# Frontier on a desktop: the weights fit, but the price cap comes from the cloud

URL: https://www.thedeepfeed.ai/posts/2026-10-17-frontier-on-a-desktop/
Category: Models
Published: 2026-10-17
Author: the-deep-feed
Tags: open-weights, local-inference, qwen, deepseek, glm, mac-studio, quantization, pricing
Kind: deep

> Seven weeks of open-weight releases put models within a benchmark generation of the frontier on a $5,499 Mac Studio or a 12 GB gaming GPU, but the measured numbers show batched cloud hosts, not desktops, now set the price of mid-market tokens.

## TL;DR

- **Qwen3.8-Flash-Next** (125B total, 6B active) writes **53 to 94 tokens/s** on a 12 GB RTX 5070 with 64 GB of system RAM at 2- to 3-bit quantization, per the Strata project's own tests. The larger releases, **DeepSeek V4.1 Flash** (a 510 GB download) and **Muse Spark Open** (about 493 GB in NVFP4), need a 256 GB Mac Studio at aggressive quantization or two machines.
- Near-frontier depends on the benchmark. In DeepSeek's own table, V4.1 Flash edges **Opus 5.0** on Terminal-Bench 2.1 (**90.6 vs 89.1**) but trails it by **20.6 points** on Terminal-Bench 4.0, and that comparator is already a generation old.
- A **$5,499** Mac Studio writing nonstop at about 54 tokens/s pays for itself in roughly **118 days** against GPT-6.1 Sol's $10 output price, but needs about **2.7 years** against DeepSeek's own $1.20 peak rate. The desk caps closed mid-tier prices; batched open-weight hosts cap everything below them.
- The licenses fence the desk off from resale. **Qwen3.8-Flash-Next** requires a separate license for any model-as-a-service or coding-assistant business at any size, and **Muse Spark Open** draws a **$20M** revenue line. DeepSeek V4.1 Flash, GLM-5.3-Flash and MiMo-V2.6 ship under **MIT**.

On October 15, Meta released the weights for [Muse Spark Open](https://research.meta.ai/blog/muse-spark-open), an 816-billion-parameter model it calls "the most capable US open-weight model to date," and pitched it as the first model in its class "to run on your desktop using two Apple M5 Ultra Mac Studios." Two days later, this morning, DeepSeek posted [V4.1-Flash-1017](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017) to Hugging Face under MIT, a 510 GB download whose September predecessor already scored at or above Claude Opus 5.0 on several agent benchmarks in DeepSeek's own table. Over the same weeks, a hobbyist inference engine called [Strata](https://github.com/Niko1221/Strata) collected roughly 26,700 GitHub stars for one promise: "Run a 125-billion-parameter AI model on your own gaming PC."

The working hypothesis behind this piece was simple. If near-frontier capability now runs on a single consumer GPU or a Mac Studio, then the desk becomes a ceiling on what API providers can charge for the middle of the market: price a token above what a $5,499 machine produces it for, and customers walk.

The evidence supports half of that. The weights did reach the desk, though "the desk" means very different machines depending on which model is meant, and "near-frontier" means very different things depending on which benchmark. But the ceiling on mid-market API prices is not coming from desks. It is coming from the same open weights served on batched cloud GPUs, at prices a single-user machine cannot match unless it runs nonstop for years. The desktop caps one thing precisely: the price of closed mid-tier models for always-on, single-tenant work. And the licenses on the most desk-friendly models are written so that the desk cannot become a competing cloud.

# Seven weeks, eight checkpoints

The run of releases since [the toll booth](/posts/2026-08-12-the-toll-booth-on-the-open-road/) and [DeepSeek's price hike](/posts/2026-08-16-deepseek-raised-prices-and-nobody-flinched/) is dense enough to need a table. The last column matters most: for each model, the smallest machine on which someone has published a measured generation speed, and who did the measuring.

| Date | Model | Total / active params | License | Smallest measured setup | Measured by |
|---|---|---|---|---|---|
| Aug 14 | Qwen3.8-27B | 27B dense | Apache 2.0 | 24 GB RTX PRO 4000, 4-bit: 50.4 tok/s | Michał Piszczek |
| Aug 26 | Qwen3.8-Flash-Next | 125B / 6B (+51B n-gram) | Qwen Community 1.0 | 12 GB RTX 5070 + 64 GB RAM, Q2_0: 94 tok/s | Strata maintainers |
| Aug 26 | GLM-5.3-Flash | 320B / 18B | MIT | 256 GB M5 Ultra, mixed 4/8-bit: 31 tok/s | Federico Viticci, MacStories |
| Aug 28 | GLM-5.3 (open weights) | ~753B, FP8 | GLM-5.3 License | None found | — |
| Sep 10 | DeepSeek V4.1 Flash | 552B backbone + 196B Engram / 8B prefill, 16B decode | MIT | 256 GB M5 Ultra: ~45 tok/s | HN commenter, quant unstated |
| Sep 21 | MiMo-V2.6 Flash / Pro | 309B / 15B; 1.02T / 42B | MIT | None found | — |
| Oct 15 | Muse Spark Open | 816B / 42B | Muse Model License | Two 512 GB M5 Ultras, NVFP4: 15.7 tok/s | Apple ML Research |
| Oct 17 | DeepSeek V4.1-Flash-1017 | Same architecture, 510 GB on disk | MIT | Not yet measured | — |

Three things stand out. Only two of the eight have been measured on a card of 24 GB or less, and both are the smallest in the set. The parameter counts behind the word "Flash" now range from 125 billion to roughly 763 billion, so the label no longer describes a size class. And only one row's speed comes from the hardware vendor's own lab: Apple's Muse Spark Open figures, published the day after the release. Every other number comes from a reviewer, a hobbyist or a forum commenter, which is the state of local-inference measurement in October 2026.

# Total parameters decide where a model lives

The most useful frame for these releases is a split model cards rarely state plainly. Total parameters set how much memory a model needs, and therefore where it can physically live. Active parameters set how much compute each token costs, and therefore what a cloud host must charge to serve it. The two pull in opposite directions, and the releases sort cleanly along that tension.

Qwen3.8-27B is dense: all 27 billion parameters work on every token. Its [4-bit file is 17 GB](https://quesma.com/blog/qwen38-27b-quantizations-benchmarked/), small enough for a 24 GB card with room for context, which makes it the easiest near-frontier model to run at home. It is also comparatively expensive to serve: [Artificial Analysis](https://artificialanalysis.ai/models/qwen3-8-27b) lists it at $0.50 per million input tokens and $3.00 per million output. GLM-5.3-Flash is the opposite shape, 320 billion total parameters with 18 billion active, and the same tracker prices it at [$0.15 input and $0.50 output](https://artificialanalysis.ai/models/glm-5-3-flash). The model that is twelve times larger on disk is six times cheaper per output token in the cloud, because a host with hundreds of gigabytes of pooled accelerator memory only pays compute for the 18 billion parameters each token touches.

That arithmetic explains why the biggest releases of the window, DeepSeek V4.1 Flash, MiMo-V2.6 Pro and Muse Spark Open, are sparse giants. Sparsity makes them cheap to sell by the token and hard to host on one machine. DeepSeek's [model card](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash) describes a 552-billion-parameter backbone plus a 196-billion-parameter "Engram" memory accessed by lookup, activating 8 billion parameters per token during prefill and 16 billion during decode. Its global KV cache is down to 890 bytes per token, which the card puts at roughly a quarter of V4-Flash's and 1/437th of DeepSeek V1's. Every one of those design choices lowers the cost of serving a busy cluster. None of them lowers the 510 GB the weights occupy.

![Schematic: models plotted by total params (memory) vs active params (serving cost), with 24 GB, 128 GB, 256 GB, 512 GB tiers](/post-images/2026-10-17-frontier-on-a-desktop/total-vs-active.jpg)

Qwen3.8-Flash-Next is the one release designed from the other end. Its [model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next) lists 125 billion parameters with 6 billion activated, plus a 51-billion-parameter n-gram embedding table that Qwen describes as "more amenable to offloading than Mixture-of-Experts" and "highly efficient for memory-constrained accelerators." The card calls the model an "experimental preview of the architecture that will underpin Qwen4." A lab that wanted its next generation to live on consumer hardware would build roughly this: very few active parameters, and a large block of weights that can sit in cheap memory rather than on the GPU.

The community built the rest. Strata's [README](https://github.com/Niko1221/Strata) describes splitting the model's 24,576 experts across the whole machine: the most-used few thousand on the graphics card, all of them in system RAM, a lookup table on the SSD. Only 10 experts fire per token, and a small draft model guesses ahead so the large one can verify several tokens at once, which the maintainers say delivers answers "1.6-1.8x sooner." A separate project, [Slotstream](https://github.com/carloslfu/slotstream), streams the same 105 GB model from SSD on Macs with 16 to 64 GB of memory, caching the busiest experts in RAM. Neither trick works on a dense model, and both depend on Qwen having drawn the architecture this way.

# What "near-frontier" measures, and against whom

Nearly every capability comparison in this window is self-reported, and most compare against models a generation behind the current frontier. That is the first thing to correct for.

DeepSeek's [V4.1 Flash card](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash) benchmarks against Claude Opus 5.0 and GPT-5.6 Sol. On the older, saturating agent benchmarks it ties or leads: 90.6 on Terminal-Bench 2.1 against Opus 5.0's 89.1, and 74.2 on DeepSWE v1.1 against 74.0. On the newest ones, it does not. Terminal-Bench 4.0 shows Opus 5.0 at 51.8 and V4.1 Flash at 31.2, a 20.6-point gap; NL2Repo shows 75.3 against 64.0. Xiaomi's [MiMo-V2.6 card](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL) shows the same pattern: 89.9 for its Pro model on Terminal-Bench 2.1, level with Claude Opus 5 at 89.1, but 34.9 on Terminal-Bench 4.0 against Opus 5's 49.0. The frontier has since moved again. OpenAI's current lineup is GPT-6 Astra and [GPT-6.1 Sol](/posts/2026-10-09-gpt-6-1-sol-prices-speed-above-astra/), and Anthropic's is Opus 5.5, none of which appear in these tables.

The desk-friendly model sits further back. Qwen3.8-Flash-Next's card compares it with Claude Opus 4.6, scoring 62.5 on SWE-bench Pro to Opus 4.6's 53.4 and 91.7 on GPQA Diamond to 91.3. Those are strong numbers against a model several Anthropic releases behind the current Opus 5.5. "Near-frontier" is defensible for the giants on saturated benchmarks. For the models that fit a gaming PC, "about where the frontier was a generation or two ago" is the precise claim.

Practitioners report the gap is smaller in daily use than the hardest benchmarks suggest. The developer behind the Jono's Corner blog, whose [post](https://www.dgt.is/blog/2026-10-07-deepseek-freek-out/) drew 1,137 points and 989 comments on [Hacker News](https://news.ycombinator.com/item?id=50000488), wrote after a month of heavy use: *"When I'm mid-session, if I don't look at the model name, I honestly could not tell you if I'm using DeepSeek or Opus."* The same writer still pulls in Opus 5.5 "to do a final code review, which will catch a few edge cases." That is the shape of the evidence everywhere: a cheap open model for the bulk of the tokens, a frontier model for the last mile.

# The quantization tax, measured

Running a model on a desk almost always means running a compressed version of it, and the useful question is how much the compression costs. The best public answer in this window comes from [Quesma](https://quesma.com/blog/qwen38-27b-quantizations-benchmarked/), which spent about $3,000 of GPU time re-running Qwen's official benchmarks on Qwen3.8-27B at several quantization levels. Its findings:

| Qwen3.8-27B build | File size | Result (Quesma's runs) |
|---|---|---|
| BF16 (full) | 55 GB | Replicated Qwen's published scores |
| Q8_0 | 29 GB | Not run on Terminal-Bench; interpolated |
| Q4_K_M | 17 GB | Matched BF16 on Terminal-Bench 2.1 |
| UD-Q2_K_XL | 10.7 GB | Noticeable drop; "the level of Opus 4.7 or Gemini 3.1 Pro" |
| UD-IQ1_S | 6.2 GB | Near random chance on GPQA Diamond |

Four bits is close to free on that test. Two bits costs a visible step. One bit collapses. [TerminalBytes](https://terminalbytes.com/run-qwen-3-8-27b-locally/) found the same cliff from the user's side on an M3 Ultra: the 1-bit file "gets facts right, but it cannot commit to an answer."

![Schematic: Qwen3.8-27B quantization ladder from BF16 55 GB to 1-bit 6.2 GB, with the quality cliff marked below 2-bit](/post-images/2026-10-17-frontier-on-a-desktop/quantization-ladder.jpg)

That matters for reading the Strata numbers precisely. The headline Hacker News [submission](https://news.ycombinator.com/item?id=49953495), with 939 points, promised Flash-Next "on consumer hardware (RTX 4090) at 100T/s." The README's own measurements are on an RTX 5070 with 12 GB of VRAM, a Ryzen 5 7600 and 64 GB of RAM: 94 tokens per second at Q2_0, 79 at IQ2_XS, 62 at IQ3_XXS and 53 at IQ3_S, with prompt reading at 1,620 to 2,650 tokens per second on a 32K-token input. On an AMD RX 9070 XT the same builds run at 44 to 60. The RTX 4090 figure in the title traces to [a commenter](https://news.ycombinator.com/item?id=49953496) reporting 124 tokens per second on a 4090 with 128 GB of DDR5, not to the maintainers.

Every Strata speed is at roughly 2 to 3 bits per weight, the zone where Quesma's 27B tests start to show losses. Strata's "Coder" variant goes further, removing half the experts; the README says it "reaches 91% of the full model's SWE-bench Verified score (measured by its authors)." Run Unsloth's 4-bit build instead and the RAM requirement rises to about 88 GB to keep every expert resident; with 64 GB, the README reports 7 to 8.5 tokens per second as the engine reads from SSD.

On a 256 GB Mac Studio the trade-off loosens. Federico Viticci's [MacStories review](https://www.macstories.net/stories/m5-ultra-mac-studio-review-the-dream-mac-for-local-ai-agents/) of the M5 Ultra ran Flash-Next at a 4-bit mixed quantization and measured 54 tokens per second on a 4,203-token prompt and 73 on a 65,235-token prompt, with prompt reading near 2,700 to 2,900 tokens per second. He reports he can run 5-bit "entirely in RAM." GLM-5.3-Flash in a mixed 4/8-bit build wrote 31 tokens per second on the same machine.

Speed per token also understates the wall-clock cost of reasoning models. Simon Willison's [first test](https://simonwillison.net/2026/Aug/16/qwen-38-27b/) of Qwen3.8-27B on local hardware took 21 minutes and 22,276 reasoning tokens to draw one SVG, because the model defaults to its highest effort setting; his advice was to "ignore that default." Tokens per task, not tokens per second, is the number a desk buyer should budget against.

# The memory bill arrived first

The hypothesis treated hardware as a fixed cost. In 2026 it is not, and the component local inference depends on most is the one in shortest supply.

[Tom's Hardware](https://www.tomshardware.com/pc-components/ram/memory-prices-climb-500-percent-in-12-months-up-to-10x-the-lowest-ever-tracked-prices-128gb-of-ddr5-now-usd3-399) reported in August that a 64 GB DDR5-5600 kit, under $200 a year earlier, now costs over $1,100, and that 128 GB kits run $3,399. Strata's README says "Your RAM decides which model fits" and recommends 64 GB to run every size. The software that makes a 12 GB card useful moved the bottleneck onto the part whose price rose roughly fivefold.

Apple's line shows the same squeeze. The new M6 Mac mini starts at $899, which [Ars Technica](https://arstechnica.com/gadgets/2026/09/apple-m6-mac-mini-review-300-price-hike-spoils-a-nice-upgrade/) notes is "a $300 increase for the same memory and storage specs" as the $599 M4 model. The [Mac Studio](https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/) with M5 Ultra starts at $5,499, with up to 512 GB of unified memory at 1.2 TB/s, but Apple's own announcement says the 512 GB configuration "is coming in late October." As of today, the machine Meta's launch post names as the way to run Muse Spark Open at home has not shipped in the memory configuration that model requires.

Demand ran ahead of supply anyway. [MacRumors](https://www.macrumors.com/2026/08/30/apple-unexpected-mac-mini-and-studio-demand/), summarizing reporting by The Information, wrote that the early Mac launch was "driven by unexpectedly strong enterprise appetite for AI hardware," that many Mac mini and Mac Studio configurations had been "out of stock for months," and that some enterprise buyers were turning to Nvidia's DGX Spark instead. Per that report, Apple had no engineering team dedicated to business customers and lacked an enterprise AI strategy.

![Schematic: the local-inference supply chain from DRAM fab to desk, with price callouts for 64 GB DDR5, Mac mini and the 512 GB Studio](/post-images/2026-10-17-frontier-on-a-desktop/memory-squeeze.jpg)

Hacker News users described the economics in fewer words. Under Apple's Muse Spark post, one [commenter](https://news.ycombinator.com/item?id=50119897) wrote: *"So, Meta will spend millions/billions on building open-weight models and then people will send Apple $20,000 to run them."* In the DeepSeek thread, [another](https://news.ycombinator.com/item?id=50012297) named the structural reason the squeeze may persist: *"locally-hosted models can't parallelize runs from many users the way cloud-hosted ones can."* Every desk holds its own full copy of the weights. A cloud host holds one copy for thousands of users.

# The break-even, worked out

This is the test the hypothesis stands or falls on, so it is worth running the numbers. Take the strongest desk in this piece, an M5 Ultra Mac Studio at its $5,499 starting price, and Viticci's measured 54 tokens per second for Flash-Next. Assume it generates output every second of every day, a deliberately generous case: that is about 4.67 million output tokens a day. Ignore input tokens, which favor the cloud further because cache hits are nearly free there, and ignore electricity, which Apple's [MLX benchmark post](https://machinelearning.apple.com/research/mlx-distributed-inference) puts at about 300 W peak for an M5 Ultra under inference, or up to 7.2 kWh a day.

| API alternative | Output price per 1M tokens | Daily value of 4.67M tokens | Days to recover $5,499 |
|---|---|---|---|
| GPT-6.1 Sol (OpenAI) | $10.00 | $46.70 | ~118 days |
| DeepSeek V4.1 Flash, peak | $1.20 | $5.60 | ~982 days (2.7 years) |
| DeepSeek V4.1 Flash, off-peak | $0.60 | $2.80 | ~1,964 days (5.4 years) |
| GLM-5.3-Flash (Artificial Analysis listing) | $0.50 | $2.33 | ~2,357 days (6.5 years) |

Prices from [OpenAI's Sol page](https://developers.openai.com/api/docs/models/gpt-6.1-sol), [DeepSeek's pricing page](https://api-docs.deepseek.com/quick_start/pricing/) and Artificial Analysis.

Against a closed mid-tier model, the desk pays for itself in about four months of continuous use. Against the open weights served in the cloud, it takes years, longer than the hardware generation will last. That is the core finding: the price ceiling on mid-market tokens is real, but it is set by batched cloud hosts running MIT-licensed weights, and they set it well below what any single-user machine can reach.

![Chart schematic: days for a $5,499 Mac Studio to repay itself against four API output prices, from 118 days to 6.5 years](/post-images/2026-10-17-frontier-on-a-desktop/break-even-meter.jpg)

DeepSeek's own pricing history makes the point. On August 16 this publication covered the company [raising V4-Pro output to $1.98 off-peak and $3.96 at peak](/posts/2026-08-16-deepseek-raised-prices-and-nobody-flinched/). Twenty-five days later it launched V4.1 Flash, which its [announcement](https://www.deepseek.com/en/news/deepseek-v4-1-flash/) says outperforms V4-Pro, and began routing all V4-Pro requests to the new model "at V4.1-Flash rates." The current [rate card](https://api-docs.deepseek.com/quick_start/pricing/) lists V4.1 Flash at $0.15 to $0.30 per million input tokens on a cache miss, $0.003 to $0.006 on a cache hit, and $0.60 to $1.20 per million output, with a concurrency limit of 2,500 requests. The price did not fall because desks threatened it. It fell because a new architecture made serving cheaper, and DeepSeek said so: "V4.1-Flash lets us serve more users at a lower cost."

> 🚀 Introducing DeepSeek-V4.1-Flash: smarter, faster, more efficient.
>
> 🔹 Introducing the smallest model in our new architecture family, with native visual understanding.
> 🔹 Designed for greater capability, faster inference, higher throughput, and scaling to larger models.
>
> — [@deepseek_ai](https://x.com/deepseek_ai/status/2097930608790167907), Sep 10, 2026

The same logic shows up in what hosted open weights do for teams. The Wagtail core team's Thibaud Colas [reported](https://wagtail.org/blog/one-month-on-glm-53-flash/) that a month-long challenge put about a billion tokens through GLM 5.3 Flash on hosted providers for $68. The Jono's Corner post put it more bluntly: *"the economics of 4.1 Flash mean self-hosting is not worth it. If saving money is your goal, you will never recoup the costs."* Calvin French-Owen's [Small Models Have Arrived](https://calv.info/small-models-have-arrived), which drew 800 points on Hacker News in late August, argued that the cost of "fast/cheap/good-enough" tokens is what finally makes consumer AI products viable. His examples were API calls costing tens of cents, not machines on desks.

Even the most-shared GLM-5.3 story of the window ran in the cloud. The [Fire HD tablet write-up](https://ericpardee.github.io/fire-hd-ownership/) that reached 706 points on Hacker News describes $266.15 in model spending, with GLM-5.3 finishing the job "in one day on day one of an $80 subscription." It is a story about cheap open-weight intelligence. It is not a story about local hardware.

# Where the desk does win

The break-even table assumes the alternative is metered by the token. For one class of workload the desk wins anyway: continuous, private, single-tenant background agents whose token volume would be absurd to pay for, and whose data should not leave the building.

Viticci's review describes exactly that. For his iOS 27 review, a team of agents built on DeepSeek V4 Flash ran on his Mac Studio "24/7, for 99 days" across 310 documents, because relying on the OpenAI or Anthropic APIs "for this kind of always-on, persistent background task would be…cost-prohibitive, to say the least." He also notes the other side: fast cloud hosts "feel extremely good to use" but are "expensive (I burned through $20 of Inco credits in literally 10 minutes last week)." TerminalBytes runs Qwen3.8-27B for RSS digests and scanned-mail filing on the same logic, where "None of this cares about tokens per second."

That workload is real, and against closed mid-tier APIs it is where the 118-day payback bites. A buyer pricing an always-on agent fleet on GPT-6.1 Sol, for work a Flash-Next-class model can handle, now has an alternative that pays back inside a fiscal year. Against open-weight hosts the case rests on privacy, control and immunity from rate limits rather than cost, and those are worth paying for. The concurrency ceiling still applies: Viticci reports stacking "up to three concurrent Flash-Next sessions" on a 256 GB machine, against DeepSeek's 2,500-request limit per account.

Vendor marketing has moved to match. Apple's Mac Studio announcement promises users can run massive models "without counting tokens or worrying about rising cloud costs." Its research team's benchmarks show what that means at the top end: Muse Spark Open in NVFP4 needs about 493 GB for weights alone, so it runs across two Mac Studios at 15.7 tokens per second with pipeline parallelism, and at 36.7 across four with tensor parallelism. A single M5 Ultra runs the dense Qwen3.8-27B at full BF16 precision at 19.7 tokens per second. Under that post, one [commenter](https://news.ycombinator.com/item?id=50119761) objected that Apple was "sandbagging Qwen pretty heavily by benchmarking Meta's model at 4-bit, against Qwen only at native BF16," and [another](https://news.ycombinator.com/item?id=50120882) reported that "single M5 Ultra with 256GiB runs DeepSeek 4.1 Flash 0910 just fine (~3200 tok/s prefill, ~45 tok/s)." That second figure carries no stated quantization and comes from one user, which is why the table above flags the quantization as unstated.

> Pretty remarkable what two M5 Ultras + MLX can do now
>
> — [@awnihannun](https://x.com/awnihannun/status/2110793280422387826), Oct 15, 2026

Cloud serving moved too, and faster. [Cerebras](https://inference-docs.cerebras.ai/models/overview) lists Qwen3.8-27B at about 1,850 tokens per second. Qwen itself highlighted Flash-Next on Nvidia hardware at a per-user speed no desk approaches:

> Congrats to @lightseekorg on the impressive TokenSpeed results! 🚀 Great to see Qwen3.8-Flash-Next reaching 685 output tokens/s per user on Nvida GPUs. Excited to see the community unlock more of Qwen’s inference potential!
>
> — [@Alibaba_Qwen](https://x.com/Alibaba_Qwen/status/2110214931559899630), Oct 14, 2026

# The licenses are written for the desk, not the reseller

If desks were going to cap API prices, the obvious route would be someone buying a rack of them and reselling tokens. The licenses on the most desk-friendly models close that route, and they extend the pattern this publication mapped at [the toll booth](/posts/2026-08-12-the-toll-booth-on-the-open-road/) in August. GLM-5.3 joined it on August 28, when [its weights shipped](/posts/2026-08-28-glm-5-3-goes-open-weight/) under a license with a $10 billion trigger.

| Model | License | Who needs permission |
|---|---|---|
| Qwen3.8-Flash-Next | [Qwen Community License 1.0](https://huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/main/LICENSE) | Any model-as-a-service or "AI Work Assistant" business, at any revenue; internal use exempt |
| Muse Spark Open | [Muse Model License](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B) | Model-as-a-service businesses above $20M in 12-month gross revenue; "Built with Muse Spark Open" display required |
| GLM-5.3 | [GLM-5.3 License](https://huggingface.co/zai-org/GLM-5.3/blob/main/LICENSE) | Model-as-a-service operators above $10B in revenue must pass a Z.ai security review |
| DeepSeek V4.1 Flash, GLM-5.3-Flash, MiMo-V2.6 | MIT | Nobody |
| Qwen3.8-27B | Apache 2.0 | Nobody |

The Qwen clause is the tightest in the set and sits on the most desk-friendly model. It defines an "AI Work Assistant" as "an independent AI-powered product primarily designed for AI-assisted coding or office productivity," and requires a separate license from Qwen for any such business regardless of size. Internal use is exempt only so long as it does not make the model's capabilities "available to any third party." The same model card points production users to Qwen Cloud, where "Qwen3.8-Flash is the official version based on Qwen3.8-Flash-Next," with a 1M-token default context. With 6 billion active parameters, Flash-Next is the model any host could serve most cheaply, and its license prevents a coding-tool company from doing so without Alibaba's agreement.

Meta's terms close a loop from [August](/posts/2026-08-10-zuckerbergs-6500-word-bet/). Muse Glimmer shipped under Apache 2.0, and the manifesto's promise of an open Muse Spark was still a promise. The promise is kept, but the license is no longer Apache: it is a custom document with a $20 million revenue line for model-as-a-service businesses, the same structure Moonshot and Alibaba adopted over the summer. Meta's model card also sets a practical threshold, stating that running it requires "a GPU with at least 700GB of VRAM for NVFP4 and 2TB for BF16."

The MIT releases point the other way, and they are the ones driving the price floor. DeepSeek, Zhipu's Flash model and Xiaomi's MiMo let anyone serve their weights. Too large for most desks, they suit competing cloud hosts, which can undercut any first-party API that drifts too high.

![Schematic: four license gates on one road, labeled ANY RESELLER, $20M, $10B and MIT, with desk and cloud lanes](/post-images/2026-10-17-frontier-on-a-desktop/license-gates.jpg)

# What the room said

The conversation on Hacker News and X through the window was dominated by speed numbers and hardware shopping. The most visible local-inference project in the research for this piece was the Strata repository, near 26,700 stars, and its Hacker News thread was largely users trading measurements. One reported 124 tokens per second on a 4090; another, running the Coder variant on a 10 GB RTX 3080 with 48 GB of RAM, reported about 30. [One reply](https://news.ycombinator.com/item?id=49957824) to the speed claims asked the question tokens per second cannot: "Speed is one thing, accuracy another. Have you benchmarked it against a reference?"

The hardware enthusiasm predates most of the models. When Flash-Next's release was announced for the next day, on August 25, one [commenter](https://news.ycombinator.com/item?id=49435021) wrote: *"Finally, a reason to own a 128GB Strix Halo or GB10 device. Or a reason to consider the new Mac Studio."* By October the skeptical case was just as clear in the DeepSeek thread: hardware is expensive, memory is scarce, and cloud hosts pool what desks must duplicate.

What the discourse almost never touched was licensing. Across the threads reviewed for this piece, the Qwen Community License's "AI Work Assistant" clause drew no visible discussion, though it governs whether the model everyone was benchmarking on gaming PCs can power a commercial coding tool. The pattern [the July map](/posts/2026-07-20-china-open-weight-surge/) found holds: the crowd's question is still "can I run it," not "can I sell it."

# What to watch before the next map

| Item | Status as of Oct 17 | What it would change |
|---|---|---|
| 512 GB Mac Studio | Apple says "late October" | Makes Muse Spark Open's two-machine setup buyable; first independent tok/s numbers |
| Community quants of V4.1-Flash-1017 | Weights posted today | Determines whether DeepSeek's strongest open model fits one 256 GB machine at usable quality |
| Independent Terminal-Bench 4.0 runs | All current scores self-reported | Tests whether the 14 to 21 point gap to Opus 5-class models holds outside lab tables |
| DRAM prices | 64 GB DDR5 over $1,100 in August | The single largest input to local break-even outside Apple's line |
| DeepSeek V4.1 Pro | Announced as pending | A new top tier could reset the open-weight API floor again |

# The desk prices privacy, the cloud prices tokens

> **The Deep Feed's position:** near-frontier open weights now run on hardware an individual can buy, and the claim deserves its precision: a generation-old frontier on a 12 GB card at 2 to 3 bits, or a current-generation contender on a 256 GB Mac Studio at aggressive quantization. But the desk does not cap what API providers charge for the middle of the market. Batched cloud hosts serving MIT weights do, at prices a single-user machine needs years to match. The desk caps something narrower: what a closed mid-tier model can charge for always-on, private work, where a $5,499 machine pays back in about four months.
>
> — The Deep Feed

The original hypothesis assumed the ceiling on token prices would come from the cheapest place to run a model. It came from the cheapest place to serve one copy to many users, a datacenter, running weights the labs gave away. The desk's role is different and still valuable: it is the option that never changes its price, its model or its terms of service, and that never sends a document out of the building. Apple is charging accordingly, memory makers more so, and the most desk-friendly license in the set makes sure the desk stays a desk.

## Sources

- [Meta AI Research — Introducing Muse Spark Open (Oct 15, 2026)](https://research.meta.ai/blog/muse-spark-open)
- [Hugging Face — meta-models/Muse-Spark-Open-816B-A42B model card and Muse Model License](https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B)
- [Apple Machine Learning Research — Local and Distributed Large LLM Inference with MLX and RDMA on Mac Studio (Oct 2026)](https://machinelearning.apple.com/research/mlx-distributed-inference)
- [Apple Newsroom — Apple introduces new Mac Studio with M5 Max and M5 Ultra (Aug 25, 2026)](https://www.apple.com/newsroom/2026/08/apple-introduces-new-mac-studio-with-m5-max-and-m5-ultra/)
- [Apple Newsroom — Apple unveils a more powerful Mac mini featuring M6 and M5 Pro (Aug 25, 2026)](https://www.apple.com/newsroom/2026/08/apple-unveils-a-more-powerful-mac-mini-featuring-the-all-new-m6-and-m5-pro/)
- [Hugging Face — Qwen/Qwen3.8-Flash-Next model card](https://huggingface.co/Qwen/Qwen3.8-Flash-Next)
- [Hugging Face — Qwen Community License 1.0 (Qwen3.8-Flash-Next)](https://huggingface.co/Qwen/Qwen3.8-Flash-Next/blob/main/LICENSE)
- [Hugging Face — Qwen/Qwen3.8-27B model card](https://huggingface.co/Qwen/Qwen3.8-27B)
- [Hugging Face — zai-org/GLM-5.3 model card](https://huggingface.co/zai-org/GLM-5.3)
- [Hugging Face — GLM-5.3 License](https://huggingface.co/zai-org/GLM-5.3/blob/main/LICENSE)
- [Hugging Face — zai-org/GLM-5.3-Flash model card](https://huggingface.co/zai-org/GLM-5.3-Flash)
- [Hugging Face — deepseek-ai/DeepSeek-V4.1-Flash model card](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash)
- [Hugging Face — deepseek-ai/DeepSeek-V4.1-Flash-1017 (Oct 17, 2026)](https://huggingface.co/deepseek-ai/DeepSeek-V4.1-Flash-1017)
- [DeepSeek — Introducing DeepSeek-V4.1-Flash (Sep 10, 2026)](https://www.deepseek.com/en/news/deepseek-v4-1-flash/)
- [DeepSeek API Docs — Models & Pricing](https://api-docs.deepseek.com/quick_start/pricing/)
- [Hugging Face — XiaomiMiMo/MiMo-V2.6-Pro-RL model card](https://huggingface.co/XiaomiMiMo/MiMo-V2.6-Pro-RL)
- [GitHub — Niko1221/Strata README](https://github.com/Niko1221/Strata)
- [GitHub — carloslfu/slotstream](https://github.com/carloslfu/slotstream)
- [Quesma — Benchmarking Qwen3.8 27B quantizations: 4-bit holds up, 1-bit collapses (Sep 2026)](https://quesma.com/blog/qwen38-27b-quantizations-benchmarked/)
- [MacStories — M5 Ultra Mac Studio review, Federico Viticci (Sep 21, 2026)](https://www.macstories.net/stories/m5-ultra-mac-studio-review-the-dream-mac-for-local-ai-agents/)
- [MacRumors — Apple caught off guard by AI demand for Mac mini and Mac Studio (Aug 30, 2026)](https://www.macrumors.com/2026/08/30/apple-unexpected-mac-mini-and-studio-demand/)
- [Ars Technica — Apple M6 Mac mini review (Sep 2026)](https://arstechnica.com/gadgets/2026/09/apple-m6-mac-mini-review-300-price-hike-spoils-a-nice-upgrade/)
- [Tom's Hardware — Memory prices climb 500% in 12 months (Aug 2026)](https://www.tomshardware.com/pc-components/ram/memory-prices-climb-500-percent-in-12-months-up-to-10x-the-lowest-ever-tracked-prices-128gb-of-ddr5-now-usd3-399)
- [TerminalBytes — Run Qwen3.8 27B locally: real numbers from my Mac Studio (Aug 2026)](https://terminalbytes.com/run-qwen-3-8-27b-locally/)
- [Michał Piszczek — Qwen3.8 27B at 256K: 50 TPS on a 24 GB GPU (Aug 2026)](https://piszczek.pl/blog/qwen38-27b-256k-50-tps-24gb-gpu)
- [Simon Willison — Qwen 3.8 27B is excellent, but it defaults to overthinking things (Aug 16, 2026)](https://simonwillison.net/2026/Aug/16/qwen-38-27b/)
- [Calvin French-Owen — Small Models Have Arrived (Aug 26, 2026)](https://calv.info/small-models-have-arrived)
- [Jono's Corner — Why isn't the industry freaking out about DeepSeek 4.1 Flash? (Oct 7, 2026)](https://www.dgt.is/blog/2026-10-07-deepseek-freek-out/)
- [Wagtail — One month on GLM 5.3 Flash, Thibaud Colas (Oct 2026)](https://wagtail.org/blog/one-month-on-glm-53-flash/)
- [Amazon kept shutting down my tablet, so I spent $266 on four AI models to own it (Aug 2026)](https://ericpardee.github.io/fire-hd-ownership/)
- [Artificial Analysis — GLM-5.3-Flash intelligence, performance and price](https://artificialanalysis.ai/models/glm-5-3-flash)
- [Artificial Analysis — Qwen3.8 27B intelligence, performance and price](https://artificialanalysis.ai/models/qwen3-8-27b)
- [Cerebras Inference Docs — Models overview](https://inference-docs.cerebras.ai/models/overview)
- [OpenAI API Docs — GPT-6.1 Sol model page](https://developers.openai.com/api/docs/models/gpt-6.1-sol)
- [Hacker News — Run Qwen 3.8 Flash Next (125B) on consumer hardware (939 points)](https://news.ycombinator.com/item?id=49953495)
- [Hacker News — Apple ML Research: Muse Spark Open with MLX on Mac Studio (345 points)](https://news.ycombinator.com/item?id=50118320)
- [Hacker News — Why isn't the industry freaking out about DeepSeek 4.1 Flash? (1,137 points)](https://news.ycombinator.com/item?id=50000488)
- [@deepseek_ai on X — Introducing DeepSeek-V4.1-Flash (Sep 10, 2026)](https://x.com/deepseek_ai/status/2097930608790167907)
- [@Alibaba_Qwen on X — Qwen3.8-Flash-Next at 685 output tokens/s per user (Oct 14, 2026)](https://x.com/Alibaba_Qwen/status/2110214931559899630)
- [@awnihannun on X — two M5 Ultras and MLX (Oct 15, 2026)](https://x.com/awnihannun/status/2110793280422387826)

---

Canonical: https://www.thedeepfeed.ai/posts/2026-10-17-frontier-on-a-desktop/
Site: https://www.thedeepfeed.ai
Full corpus: https://www.thedeepfeed.ai/llms-full.txt