse Spark Open" attribution and gated sign-up, and API resellers above US$20M in revenue need a separate agreement with Meta.
- Our advice: for most Indian products, use DeepSeek 1017 through an API or call Muse through a hosted provider. Self-host only if your data cannot leave your own infrastructure, because either model needs a multi-GPU server costing crores.
1. What shipped, in one paragraph each
Muse Spark Open is Meta's first open-weight release from the Muse line. It shares a base model with the closed Muse Spark 1.3. It is a sparse mixture-of-experts (MoE) model and accepts text, image, video and audio input. Meta pitches it at coding and agent work: multi-file refactors, subagents, and browser and computer use (PyTorch blog). The weights are on Hugging Face. It is also in the SageMaker, Google Cloud Model Garden and Microsoft Foundry catalogues, and hosted by Together, Fireworks, Baseten, OpenRouter and others.
DeepSeek V4.1 Flash 1017 is a snapshot update of September's V4.1 Flash. It runs as deepseek-flash in DeepSeek's app and API. DeepSeek reports 77.0 on DeepSWE v1.1, up from 74.2 for the previous version. It also launched paid Search and Fetch APIs at US$0.75 per 1,000 successful calls (MarkTechPost). The Decoder reports that on DeepSeek's own chart a DeepSWE task costs just under US$0.20. It also notes that this figure is vendor-reported and not yet independently reproduced (The Decoder).
2. Side-by-side
| Muse Spark Open | DeepSeek V4.1 Flash 1017 | |
|---|---|---|
| Released | 15 Oct 2026 | 17 Oct 2026 |
| Size | 816B total, ~42B active (MoE) | 552B backbone + 196B "Engram"; 8B active in prefill, 16B in decode (reported) |
| Context | 1M tokens | 1M tokens, up to 384K output |
| Inputs | Text, image, video, audio | Multimodal (reported) |
| Artificial Analysis Intelligence Index | 39 | 45 |
| Output speed (Artificial Analysis) | ~103 tokens/s | ~203 tokens/s |
| Licence | Muse Model License (custom, gated) | MIT, per Artificial Analysis |
| Typical API price (in / out, per 1M tokens) | US$1.25 / US$4.25 | US$0.30 / US$1.20 list; off-peak is half |
| Smallest practical self-host | ~530 GB NVFP4 on 8× H200 or 4× Blackwell Ultra | Not yet published for 1017; September's V4.1 Flash weights were 763B in FP8 |
Two cautions. First, many benchmark figures in vendor posts come from each company's own agent harness, so they are not directly comparable. Meta lists Terminal-Bench 2.1 at 76.4, while DeepSeek reports Terminal-Bench 4.0 at 42.3. Those are different versions of the benchmark. Use the Artificial Analysis row, which applies the same method to both, for comparisons. Second, Meta says audio and video benchmarks and its safety evaluation report are still to come.
3. What it costs in rupees
Take a mid-sized customer-support or internal-docs assistant: 10 million input tokens and 2 million output tokens a month. At ₹94 per US$ (Pound Sterling Live, 21 Oct):
| Route | Monthly token cost | Approx. ₹ / month |
|---|---|---|
| DeepSeek API, peak list price | US$5.40 | ~₹510 |
| DeepSeek API, off-peak (half price) | US$2.70 | ~₹255 |
| DeepSeek 1017 via cheapest OpenRouter providers | from ~US$1.24 | ~₹115+ |
| Muse Spark Open, hosted (US$1.25 / US$4.25) | US$21.00 | ~₹1,975 |
Third-party prices for DeepSeek 1017 on OpenRouter range from US$0.049 to US$0.33 per million input tokens, depending on provider. The cheapest routes can be slower or have lower rate limits, so test before you commit.
At this volume, the model bill is smaller than one engineer-day. What drives cost is how many tokens each request uses: retrieved context, long chat histories and agent loops that retry. A 1M-token context window does not mean you should send 1M tokens. If a support bot sends 50K tokens of retrieved documents with every request, it will cost 10–20× more than one that sends 5K. Fix that first.
4. Self-hosting: when it makes sense in India
Self-hosting a model of this size is not a weekend project. Muse Spark Open's smallest official format (NVFP4, ~530 GB) needs eight H200s or four Blackwell Ultra GPUs. Unsloth's ~2-bit GGUF still needs around 256–300 GB of combined RAM and VRAM, and quality drops at that compression (Meta for Developers). Buying an 8-GPU H200 server costs crores of rupees. Renting one from an Indian GPU cloud costs lakhs a month at typical on-demand rates.
It makes sense when:
- Data cannot leave your own environment. Examples are health records, bank KYC documents or defence-adjacent work, where your DPDP Act obligations or client contracts rule out a foreign API. With open weights you can run inference in an Indian data centre you control.
- Volume is very high. Billions of tokens a month, steady around the clock, can justify reserved GPUs.
- You need to fine-tune. Muse ships with TorchTitan SFT/LoRA support, and fine-tuning through managed APIs (Fireworks, Together) is also listed.
Otherwise, a hosted endpoint is cheaper and faster to set up. For the India-region-but-managed option, check whether your cloud's Model Garden or Foundry catalogue serves the model in an Indian region before you assume it does.
5. Licence: read this before you ship
DeepSeek (MIT, per Artificial Analysis): you can use it commercially, modify it and redistribute it, with no attribution in your product UI. Check the licence file on the 1017 Hugging Face card before release. Early reports on 17 October did not state the snapshot's terms.
Muse Model License (summary):
- You must show "Built with Muse Spark Open" prominently in the product UI or its documentation.
- If you resell API access to the model and your group's revenue is above US$20M in any 12 months, you must sign a separate agreement with Meta. Building it into your own product is exempt.
- Downloading the weights is gated. The form asks for legal name, date of birth and company identifiers, and you cannot edit it after submitting.
- Fine-tuning, distillation and selling derivatives are allowed.
For an Indian SaaS company building features into its own app, neither licence is a blocker. An AI-API reseller or a GPU cloud should read Meta's clause carefully.
6. What to do this week
| If you are… | Do this |
|---|---|
| A startup on GPT/Claude/Gemini APIs | Add DeepSeek 1017 as a second model behind a router. Run your 50 hardest real prompts through both models and compare cost per resolved ticket, not benchmark scores. |
| Building coding agents | Try 1017 on your own repo tasks. Also test Muse on multimodal work (screenshots, UI) where DeepSeek has less evidence. |
| In a regulated sector | Shortlist Muse or DeepSeek for private hosting, estimate GPU cost against your volume, and do a DPDP data-flow review first. |
| A student or fresher | Run the small GGUF builds of earlier models locally, and learn vLLM/SGLang serving and MoE basics. These skills are now asked for in AI-infra job descriptions. |
Whichever you pick, track three numbers from day one: average tokens per request, cache hit rate, and cost per successful outcome. Those predict your bill much better than a leaderboard does.
Need help choosing, routing or self-hosting an open model? See our AI & automation development service, or read how to choose an LLM for your product.