Hardcore Reviews
Hardcore Reviews

Open Weight, But How Open? Five Model Licenses Compared

Five open-weight models crowded into the same mid-October 2026 window, all saying 'open source' - and the substance spans three tiers. Tier A, genuinely open: Reflection Beam under Apache 2.0 with zero riders and ungated downloads; DeepSeek V4.1 Flash 1017 under MIT with zero riders; NVIDIA Nemotron 3.5 Bolt 74B-A9B under OpenMDW-1.1 and the only one of the five to publish weights, training data, and recipes together. Tier B, open weights with riders: Meta Muse Spark Open carries an attribution duty that extends to distilled products (models trained on its outputs must also display Built with Muse Spark Open), a 20-million-dollar annual-revenue threshold for MaaS resale, and gated HF downloads - legally not open source. Tier C, promise unfulfilled: Mistral Large 4 had no weights as of October 23, a license with no name, and an AA classification of proprietary. On the shared ruler of AA Intelligence Index v4.3.2: GLM-5.3 at 45, Kimi K3 at 44, DeepSeek 39.5 about even with Muse at 39, Mistral 38, and Beam 30 (Beam has no public API; its score comes from self-hosted inference). Self-reported DeepSWE and Terminal-Bench numbers use different harnesses and versions and cannot be compared directly. Hardware accessibility spans two orders of magnitude, from Bolt's single 24GB card to Muse's 700GB+ NVFP4 and Beam's 294GB checkpoint (about 400GB of memory to start). The piece includes an eight-dimension license table, a watering-down spectrum (attribution > revenue threshold > gated download > late weights > unnamed license > closed data), and per-use-case protocol advice; single-model deep dives for Muse and Beam live in their own articles.

Published October 23, 20268 min read
<!-- openweight-license-comparison-review | review | Open Weight, But How Open? Five Model Licenses Compared -->

Mid-October packed five major model launches into one window: Meta Muse Spark Open (weights landed October 15), Reflection AI Beam (October 16), DeepSeek V4.1 Flash 1017 (October 17), NVIDIA Nemotron 3.5 Bolt (October 15), and Mistral Large 4 (API preview October 6). Every launch deck said "open." But set the license terms side by side and the substance splits into three tiers: Beam ships Apache 2.0 with zero addenda and no download gate; DeepSeek 1017 ships MIT with zero addenda; Nemotron Bolt is the only one of the five to release weights, data, and recipe together. Muse Spark Open has public weights, but its attribution clause covers distilled descendants, on top of a $20 million revenue threshold and gated downloads — legally, not open source. Mistral Large 4 is blunter: no weights as of this writing, no named license, and third-party leaderboards still file it as proprietary.

Scope first: Muse Spark Open and Beam already have their own deep dives here; this piece compares license terms and accessibility only. Figures come from official model cards, official blogs, and the Artificial Analysis Intelligence Index v4.3.2 ("AA"), as of 2026-10-23.

1. The five at a glance

ModelRelease and weightsParametersLicenseHardware bar
Muse Spark OpenOct 15, weights same day816B total, ~42B active, 1M contextMuse Model License (Meta custom)BF16 ~1.5TB; NVFP4 needs 700GB+ VRAM, or two 512GB M5 Ultra machines
BeamAnnounced Oct 5, weights Oct 16501.3B total, 23.4B activeApache 2.0, zero addendaNVFP4 checkpoint ~294GB; vendor suggests 4x B200
DeepSeek V4.1 Flash 1017Oct 17, web, API, weights same day763B total (third-party config analysis)MIT, zero addendaRepo ~510GB (unconfirmed); official FP8 and INT4 quants
Nemotron 3.5 Bolt 74B-A9BOct 15, weights same day74B total, 9B active, 256K contextOpenMDW-1.1 (NVIDIA's own, not OSI)Native 2-bit QAT ~24GB, one 24GB card
Mistral Large 4Oct 6 API preview, no weights1.05T total, 52B active (with embedding and output, per A52B repo name; docs cite 49B)Unnamed, docs say "Open"No weights, no bill

Calibration notes: DeepSeek's parameter count and repo size are third-party analyses; Mistral's active parameters have two variants — 52B counts embedding and output, 49B does not — this piece uses 52B and flags 49B.

Repositories: Muse Spark Open (plain URL https://huggingface.co/meta-models/Muse-Spark-Open-816B-A42B , with an NVFP4 branch), Beam (plus FP8, NVFP4, and a DFlash2 drafter), DeepSeek V4.1 Flash 1017 (on ModelScope too), Nemotron 3.5 family; Mistral Large 4 has no checkpoint under the mistralai org as of this writing.

2. License terms, dimension by dimension

TermMuse Spark OpenBeamDeepSeek 1017Nemotron BoltMistral Large 4
License nameMuse Model License (custom)Apache 2.0MITOpenMDW-1.1Unnamed
OSI-certified open sourceNoYesYesNo (permissive, not OSI)Unknown
AttributionProminent "Built with Muse Spark Open" label; distilled models countStandard Apache NOTICE onlyNoneNoneUnpublished
Revenue threshold$20M over any 12 months (MaaS resale only)NoneNoneNoneUnpublished
DownloadHF gated formUngatedUngatedUngatedWeights unreleased
Derivatives and distillationAllowed, but derivatives carry the same obligationsFully freeFully freeFully freeUnpublished
Use restrictionsBans military, nuclear, critical infrastructure, under-18 useNoneNoneNoneUnpublished
Data and recipeTraining data closedData closed; SFT and RL recipe publicData and recipe closedWeights, data, and recipe all publicUnpublished

The gaps are wider than the names suggest.

Attribution is Muse's heaviest addendum. Any product using the model or its derivatives must prominently display "Built with Muse Spark Open" — and "derivative model" is defined expansively: models trained or distilled on its outputs count, so the obligation travels with distillates. Third-party analysts read this as aimed at clouds that would resell the model wholesale. The $20M threshold covers MaaS resale only; internal use, research, and your own commercial products are exempt. The structure resembles Kimi K3's threshold (press-reported).

Beam and DeepSeek form the clean control group: zero addenda, ungated, fully free derivatives and distillation. A clean license also buys ecosystem speed: Beam launched with 40-plus partners and day-zero support across Transformers, vLLM, SGLang, Unsloth, Axolotl, and Dynamo. Muse's hosting coverage is fine (Baseten, Fireworks, Groq, Together), so the burden lands on self-hosters and distillers.

Nemotron Bolt's OpenMDW-1.1 is NVIDIA's own license: permissive, no revenue threshold, no OEM restriction, no MaaS clause — but not submitted for OSI certification. What makes it special is scope: training data and recipe are fully public, the only one of the five. Third-party analysts call this infrastructure strategy rather than generosity: NVIDIA sells hardware, not tokens — the wider the models spread, the more GPUs sell.

Mistral Large 4's license text arrives with the weights. Large 3 was Apache 2.0; Large 4 is not, the company has made clear. Some coverage cites a $20M monthly threshold in Mistral's older licenses, but that describes prior terms and is unconfirmed for Large 4.

3. The hardware bill

Licenses decide whether you may; hardware, whether you can.

Beam: the NVFP4 minimum checkpoint is roughly 294GB (another source says ~280GB, unconfirmed). With KV cache and long-context headroom, ~400GB of memory is a comfortable start. All 188 safetensors shards are ungated; the vendor suggests 4x B200.

Muse Spark Open: BF16 weights run about 1.5TB; NVFP4 needs 700GB-plus of VRAM, effectively a B200 node. The official demo chains two 512GB M5 Ultra Mac Studio machines via MLX into one system — Meta calls it the first of its class to run on a desktop. Checkpoint formats are thorough: official BF16 and NVFP4, AMD and Intel quant flavors, plus Unsloth's GGUF.

DeepSeek V4.1 Flash 1017: the repo is ~510GB (still unconfirmed), with official FP8 and INT4 quants; MLA compression reportedly keeps 128K-context KV cache under 2.8GB. A circulating claim that INT4 runs on a 24GB consumer GPU conflicts with the 763B total parameters and reads as exaggerated.

Nemotron Bolt: the only true small-VRAM option. Native 2-bit QAT (effectively 2.0625 bit) packs to about 24GB, verified on an RTX 5090, and fits DGX Spark and RTX Spark laptops ($2,599 and up in the US from October 16). vLLM supports A100, H100, and Blackwell; Unsloth GGUF ~22GB, Ollama ~25GB. The cost: no official Apple Silicon support (Ollama confirms Metal will not work), and no independent AA score yet.

Mistral Large 4: no weights, no bill. A 1.05T BF16 deployment would start around 2TB by class analogy.

4. Capability, measured with one ruler

Self-reported numbers are all over the map. On AA's v4.3.2 third-party ruler: GLM-5.3 scores 45, Kimi K3 44, DeepSeek 39.5 and Muse Spark Open 39 effectively tied, Mistral Large 4 38, Beam 30 — clearly low for its class. Two footnotes: Beam has no public API, so its AA score comes from self-hosted inference; Mistral Large 4 was still classified as proprietary by AA when it debuted.

Self-reports are not useless — just not comparable. Beam's SWE-bench Verified appears as 80.9% in the official blog and 81.6% on the model card; both coexist, likely different harnesses or precisions. On DeepSWE, Beam self-reports 44.4, DeepSeek 77.0, Muse 60 — each under its own harness; Terminal-Bench spans 2.1 and 4.0. Muse's card shows MCP-Atlas 89.2 and standout safety numbers, but vendor-run. DeepSeek's self-reported DeepSWE 77.0 contrasts with AA's own tests (HLE 39.2%, Terminal-Bench 26.8%). Mistral self-reports Cybench 93% and DeepSWE 62% — with GLM-5.3 at just 61 in its table while public boards put GLM-5.3 near 69. AA also flags Mistral's verbosity: 200M output tokens to finish the index (Muse used 110M), costing $1,607 at list prices versus Muse's $1,183. Bolt leads its class on self-reported PinchBench 87.82, but AA has not scored it separately. Whenever you see a "number one," ask which harness.

5. Tiering, and a cheat sheet for dilution

A third-party framework maps well onto this batch: restricted open weights, genuine open source, and hardware-vendor open source. Adding a slot for unfulfilled promises:

Tier A, fully open. Beam: Apache 2.0, zero addenda, ungated, with the SFT and RL recipe and Mirror CLI open, plus a transparent training account (23.8T pretraining tokens; RL on 10,500 GB300 GPUs for four weeks); the weakness is a low AA score for its class. DeepSeek 1017: MIT, zero addenda; the weakness is closed data and recipe. Nemotron Bolt: OpenMDW-1.1 with no addenda, and the only one to open weights, data, and recipe together; the weakness is the missing OSI certification and deep binding to NVIDIA's stack.

Tier B, open weights with strings. Muse Spark Open: weights are out, but attribution covers distillates, the $20M MaaS threshold and gated download apply, and the usage policy bans military, nuclear, critical-infrastructure, and under-18 use. For most individual and internal commercial developers, practical freedom is close to MIT — legally it is not open source, and the official safety report had not shipped.

Tier C, promise unfulfilled. Mistral Large 4: seventeen days after launch (as of October 23), no weights and no license name. Even the context window has three versions: 1M in official docs, 524,288 recorded by AA, 512K on OpenRouter. Its headline CyberGym 81.7% rests on rival models refusing 98% of tasks, which caps how much it proves.

Dilution dimensions, heaviest first: attribution obligations (Muse), revenue threshold (Muse, $20M per year), gated downloads (Muse), late weights (Mistral), license opacity (Mistral), closed data and recipe (Muse, Beam, and DeepSeek all close training data; NVIDIA is the sole exception).

6. Picking a license by use case

Your use casePreferWhy
Personal projects, internal commercial useAny of them, Muse includedMuse's threshold targets MaaS resale; internal use is exempt — mind the attribution
Running MaaS or reselling APIBeam, DeepSeek, NemotronNo revenue thresholds; Muse requires a separate agreement past $20M
Fine-tuning or distilling for redistributionBeam, DeepSeek, NemotronApache and MIT carry no derivative entanglement; Muse's attribution spreads through distillates
Reproducing training, studying recipesNemotron BoltThe only one with data and recipe fully open
One 24GB card, local deploymentNemotron BoltNative 2-bit QAT, ~24GB packed
Waiting on Mistral Large 4Keep waitingReassess when weights and license text land; only the API exists today

The short version: only the Apache 2.0 and MIT pair lets both legal and procurement sleep easy; Nemotron Bolt trades ecosystem lock-in for openness; Muse reserves its freedoms for those below the threshold; Mistral tells an open-source story with a license not yet written down. Read the license first, then do the hardware math.

FAQ

Q1: Muse Spark Open published its weights — why is it not open source? A: Open source is judged by the license, not by whether weights are downloadable. The Muse Model License requires a prominent "Built with Muse Spark Open" label, defines derivatives broadly enough to cover distillates, and adds a $20M MaaS threshold plus gated downloads. The OSI definition requires no discrimination and no field-of-use restrictions; these addenda fail that bar.

Q2: Do Apache 2.0, MIT, and OpenMDW-1.1 differ in practice? A: For most developers, barely: all permit commercial use, modification, redistribution, and distillation, with no revenue thresholds. The differences are compliance ones. Apache 2.0 and MIT are mature, OSI-certified licenses with explicit patent grants; OpenMDW-1.1 is NVIDIA's own, permissive but uncertified, which may add a legal review cycle at larger companies.

Q3: Can I self-host Mistral Large 4 today? A: No. As of 2026-10-23 there is no Large 4 checkpoint under the mistralai org on Hugging Face, the license is unnamed, and AA still files it as proprietary. The company promises weights by end of October, citing red-team work with cybersecurity and government partners. Until weights and license text land together, the API is all there is.

Q4: Which of the five can run on 24GB of VRAM? A: Only Nemotron 3.5 Bolt, trained with native 2-bit QAT, packing to about 24GB, verified on an RTX 5090. The circulating claim that DeepSeek 1017's INT4 fits a 24GB card conflicts with its 763B total parameters and reads as hype. Beam and Muse's smallest checkpoints are roughly 294GB and 700GB-plus — a different universe.

Q5: Why can't the five models' self-reported scores be compared directly? A: Different harnesses and benchmark versions. Beam's SWE-bench Verified exists as 80.9% (blog) and 81.6% (model card); DeepSWE self-reports come from different harnesses, so Beam 44.4, DeepSeek 77.0, and Muse 60 in one table means nothing; Terminal-Bench spans 2.1 and 4.0. The only reliable cross-vendor ruler today is the AA Intelligence Index v4.3.2 — noting Beam's score came from self-hosted inference.

This article is AI-assisted and human-edited. Last updated: 2026-10-23

FAQ

Muse Spark Open published its weights — why is it not open source?
Open source is judged by the license, not by whether weights are downloadable. The Muse Model License requires a prominent "Built with Muse Spark Open" label, defines derivatives broadly enough to cover distillates, and adds a $20M MaaS threshold plus gated downloads. The OSI definition requires no discrimination and no field-of-use restrictions; these addenda fail that bar.
Do Apache 2.0, MIT, and OpenMDW-1.1 differ in practice?
For most developers, barely: all permit commercial use, modification, redistribution, and distillation, with no revenue thresholds. The differences are compliance ones. Apache 2.0 and MIT are mature, OSI-certified licenses with explicit patent grants; OpenMDW-1.1 is NVIDIA's own, permissive but uncertified, which may add a legal review cycle at larger companies.
Can I self-host Mistral Large 4 today?
No. As of 2026-10-23 there is no Large 4 checkpoint under the mistralai org on Hugging Face, the license is unnamed, and AA still files it as proprietary. The company promises weights by end of October, citing red-team work with cybersecurity and government partners. Until weights and license text land together, the API is all there is.
Which of the five can run on 24GB of VRAM?
Only Nemotron 3.5 Bolt, trained with native 2-bit QAT, packing to about 24GB, verified on an RTX 5090. The circulating claim that DeepSeek 1017's INT4 fits a 24GB card conflicts with its 763B total parameters and reads as hype. Beam and Muse's smallest checkpoints are roughly 294GB and 700GB-plus — a different universe.
Why can't the five models' self-reported scores be compared directly?
Different harnesses and benchmark versions. Beam's SWE-bench Verified exists as 80.9% (blog) and 81.6% (model card); DeepSWE self-reports come from different harnesses, so Beam 44.4, DeepSeek 77.0, and Muse 60 in one table means nothing; Terminal-Bench spans 2.1 and 4.0. The only reliable cross-vendor ruler today is the AA Intelligence Index v4.3.2 — noting Beam's score came from self-hosted inference.

Related

Hardcore Reviews

Four Open Models, One 16B Base: Why Four Paths?

NVIDIA's four Cosmos 3 Nano checkpoints are called four new open-weight checkpoints in the official group blog. All four share one 16B base, the Mixture-of-Transformers (MoT) backbone and the OpenMDW1.1 license, so parameter count, license and backbone are identical cell by cell; the differences lie in input contracts, generated output and who holds physical ground truth. This comparison follows four post-training paths, Transfer-Auto, Sim-Auto, Sim-Transfer and Sim-Bimanual: from text prompts plus World Scenario Map controls to synchronized multi-view camera video and LiDAR, from the causally distilled streaming variant, through Depth, Segmentation, Edges and Blur spatial controls rendered into color observations, to video predicted from actions. Post-training is described only in the model cards' two stages, post-training then causal distillation, not as four separate recipes. All four cards list Release Date Hugging Face 10/20/2026, kept alongside the NVIDIA blog and the Hugging Face group blog as three date frames rather than merged into one launch day. No leaderboard scores, no repository metadata registry and no single-product launch report.

Nov 9, 20269 min read
Hardcore Reviews

Cheaper Per Token, Same Bill: LLM API Pricing Compared

Per-token list prices keep falling, but your bill may not: this comparison unpacks the unit-price illusion in LLM APIs - verbose output (Artificial Analysis measured K3.1 Pro at 1.7x the median token spend), cache-hit vs miss gaps up to 50x (DeepSeek 0.0028 vs 0.14 dollars), about 5.5 percent aggregator markups, and assorted surcharges. Includes a multi-source pricing table and workload-based picks, with a reminder that self-hosting moves cost rather than erasing it.

Nov 4, 20269 min read
Hardcore Reviews

Eight Flagship LLMs, Only Half Accept Your Image

A November 4, 2026 snapshot of the OpenRouter API lists 453 models, 308 of them accepting image input. This comparison puts eight flagships side by side on the same API fields: input modality, context length, and per-million-token input/output/cache-read prices. tencent/hy4 is text+image with a 1,048,576 context at 0.417/1.2505/0.021 dollars, while tencent/hy4-preview, GLM-4.6, Kimi K2 Thinking and DeepSeek V3.2 remain text-only on that channel. Includes a full table, scenario picks, and three caveats: no cross-leaderboard scoring, aggregator prices are not vendor list prices, and modality tags are not capability measurements.

Nov 4, 20269 min read