Inkling 2 Small vs GLM-5.3: The Open-Weights Shakeout — 17B-Active FP4 MoE vs the 753B RL Flagship

Posted by Reda Fornera on 2026-11-15
Estimated Reading Time 20 Minutes
Words 3.3k In Total

Dark circuit board with glowing traces, visualizing the two open-weights camps colliding

Inkling 2 Small vs GLM-5.3 turned out to be the open-weights story of the week — and the two models are pointed in opposite directions on capability and on price.

Here’s the collision in brief. On November 12, Thinking Machines Lab shipped open weights for Inkling 2 Small, its 594B-parameter MoE, after the model had spent its first week available only through Tinker. Two days later, llama.cpp merged full Inkling 2 support (text and vision) via pull request #32217, closing the distance between “frontier checkpoint” and “runs on my box.” Meanwhile, Z.ai’s 753B flagship GLM-5.3 saw one of its reseller-facing API cards — OpenRouter’s aggregation card — raise its output price from $1.40 to $3.60 per million tokens (+157%) in a change recorded on November 13 — the cheapest route on OpenRouter (Wafer) kept its $1.40 output price throughout. NVIDIA, meanwhile, published a guide showing GLM-5.3’s ~464 GB NVFP4 checkpoint served at 607.8 tok/s… across two DGX Stations.

So the same week that one camp made a 753B model servable and, in one channel, more expensive, another camp put a natively-quantized MoE on the shelf with a sub-200 GB quant floor and a $1.00 output price. The price war inverted: the incumbents are pushing API prices up while making weights heavier to serve; the challenger is pushing API prices down while making weights radically cheaper to run.

Let’s take them apart.

Two open-weights philosophies: FP4-native MoE vs RL-stacked flagship

Inkling 2 Small released on Tinker on November 5 (per Artificial Analysis), with Apache-2.0 weights following on the 12th (per third-party release trackers; the HF repo carries the license: apache-2.0 tag and links the Apache 2.0 text). One precision worth knowing before you assume a “do anything” license: there is no LICENSE file at the repo root — the grant lives in the model card metadata, with a separate Acceptable Use Policy linked from thinkingmachines.ai.

The architecture is the headline. Per the model card: a 48-layer decoder-only transformer with a sparse MoE feed-forward backbone — 384 experts, top-8 routing — at 594B total / 17B active parameters, hybrid local/global attention, natively multimodal (text + image in via a vision transformer), a 1M-token context, and “pretrained, midtrained and post-trained in NVFP4.” That last part matters enormously below: this isn’t a BF16 model quantized afterward. FP4 is native precision, and Thinking Machines recommends the NVFP4 checkpoint unconditionally because the higher-precision MXFP8 variant buys “a very marginal 1–2 percentage points” on evals. A reasoning_effort dial spans 0.0 to 0.99 in the chat template, from near-no-thinking to near-max.

GLM-5.3 is a different bet entirely. Z.ai’s card (created August 25) says it uses the same base model as GLM-5.2 — all improvements come from post-training. It’s a 753B-parameter MoE (40B active, per the parameter table in Thinking Machines’ card and NVIDIA’s write-up), takes text only, matches the 1M context, and always reasons — effort levels are low–high–max (per Z.ai’s docs, “disabling reasoning is no longer supported”). Where Inkling’s story is “efficient architecture, cheap tokens,” GLM-5.3’s story is “expensive RL environments, maximal effort” — days-long ML-infra-scale tasks, synthesized and verified at scale.

The headline results of that investment [VENDOR-REPORTED]: Terminal Bench 3.0 jumped from 4.6 (GLM-5.2) to 28.3; DeepSWE v1.1 from 46.2 to 66.9; CyberGym vulnerability discovery at 84.5%, claimed SOTA, with exploitation-chain scores that Z.ai says more than doubled versus GLM-5.2. Z.ai also says a GLM-powered effort flagged 2,436 vulnerabilities across 269 real-world projects — 1,097 medium-to-high severity — published via a public disclosure ledger.

License note [VENDOR-REPORTED]: GLM-5.3 on Hugging Face carries a license: other tag — Z.ai’s own model terms rather than Apache-2.0 — and NVIDIA’s NVFP4 distribution of it requires accepting both the NVIDIA Open Model Agreement and Z.ai’s original license terms. Inkling’s Apache-2.0 is, on paper, the freer-hand deal for commercial self-hosters.

Inkling 2 Small vs GLM-5.3 head-to-head benchmarks: read the small print

Here’s the core comparison table — and before you read a single number, note the framing. These rows come from Thinking Machines’ own model card, which means the comparator columns were hand-picked by Inkling’s vendor. Treat this as [VENDOR-TML] throughout: instructive, but self-interested.

Benchmark (harness) Inkling 2 Small GLM-5.3 Provenance
AA Intelligence Index (v5.0) 30.6 31.7 [VENDOR-TML, third-party data]
Vals Index (v2.2) 56.1% 59.7% [VENDOR-TML]
DeepSWE v1.1 (mini-swe-agent) 72.1% 69.0% [VENDOR-TML]
SWEBench Pro v2 (Hard) 86.3% 84.3% [VENDOR-TML]
AA-Briefcase v1.1 1527 1503 [VENDOR-TML]
DesignArena 3D Design 1457 1304 [VENDOR-TML]
DesignArena Game Dev 1424 1292 [VENDOR-TML]
BrowseComp (with compaction) 90.2% 75.1% [VENDOR-TML]

Illustrative macro photo of GPU chips — note: this is a hardware photo for illustration, not an actual benchmark chart

Two things to notice. First, GLM wins cells inside Inkling’s own table — the AA Index (31.7 vs 30.6), the Vals Index (59.7 vs 56.1), and GDPval-AA (1657 vs 1630) — a coding-agency-leaning spread consistent with Z.ai’s claims. But Inkling takes SWEBench Pro v2 Hard (86.3% vs 84.3%) and DeepSWE v1.1 (72.1% vs 69.0%) right back, so the coding split is genuinely mixed inside the vendor’s own table. GLM also sweeps the terminal benchmark row where it counts for its identity story: Z.ai puts Terminal Bench 2.1 at 88.2 and Agents’ Last Exam (CLI) at 28.5 [VENDOR-REPORTED]. Second, Inkling’s visual/creative margins are enormous — 1457 vs 1304 in 3D design — because GLM-5.3 is text-only, making the multimodal rows a category advantage rather than a quality advantage.

Now the independent reads. DesignArena’s live leaderboard (updated Nov 15, 6:12 UTC) has Inkling 2 Small at #6 overall, Elo 1314, 57.2% win rate, sandwiched between Claude Fable 5.1 and Gemini 4 Carbon [THIRD-PARTY]. That’s down from its announcement-day figure — DesignArena’s launch claim put it at 1376, #1 among open-weights models, as circulated on November 5 and relayed via Techmeme [THIRD-PARTY, announcement snapshot — not the live board]. GLM-5.3’s live category Elos sit in the 1265–1295 range (Website 1279, Web Apps 1270, Game Dev 1288, 3D 1295, per DesignArena data on OpenRouter) [THIRD-PARTY].

Artificial Analysis scores Inkling 2 Small at 31 on its Intelligence Index (#4 of 29) [THIRD-PARTY]. GLM-5.3 (Max) sits at 31.7 in AA’s data, and TML’s table for GLM-5.3 — 31.7 — matches that figure exactly; there is no AA-vs-card gap on the GLM side. The only two-way gap is Inkling’s: AA rounds to 31, the card lists 30.6. Prefer the current AA figure; hold the vendor table loosely. One AA caveat worth keeping: Inkling 2 Small is flagged “notably slow per task and very verbose” — 28.5 minutes per Index task, ~88 tok/s output as of November 15 [THIRD-PARTY]. Speed is a real cost.

The honest skew: there is no deep third-party eval suite of Inkling 2 Small’s shipped weights yet — the coverage that exists is vendor-posted plus early community results. Community-attached evaluation results landed in the HF repo on November 13 (Terminal-Bench 2.1, Terminal-Bench v4.0, SWE-bench Multilingual, GPQA Diamond, HMMT February 2026), though most of those figures trace back to the vendor’s own card rather than new independent runs. Beyond that, the model-card table plus Unsloth’s quant-quality curves (agreement against the model’s own BF16 GGUF, not real benchmarks) are what we have. Treat every number above that originated with TML as a claim pending replication.

The self-host gate: 171.7 GB GGUF vs two DGX Stations

This is where the philosophies stop being abstract. Inkling’s NVFP4 checkpoint is 393.5 GB of files: 24 safetensors shards summing to 390.06 GB plus a 3.42 GB MTP head (verified against the Hugging Face tree API; the model index counts 594,050,672,720 parameters). But files are not VRAM: the card’s operational gate for NVFP4 is ≥500 GB of aggregated VRAM — two B300s or four H200s — and the MXFP8 variant needs ≥800 GB.

The quant ecosystem is where Inkling’s story gets fun. Unsloth’s dynamic GGUF set starts at UD-IQ1_S, 171.7 GB, with 89.7% same-top-token agreement against their BF16 reference — through UD-IQ2_XXS at 190.6 GB and a 484.4 GB BF16 rung (sizes verified from the repo table; file size ≠ total inference memory — context, KV cache, and OS overhead add headroom). That puts a quantized Inkling inside a 256 GB-class workstation, with vision via a separate projector file. The catch, per Unsloth’s card: the fresh inkling2 architecture “needs a llama.cpp build with Inkling 2 support.”

Datacenter GPU server racks — the two-DGX-Station class of hardware GLM-5.3's 753B checkpoint needs to serve

That build arrived this week. Per the GitHub API, llama.cpp PR #32217 (“feat: add Inkling 2 support (text + vision)”) opened November 13 and merged November 14, 2026, at 23:51 UTC — 1,667 additions across 45 files. One footnote: the PR states it adds conversion and model support without the DSpark speculator (“a separate PR stacked on top of this”), so day-zero llama.cpp runs get no draft-model speedup. Expect the usual new-model-day jank.

GLM-5.3’s self-host math is a different sport. NVIDIA’s tutorial (published November 13) serves the ~464 GB NVFP4 checkpoint across two GB300 DGX Stations over ConnectX-8 RDMA, because one GB300 carries 252 GB of HBM3e and the checkpoint exceeds that alone. The recipe: ~218 GiB of weights per GPU, tensor parallelism TP=2 with both Stations computing every token, FP8 KV cache, prefix caching, a 4-token MTP speculative-decoding draft, and a 98K context ceiling. NVIDIA’s measurements on SPEED-Bench coding prompts: 607.8 tok/s overall at concurrency 16 (mean TTFT 482.9 ms), falling to 385.0 tok/s at C32 with 453.5-second mean request latency — per-user throughput drops from 115.3 tok/s at C1 to 13.0 tok/s at C32 [VENDOR-REPORTED, NVIDIA]. That’s the MoE-vs-flagship batching story in one table: 40B parameters hot per token means the per-user experience decays fast under load.

What about a GLM-5.3 GGUF floor? A community quant ecosystem exists — Unsloth publishes full-model GGUFs for GLM-5.3, there are expert-pruned community builds aimed at single-box targets like DGX Spark, and a slim-GGUF variant is circulating — but unlike Inkling’s card, which publishes an accuracy-vs-size table with a documented 171.7 GB floor, none of the pages fetched for this piece publish a vendor-style quality-vs-quant-size table for full GLM-5.3. The asymmetry is real: Inkling’s FP4-native design plus a published quant table makes “small enough for me” quantifiable; GLM-5.3’s community path involves more trial, more error, and heavier base math. [UNVERIFIED: smallest practical GLM-5.3 GGUF size with quality data]

The VRAM-gate economics here rhyme with the open-weights self-hosting breakdown we did earlier this month — Mistral Large 4 vs Kimi K3.1 Pro: the self-hosting economics — where the same pattern held: the model with the smaller quant floor won the GPU-hours argument regardless of which one won the benchmark table.

API pricing: Tinker’s simple card vs Z.ai’s three-card puzzle

Inkling’s card is refreshingly boring. Tinker’s pricing: $0.30/M prefill, $0.03/M cached prefill, $1.00/M sample ($1.50/M to train at 64K, $3.00 at 256K) — verified from Tinker’s published pricing page, with the same $0.30/$1.00 rates at third-party providers like Cloudflare. Three numbers, done.

GLM-5.3 is the puzzle. The three cards disagree — because they’re different channels:

  • Z.ai direct (fetched Nov 15 from Z.ai’s official pricing page): $1.40 input / $4.40 output, $0.26 cached. Still at these levels after the week’s drama.
  • The OpenRouter aggregation card: the comparison/aggregation price on OpenRouter posts $0.084 input / $3.60 output. Third-party price trackers (LLM Stats’ pricing comparison) record this card moving from $1.40 to $3.60 output (input $0.07 → $0.084) in a change logged November 13 — the +157% output hike [THIRD-PARTY]. This is the aggregation card’s price, not the market floor: OpenRouter’s cheapest actual GLM-5.3 route is Wafer at $0.0699/$1.40 [THIRD-PARTY]. The nuance: even post-hike, the aggregation card is cheaper than Z.ai’s own. The “price increase” is a reseller margin correction, not a vendor repricing.
  • The undercutting cards: other resellers price GLM-5.3 lower still — DeepInfra’s cheapest listing at $0.45/$2.00, Decart at $0.56/$1.80 [THIRD-PARTY]. And Z.ai’s own GLM-5.3-Flash runs $0.15/$0.50, FlashX $0.37/$1.25 direct — the family discounts beneath the flagship.

The honest version of the headline: the direct card didn’t move; the discounted channel repriced toward profitability, and GLM-5.3 still effectively undercuts its own card via plumbing. Supporting color: DeepInfra resells GLM-4.7 at $0.40/$1.75 [THIRD-PARTY]. Against that backdrop, Inkling’s $0.30/$1.00 isn’t just low — it’s priced like an efficiency play: 17B active parameters per token, cheap to serve rather than cheap to license.

If you want the full playbook on how open-weights API pricing vs self-hosting math settles out over a model’s lifetime, that’s exactly the lens of our earlier head-to-head — Mistral Large 4 vs Kimi K3.1 Pro — and the same reseller-vs-direct dynamics apply here.

Providers and runtimes on day 0 vs day 90

Inkling 2 Small’s rollout is the widest day-zero distribution we’ve covered. The model card lists local inference in SGLang, vLLM, TokenSpeed, and Unsloth; post-training with Miles, SkyRL, and Snowflake Cortex; and API access through NVIDIA, Fireworks, Together, Baseten, Nebius, Modal, and Cloudflare — the whole spectrum from single-node quant tinkerers to managed serverless, day one.

GLM-5.3 sits at day 90 of a longer arc, and its distribution depth shows: OpenRouter alone exposes dozens of reseller routes with published latency, throughput, and uptime stats, plus EU in-region routing. The infrastructure story is genuinely unique: Z.ai’s blogs say all production inference for GLM-5.3-Flash runs on a cluster of 100,000+ Chinese-made AI accelerators, with a GLM-5.3-powered infrastructure agent assisting engineers — the September technical account claims a roughly 3× end-to-end throughput gain [VENDOR-REPORTED]. Whatever your priors on that claim, a model family helping to build its own serving fleet on domestic silicon is a strategic statement nothing in Inkling’s distribution matches.

Runtime maturity, though, favors GLM: months of vLLM/SGLang production hardening and mature quant tooling. Inkling’s stack is nine days old, and its native speculative decoding path (DSpark) hasn’t landed in llama.cpp yet.

Which one for what: Inkling 2 Small vs GLM-5.3

Open-weights self-hosters and efficiency builders: Inkling 2 Small. Apache-2.0-tagged weights, FP4-native at 393.5 GB with a 171.7 GB quant floor and a published accuracy table, natively multimodal, a tunable thinking dial, a $0.30/$1.00 API card, and seven day-zero cloud providers with four runtimes plus training integrations. The tradeoffs are real: day-one runtime jank, no DSpark yet, no deep third-party eval suite of the shipped weights yet, and AA’s “slow and verbose” flag. If your budget line is GPU-hours rather than tokens, Inkling is the math that fits a desk instead of a datacenter.

API-first coding-agency and security shops: GLM-5.3. The proven long-horizon agent — the RL-post-trained flagship with the best public terminal and vulnerability-discovery numbers among open-weights models, months of runtime stability, a deep reseller market driving your effective price down (often below Z.ai’s list), and cheaper Flash tiers beneath it. Text-only is the main capability caveat; the license terms are the legal one; and the direction of travel — NVIDIA documenting two-Station serving while Z.ai raises sibling list prices — suggests GLM-class flags are becoming infrastructure, not commodities.

The tiebreaker question: do you need to own the inference? If yes, Inkling’s floor is roughly a quarter of GLM-5.3’s two-Station gate. If no, the reseller market gives you cheaper effective prices today than Inkling’s own card — the one place this article’s economics cut against its headline.

For a wider sweep of everything that landed this month — this matchup included — see our monthly roundup, Best new AI tools: November 2026.

What to watch next

  1. The independent-evaluation gap: Inkling 2 Small’s shipped weights have vendor-posted plus early community results (including the November 13 repo-attached batch) but no deep third-party suite yet. The first replication runs — and llama.cpp quality reports from real checkpoints — will either validate the model card or deflate it. Watch both.
  2. The DSpark PR: llama.cpp merged base support without Inkling’s native speculator. When the stacked PR lands, single-box throughput could move noticeably.
  3. Price-war follow-through: does Z.ai defend the direct $1.40/$4.40 card, cut in response to Inkling’s entry, or let the reseller market keep repricing underneath it? The Nov 13 aggregation-card hike suggests the reseller channel believes in its own margins even with the direct card holding. Track the cheapest OpenRouter card weekly — it’s the market’s live consensus.
  4. The next GLM generation: whether Z.ai’s post-training RL thesis scales to a larger base will tell us if the flagship track can close the price-locality gap Inkling just opened. [UNVERIFIED: any announced Z.ai next-gen model size or roadmap detail]
  5. DesignArena drift: Inkling’s announcement-week 1376 has already drifted to 1314 on the live board. If the drift continues, the “open-weights crown” narrative needs re-writing — and GLM-5.3’s category scores will still be sitting in the 1270s waiting.

The Inkling 2 Small vs GLM-5.3 matchup isn’t settled — it’s week one of a fight the pricing data just made interesting.

References and further reading


Please let us know if you enjoyed this blog post. Share it with others to spread the knowledge! If you believe any images in this post infringe your copyright, please contact us promptly so we can remove them.



// adding consent banner