The most permissive licence anyone has attached to a model this size. Three hundred and twenty billion parameters with eighteen billion active per token, multimodal, a 300K context, and 84.3 on Terminal-Bench 2.1, four points behind its own 753B flagship. For a commercial deployment the licence is doing more work here than any benchmark on the page.
MIT is about as permissive as software licensing gets: use it, modify it, ship it, sell it, no field-of-use restriction and no user threshold. Attached to a 320-billion-parameter multimodal model, that is a genuinely unusual thing to be able to write down, and it removes an entire category of commercial risk.
Terminal-Bench 2.1 comes in at 84.3 against 88.2 for the 753B GLM-5.3. For that you get a model you can actually deploy, a licence your legal team already knows, and image input the flagship does not have. On most commercial work that trade is not a close call.
784k downloads and 2,131 likes against 442k and 1,746. The community has already voted, and it voted for the smaller, permissively licensed, multimodal one. That is a useful signal about what practitioners actually deploy rather than what benchmarks best.
Compute per token sits near an 18B dense model while the full 320B must stay resident. Cheap tokens, substantial machine, and a break-even that depends on utilisation rather than on peak throughput. The usual sparse trade, at a size that is reachable rather than aspirational.
reasoning_effort at low, high, or max. Set it per call site rather than globally: max on planning and hard debugging, low on the mechanical turns. On a long agent run that single setting moves cost more than the choice between these two models does.
The comparison that decides this is internal to the family. Flash against the 753B flagship on licence, hardware and capability, and then both against a frontier API.
| Dimension | vs GLM-5.3 (753B) | vs a frontier API |
|---|---|---|
Licence | MIT against a bespoke glm-5.3 licence. This is the difference that matters most and it points one way: MIT needs no review, carries no field-of-use restriction, and cannot be reinterpreted against you later. On client work it is usually decisive on its own. | An API has commercial terms you accept and a counterparty you can escalate to. MIT weights have neither and need neither, which is a simpler position than either alternative. |
Hardware | 320B resident against 753B, with 18B active. Still a serious deployment, but it is the difference between reachable and multi-node-only. Combined with the FP8 shipping format, this is the GLM most teams can realistically self-host. | An API needs no hardware. Self-hosting has to win on privacy, latency or sustained volume, and Z.ai's own hosted rates at $1.40 / $4.40 are low enough that the bar is higher than it would be against a frontier vendor. |
Capability | 84.3 against 88.2 on Terminal-Bench 2.1, so roughly four points for less than half the parameters. Against that, Flash adds image input, which the flagship does not have. Whether four points outweigh a modality depends entirely on what you are building. | Behind the frontier on the hardest work, as its flagship is. It closes most of the gap on mainstream coding and does not close it at the top end. Route the hardest calls up and keep the volume here. |
Multimodal | Image and text in, where the 753B flagship is text only. For document work, screenshots in a coding loop, or anything with a visual input, Flash is not the compromise choice, it is the only choice in this family. | The frontier models still lead on dense document and chart reading. For ordinary screenshots and diagrams inside an agent loop, an MIT-licensed model you host yourself is a very different proposition from a per-image API call. |
This is the GLM we reach for first. The licence removes a review, the size makes self-hosting realistic, and image input covers cases the flagship cannot. We route up to the 753B or to a frontier model when an eval on customer tasks shows the gap is real, and not before.
The permissive-licence frontier moved. Apache 2.0 at 27B from Qwen was already notable; MIT at 320B is a different statement. No field-of-use restriction, no acceptable-use policy that can be revised, no user threshold that triggers a negotiation.
The architecture pairs sparse and linear attention with mHC, a connection scheme Z.ai document in arXiv 2602.15763. It is how the model holds a 300K context at 18B active without the memory curve becoming impractical.
Flash is image-text-to-text where GLM-5.3 is text-only. The cheaper, more permissively licensed model in the family is also the more capable one on modality, which is an unusual and welcome inversion.
An unusually large output ceiling, well beyond the 128K that has become standard. For document generation, long refactors, or anything that produces a large artefact in one pass, that ceiling is the constraint that usually bites.
Three explicit levels rather than an opaque default. Set it per call site: max for planning and hard debugging, low for mechanical turns. On a long agent run this moves cost more than most model choices do.
Already quantised in the published checkpoint, so the memory bill is lower than a BF16 release of the same size and the artefact you download is the artefact that was measured. That removes the most common open-weight evaluation trap before you hit it.
The limits, the endpoints, and the lines that decide whether this model fits your transport and your budget. Kept here so nobody has to reconstruct it from three vendor pages.
| Hugging Face repo | zai-org/GLM-5.3-Flash |
|---|---|
| Licence | MIT. No field-of-use restriction, no user threshold, no bespoke terms |
| Parameters | 321,323,031,390 total, 320B nominal. Shipped in FP8 E4M3 with BF16 components |
| Active parameters | 18B per token |
| Architecture | Glm5NextForConditionalGeneration. Hybrid sparse and linear attention with Manifold-Constrained Hyper-Connections |
| Context window | 300,000 tokens, as evaluated |
| Max output | 163,840 tokens |
| Modalities | Image and text in, text out. English and Chinese |
| Reasoning control | reasoning_effort at low, high, or max |
| Serving | SGLang, vLLM, TokenSpeed, Transformers, KTransformers, Unsloth |
| Hugging Face signal | 784k downloads and 2,131 likes, ahead of the 753B flagship on both |
| Paper | arXiv 2602.15763 |
| Flagship sibling | GLM-5.3 at 753B, bespoke licence, 88.2 on Terminal-Bench 2.1 |
The weights are MIT and free. These are the numbers a self-hosted deployment gets measured against.
| Flash weights | Free | MIT. No commercial restriction and no legal review needed. |
|---|---|---|
| Resident memory | 320B params | FP8 in the published checkpoint, so lower than a BF16 release of the same size. |
| Active per token | 18B | Compute per token near an 18B dense model. |
| GLM-5.3 hosted API | $1.40 / $4.40 | The flagship through Z.ai. Cached input $0.26. |
| GLM Coding Plan | $18 to $168 / mo | Flat rate. About $12.60 to $117.60 billed annually. |
| For comparison | $10 / $50 | GPT-6 Astra and Claude Fable 5.1 per million tokens. |
The honest framing is that Z.ai's hosted pricing is low enough to make self-hosting a harder case than it would be against a frontier vendor. At $1.40 input you need real sustained volume, a privacy constraint, or a latency floor before your own GPUs win on a spreadsheet that includes engineering time. What MIT weights buy you is not cheaper inference, it is the right to deploy on your own terms, inside your own boundary, with no counterparty. That is worth a great deal on some engagements and nothing at all on others, and it is worth being clear about which one you are on.
That is the substantive benefit, and it is larger than it sounds. Community and bespoke licences carry acceptable-use policies that can be revised, thresholds that trigger renegotiation, and field-of-use limits that are read differently by different lawyers. MIT settles all of it once, and the settlement survives the vendor.
No vendor classifier between prompt and model, no refusal path, no abuse monitoring. Input filtering, output review, rate limiting and audit logging become your build. Teams consistently underestimate this and it is the usual reason a self-hosted pilot stops short of production.
Z.ai aim GLM at security research and the flagship tops CyberGym. Flash sits in the same family and behind the same absence of vendor gating. Agree scope, tool permissions, logging and a human gate on consequential actions before it goes into a client environment, rather than after.
MIT at 320B multimodal removes the conversation that stalls most open-weight plans. We have watched capable models lose to less capable ones purely on licence terms, and this is the first at this scale where that argument does not arise.
More downloads and more likes than the 753B, despite scoring four points lower. Practitioners are optimising for deployability and licence rather than for the top line, which is the same calculation we make on client work.
The flagship is text-only. A coding agent that can look at a screenshot, a document pipeline that reads a scan, a support flow that handles a photo, all of those need Flash and none of them can use GLM-5.3 at all.
Three explicit levels is a gift and leaving it at one setting for the whole session wastes it. High on planning and debugging, low on mechanical turns. This single decision moves an agent's cost more than the choice between these two models does.
At $1.40 / $4.40 the hosted path undercuts most self-hosting maths once engineering time is counted honestly. Self-host these weights for privacy, latency or genuine sustained volume. Do not self-host them to save money without doing the sums first.
Every model we integrate runs through the same operating system. Three pillars, sixteen layers, one Compound Growth Loop. The methodology that keeps AI work from rotting after the first ship.
Read the K-FrameworkDirect API integration with the model. No LangChain, no orchestration vendor, no agent framework built on quicksand. Typed contracts, the same way we wire up Postgres.
An eval suite built from your real tasks gates every prompt and model change. Quality is measured before it ships, not vibed in a demo.
Governance, audit, and oversight wired in from day one. Who called what, with which prompt version, at what cost. Your auditors get answers, not screenshots.
A model in production without observability is roulette. We instrument every integration so engineering and finance can see the same numbers, and so a regression at 3am surfaces before a customer opens a ticket.
Tokens in, tokens out, dollars spent. Sliced by feature, tenant, and route. Budgets enforced where it matters.
Real distributions, not averages. We know which routes are slow, and why.
The same eval suite that gates a release runs continuously in production. A regression on real traffic surfaces fast.
PII scrubbed at the proxy, shipped to your SIEM. Retention controls match your compliance window.
Dashboards your team owns, not ours. At handoff you get the queries, the alerts, and the runbook. We are not in the path to read your metrics.
